From 9324cdaa3199d65c182858785dd1eca42b192b8e Mon Sep 17 00:00:00 2001 From: MalumaDev Date: Sun, 16 Oct 2022 17:53:56 +0200 Subject: ui fix, re organization of the code --- modules/sd_models.py | 5 ++++- 1 file changed, 4 insertions(+), 1 deletion(-) (limited to 'modules/sd_models.py') diff --git a/modules/sd_models.py b/modules/sd_models.py index 3aa21ec1..8e4ee435 100644 --- a/modules/sd_models.py +++ b/modules/sd_models.py @@ -20,7 +20,7 @@ checkpoints_loaded = collections.OrderedDict() try: # this silences the annoying "Some weights of the model checkpoint were not used when initializing..." message at start. - from transformers import logging + from transformers import logging, CLIPModel logging.set_verbosity_error() except Exception: @@ -196,6 +196,9 @@ def load_model(): sd_hijack.model_hijack.hijack(sd_model) + if shared.clip_model is None or shared.clip_model.transformer.name_or_path != sd_model.cond_stage_model.wrapped.transformer.name_or_path: + shared.clip_model = CLIPModel.from_pretrained(sd_model.cond_stage_model.wrapped.transformer.name_or_path) + sd_model.eval() print(f"Model loaded.") -- cgit v1.2.3 From 8e7097d06a6a261580d34375c9d2a9e4ffc63ffa Mon Sep 17 00:00:00 2001 From: random_thoughtss Date: Wed, 19 Oct 2022 13:47:45 -0700 Subject: Added support for RunwayML inpainting model --- modules/processing.py | 34 ++++++- modules/sd_hijack_inpainting.py | 208 ++++++++++++++++++++++++++++++++++++++++ modules/sd_models.py | 16 +++- modules/sd_samplers.py | 50 +++++++--- 4 files changed, 293 insertions(+), 15 deletions(-) create mode 100644 modules/sd_hijack_inpainting.py (limited to 'modules/sd_models.py') diff --git a/modules/processing.py b/modules/processing.py index bcb0c32c..a6c308f9 100644 --- a/modules/processing.py +++ b/modules/processing.py @@ -546,7 +546,16 @@ class StableDiffusionProcessingTxt2Img(StableDiffusionProcessing): if not self.enable_hr: x = create_random_tensors([opt_C, self.height // opt_f, self.width // opt_f], seeds=seeds, subseeds=subseeds, subseed_strength=self.subseed_strength, seed_resize_from_h=self.seed_resize_from_h, seed_resize_from_w=self.seed_resize_from_w, p=self) - samples = self.sampler.sample(self, x, conditioning, unconditional_conditioning) + + # The "masked-image" in this case will just be all zeros since the entire image is masked. + image_conditioning = torch.zeros(x.shape[0], 3, self.height, self.width, device=x.device) + image_conditioning = self.sd_model.get_first_stage_encoding(self.sd_model.encode_first_stage(image_conditioning)) + + # Add the fake full 1s mask to the first dimension. + image_conditioning = torch.nn.functional.pad(image_conditioning, (0, 0, 0, 0, 1, 0), value=1.0) + image_conditioning = image_conditioning.to(x.dtype) + + samples = self.sampler.sample(self, x, conditioning, unconditional_conditioning, image_conditioning=image_conditioning) return samples x = create_random_tensors([opt_C, self.firstphase_height // opt_f, self.firstphase_width // opt_f], seeds=seeds, subseeds=subseeds, subseed_strength=self.subseed_strength, seed_resize_from_h=self.seed_resize_from_h, seed_resize_from_w=self.seed_resize_from_w, p=self) @@ -714,10 +723,31 @@ class StableDiffusionProcessingImg2Img(StableDiffusionProcessing): elif self.inpainting_fill == 3: self.init_latent = self.init_latent * self.mask + if self.image_mask is not None: + conditioning_mask = np.array(self.image_mask.convert("L")) + conditioning_mask = conditioning_mask.astype(np.float32) / 255.0 + conditioning_mask = torch.from_numpy(conditioning_mask[None, None]) + + # Inpainting model uses a discretized mask as input, so we round to either 1.0 or 0.0 + conditioning_mask = torch.round(conditioning_mask) + else: + conditioning_mask = torch.ones(1, 1, *image.shape[-2:]) + + # Create another latent image, this time with a masked version of the original input. + conditioning_mask = conditioning_mask.to(image.device) + conditioning_image = image * (1.0 - conditioning_mask) + conditioning_image = self.sd_model.get_first_stage_encoding(self.sd_model.encode_first_stage(conditioning_image)) + + # Create the concatenated conditioning tensor to be fed to `c_concat` + conditioning_mask = torch.nn.functional.interpolate(conditioning_mask, size=self.init_latent.shape[-2:]) + conditioning_mask = conditioning_mask.expand(conditioning_image.shape[0], -1, -1, -1) + self.image_conditioning = torch.cat([conditioning_mask, conditioning_image], dim=1) + self.image_conditioning = self.image_conditioning.to(shared.device).type(self.sd_model.dtype) + def sample(self, conditioning, unconditional_conditioning, seeds, subseeds, subseed_strength): x = create_random_tensors([opt_C, self.height // opt_f, self.width // opt_f], seeds=seeds, subseeds=subseeds, subseed_strength=self.subseed_strength, seed_resize_from_h=self.seed_resize_from_h, seed_resize_from_w=self.seed_resize_from_w, p=self) - samples = self.sampler.sample_img2img(self, self.init_latent, x, conditioning, unconditional_conditioning) + samples = self.sampler.sample_img2img(self, self.init_latent, x, conditioning, unconditional_conditioning, image_conditioning=self.image_conditioning) if self.mask is not None: samples = samples * self.nmask + self.init_latent * self.mask diff --git a/modules/sd_hijack_inpainting.py b/modules/sd_hijack_inpainting.py new file mode 100644 index 00000000..7e5670d6 --- /dev/null +++ b/modules/sd_hijack_inpainting.py @@ -0,0 +1,208 @@ +import torch +import numpy as np + +from tqdm import tqdm +from einops import rearrange, repeat +from omegaconf import ListConfig + +from types import MethodType + +import ldm.models.diffusion.ddpm +import ldm.models.diffusion.ddim + +from ldm.models.diffusion.ddpm import LatentDiffusion +from ldm.models.diffusion.ddim import DDIMSampler, noise_like + +# ================================================================================================= +# Monkey patch DDIMSampler methods from RunwayML repo directly. +# Adapted from: +# https://github.com/runwayml/stable-diffusion/blob/main/ldm/models/diffusion/ddim.py +# ================================================================================================= +@torch.no_grad() +def sample( + self, + S, + batch_size, + shape, + conditioning=None, + callback=None, + normals_sequence=None, + img_callback=None, + quantize_x0=False, + eta=0., + mask=None, + x0=None, + temperature=1., + noise_dropout=0., + score_corrector=None, + corrector_kwargs=None, + verbose=True, + x_T=None, + log_every_t=100, + unconditional_guidance_scale=1., + unconditional_conditioning=None, + # this has to come in the same format as the conditioning, # e.g. as encoded tokens, ... + **kwargs + ): + if conditioning is not None: + if isinstance(conditioning, dict): + ctmp = conditioning[list(conditioning.keys())[0]] + while isinstance(ctmp, list): + ctmp = elf.inpainting_fill == 2: + self.init_latent = self.init_latent * self.mask + create_random_tensors(self.init_latent.shape[1:], all_seeds[0:self.init_latent.shape[0]]) * self.nmask + elif self.inpainting_fill == 3: + self.init_latent = self.init_latent * self.mask + + if self.image_mask is not None: + conditioning_mask = np.array(self.image_mask.convert("L")) + conditioning_mask = conditioning_mask.astype(np.float32) / 255.0 + conditioning_mask = torch.from_numpy(conditioning_mask[None, None]) + + # Inpainting model uses a discretized mask as input, so we round to either 1.0 or 0.0 + conditioning_mask = torch.round(conditioning_mask) + else: + conditioning_mask = torch.ones(1, 1, *image.shape[-2:]) + + # Create another latent image, this time with a masked version of the original input. + conditioning_mask = conditioning_mask.to(image.device) + conditioning_image = image * (1.0 - conditioning_mask) + conditioning_image = self.sd_model.get_first_stage_encoding(self.sd_model.encode_first_stage(conditioning_image)) + + # Create the concatenated conditioning tensor to be fed to `c_concat` + conditioning_mask = torch.nn.functional.interpolate(conditioning_mask, size=self.init_latent.shape[-2:]) + conditioning_mask = conditioning_mask.expand(conditioning_image.shape[0], -1, -1, -1) + self.image_conditioning = torch.cat([conditioning_mask, conditioning_image], dim=1) + self.image_conditioning = self.image_conditioning.to(shared.device).type(self.sd_model.dtype) + + def sample(self, conditioning, unconditional_conditioning, seeds, subseeds, subseed_strength): + x = create_random_tensors([opctmp[0] + cbs = ctmp.shape[0] + if cbs != batch_size: + print(f"Warning: Got {cbs} conditionings but batch-size is {batch_size}") + else: + if conditioning.shape[0] != batch_size: + print(f"Warning: Got {conditioning.shape[0]} conditionings but batch-size is {batch_size}") + + self.make_schedule(ddim_num_steps=S, ddim_eta=eta, verbose=verbose) + # sampling + C, H, W = shape + size = (batch_size, C, H, W) + print(f'Data shape for DDIM sampling is {size}, eta {eta}') + + samples, intermediates = self.ddim_sampling(conditioning, size, + callback=callback, + img_callback=img_callback, + quantize_denoised=quantize_x0, + mask=mask, x0=x0, + ddim_use_original_steps=False, + noise_dropout=noise_dropout, + temperature=temperature, + score_corrector=score_corrector, + corrector_kwargs=corrector_kwargs, + x_T=x_T, + log_every_t=log_every_t, + unconditional_guidance_scale=unconditional_guidance_scale, + unconditional_conditioning=unconditional_conditioning, + ) + return samples, intermediates + + +@torch.no_grad() +def p_sample_ddim(self, x, c, t, index, repeat_noise=False, use_original_steps=False, quantize_denoised=False, + temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None, + unconditional_guidance_scale=1., unconditional_conditioning=None): + b, *_, device = *x.shape, x.device + + if unconditional_conditioning is None or unconditional_guidance_scale == 1.: + e_t = self.model.apply_model(x, t, c) + else: + x_in = torch.cat([x] * 2) + t_in = torch.cat([t] * 2) + if isinstance(c, dict): + assert isinstance(unconditional_conditioning, dict) + c_in = dict() + for k in c: + if isinstance(c[k], list): + c_in[k] = [ + torch.cat([unconditional_conditioning[k][i], c[k][i]]) + for i in range(len(c[k])) + ] + else: + c_in[k] = torch.cat([unconditional_conditioning[k], c[k]]) + else: + c_in = torch.cat([unconditional_conditioning, c]) + e_t_uncond, e_t = self.model.apply_model(x_in, t_in, c_in).chunk(2) + e_t = e_t_uncond + unconditional_guidance_scale * (e_t - e_t_uncond) + + if score_corrector is not None: + assert self.model.parameterization == "eps" + e_t = score_corrector.modify_score(self.model, e_t, x, t, c, **corrector_kwargs) + + alphas = self.model.alphas_cumprod if use_original_steps else self.ddim_alphas + alphas_prev = self.model.alphas_cumprod_prev if use_original_steps else self.ddim_alphas_prev + sqrt_one_minus_alphas = self.model.sqrt_one_minus_alphas_cumprod if use_original_steps else self.ddim_sqrt_one_minus_alphas + sigmas = self.model.ddim_sigmas_for_original_num_steps if use_original_steps else self.ddim_sigmas + # select parameters corresponding to the currently considered timestep + a_t = torch.full((b, 1, 1, 1), alphas[index], device=device) + a_prev = torch.full((b, 1, 1, 1), alphas_prev[index], device=device) + sigma_t = torch.full((b, 1, 1, 1), sigmas[index], device=device) + sqrt_one_minus_at = torch.full((b, 1, 1, 1), sqrt_one_minus_alphas[index],device=device) + + # current prediction for x_0 + pred_x0 = (x - sqrt_one_minus_at * e_t) / a_t.sqrt() + if quantize_denoised: + pred_x0, _, *_ = self.model.first_stage_model.quantize(pred_x0) + # direction pointing to x_t + dir_xt = (1. - a_prev - sigma_t**2).sqrt() * e_t + noise = sigma_t * noise_like(x.shape, device, repeat_noise) * temperature + if noise_dropout > 0.: + noise = torch.nn.functional.dropout(noise, p=noise_dropout) + x_prev = a_prev.sqrt() * pred_x0 + dir_xt + noise + return x_prev, pred_x0 + + +# ================================================================================================= +# Monkey patch LatentInpaintDiffusion to load the checkpoint with a proper config. +# Adapted from: +# https://github.com/runwayml/stable-diffusion/blob/main/ldm/models/diffusion/ddpm.py +# ================================================================================================= + +@torch.no_grad() +def get_unconditional_conditioning(self, batch_size, null_label=None): + if null_label is not None: + xc = null_label + if isinstance(xc, ListConfig): + xc = list(xc) + if isinstance(xc, dict) or isinstance(xc, list): + c = self.get_learned_conditioning(xc) + else: + if hasattr(xc, "to"): + xc = xc.to(self.device) + c = self.get_learned_conditioning(xc) + else: + # todo: get null label from cond_stage_model + raise NotImplementedError() + c = repeat(c, "1 ... -> b ...", b=batch_size).to(self.device) + return c + +class LatentInpaintDiffusion(LatentDiffusion): + def __init__( + self, + concat_keys=("mask", "masked_image"), + masked_image_key="masked_image", + *args, + **kwargs, + ): + super().__init__(*args, **kwargs) + self.masked_image_key = masked_image_key + assert self.masked_image_key in concat_keys + self.concat_keys = concat_keys + +def should_hijack_inpainting(checkpoint_info): + return str(checkpoint_info.filename).endswith("inpainting.ckpt") and not checkpoint_info.config.endswith("inpainting.yaml") + +def do_inpainting_hijack(): + ldm.models.diffusion.ddpm.get_unconditional_conditioning = get_unconditional_conditioning + ldm.models.diffusion.ddpm.LatentInpaintDiffusion = LatentInpaintDiffusion + ldm.models.diffusion.ddim.DDIMSampler.p_sample_ddim = p_sample_ddim + ldm.models.diffusion.ddim.DDIMSampler.sample = sample \ No newline at end of file diff --git a/modules/sd_models.py b/modules/sd_models.py index eae22e87..47836d25 100644 --- a/modules/sd_models.py +++ b/modules/sd_models.py @@ -9,6 +9,7 @@ from ldm.util import instantiate_from_config from modules import shared, modelloader, devices from modules.paths import models_path +from modules.sd_hijack_inpainting import do_inpainting_hijack, should_hijack_inpainting model_dir = "Stable-diffusion" model_path = os.path.abspath(os.path.join(models_path, model_dir)) @@ -211,6 +212,19 @@ def load_model(): print(f"Loading config from: {checkpoint_info.config}") sd_config = OmegaConf.load(checkpoint_info.config) + + if should_hijack_inpainting(checkpoint_info): + do_inpainting_hijack() + + # Hardcoded config for now... + sd_config.model.target = "ldm.models.diffusion.ddpm.LatentInpaintDiffusion" + sd_config.model.params.use_ema = False + sd_config.model.params.conditioning_key = "hybrid" + sd_config.model.params.unet_config.params.in_channels = 9 + + # Create a "fake" config with a different name so that we know to unload it when switching models. + checkpoint_info = checkpoint_info._replace(config=checkpoint_info.config.replace(".yaml", "-inpainting.yaml")) + sd_model = instantiate_from_config(sd_config.model) load_model_weights(sd_model, checkpoint_info) @@ -234,7 +248,7 @@ def reload_model_weights(sd_model, info=None): if sd_model.sd_model_checkpoint == checkpoint_info.filename: return - if sd_model.sd_checkpoint_info.config != checkpoint_info.config: + if sd_model.sd_checkpoint_info.config != checkpoint_info.config or should_hijack_inpainting(checkpoint_info) != should_hijack_inpainting(sd_model.sd_checkpoint_info): checkpoints_loaded.clear() shared.sd_model = load_model() return shared.sd_model diff --git a/modules/sd_samplers.py b/modules/sd_samplers.py index b58e810b..9d3cf289 100644 --- a/modules/sd_samplers.py +++ b/modules/sd_samplers.py @@ -136,9 +136,15 @@ class VanillaStableDiffusionSampler: if self.stop_at is not None and self.step > self.stop_at: raise InterruptedException + # Have to unwrap the inpainting conditioning here to perform pre-preocessing + image_conditioning = None + if isinstance(cond, dict): + image_conditioning = cond["c_concat"][0] + cond = cond["c_crossattn"][0] + unconditional_conditioning = unconditional_conditioning["c_crossattn"][0] conds_list, tensor = prompt_parser.reconstruct_multicond_batch(cond, self.step) - unconditional_conditioning = prompt_parser.reconstruct_cond_batch(unconditional_conditioning, self.step) + unconditional_conditioning = prompt_parser.reconstruct_cond_batch(unconditional_conditioning, self.step) assert all([len(conds) == 1 for conds in conds_list]), 'composition via AND is not supported for DDIM/PLMS samplers' cond = tensor @@ -157,6 +163,10 @@ class VanillaStableDiffusionSampler: img_orig = self.sampler.model.q_sample(self.init_latent, ts) x_dec = img_orig * self.mask + self.nmask * x_dec + if image_conditioning is not None: + cond = {"c_concat": [image_conditioning], "c_crossattn": [cond]} + unconditional_conditioning = {"c_concat": [image_conditioning], "c_crossattn": [unconditional_conditioning]} + res = self.orig_p_sample_ddim(x_dec, cond, ts, unconditional_conditioning=unconditional_conditioning, *args, **kwargs) if self.mask is not None: @@ -182,7 +192,7 @@ class VanillaStableDiffusionSampler: self.mask = p.mask if hasattr(p, 'mask') else None self.nmask = p.nmask if hasattr(p, 'nmask') else None - def sample_img2img(self, p, x, noise, conditioning, unconditional_conditioning, steps=None): + def sample_img2img(self, p, x, noise, conditioning, unconditional_conditioning, steps=None, image_conditioning=None): steps, t_enc = setup_img2img_steps(p, steps) self.initialize(p) @@ -202,7 +212,7 @@ class VanillaStableDiffusionSampler: return samples - def sample(self, p, x, conditioning, unconditional_conditioning, steps=None): + def sample(self, p, x, conditioning, unconditional_conditioning, steps=None, image_conditioning=None): self.initialize(p) self.init_latent = None @@ -210,6 +220,11 @@ class VanillaStableDiffusionSampler: steps = steps or p.steps + # Wrap the conditioning models with additional image conditioning for inpainting model + if image_conditioning is not None: + conditioning = {"c_concat": [image_conditioning], "c_crossattn": [conditioning]} + unconditional_conditioning = {"c_concat": [image_conditioning], "c_crossattn": [unconditional_conditioning]} + # existing code fails with certain step counts, like 9 try: samples_ddim = self.launch_sampling(steps, lambda: self.sampler.sample(S=steps, conditioning=conditioning, batch_size=int(x.shape[0]), shape=x[0].shape, verbose=False, unconditional_guidance_scale=p.cfg_scale, unconditional_conditioning=unconditional_conditioning, x_T=x, eta=self.eta)[0]) @@ -228,7 +243,7 @@ class CFGDenoiser(torch.nn.Module): self.init_latent = None self.step = 0 - def forward(self, x, sigma, uncond, cond, cond_scale): + def forward(self, x, sigma, uncond, cond, cond_scale, image_cond): if state.interrupted or state.skipped: raise InterruptedException @@ -239,28 +254,29 @@ class CFGDenoiser(torch.nn.Module): repeats = [len(conds_list[i]) for i in range(batch_size)] x_in = torch.cat([torch.stack([x[i] for _ in range(n)]) for i, n in enumerate(repeats)] + [x]) + image_cond_in = torch.cat([torch.stack([image_cond[i] for _ in range(n)]) for i, n in enumerate(repeats)] + [image_cond]) sigma_in = torch.cat([torch.stack([sigma[i] for _ in range(n)]) for i, n in enumerate(repeats)] + [sigma]) if tensor.shape[1] == uncond.shape[1]: cond_in = torch.cat([tensor, uncond]) if shared.batch_cond_uncond: - x_out = self.inner_model(x_in, sigma_in, cond=cond_in) + x_out = self.inner_model(x_in, sigma_in, cond={"c_crossattn": [cond_in], "c_concat": [image_cond_in]}) else: x_out = torch.zeros_like(x_in) for batch_offset in range(0, x_out.shape[0], batch_size): a = batch_offset b = a + batch_size - x_out[a:b] = self.inner_model(x_in[a:b], sigma_in[a:b], cond=cond_in[a:b]) + x_out[a:b] = self.inner_model(x_in[a:b], sigma_in[a:b], cond={"c_crossattn": [cond_in[a:b]], "c_concat": [image_cond_in[a:b]]}) else: x_out = torch.zeros_like(x_in) batch_size = batch_size*2 if shared.batch_cond_uncond else batch_size for batch_offset in range(0, tensor.shape[0], batch_size): a = batch_offset b = min(a + batch_size, tensor.shape[0]) - x_out[a:b] = self.inner_model(x_in[a:b], sigma_in[a:b], cond=tensor[a:b]) + x_out[a:b] = self.inner_model(x_in[a:b], sigma_in[a:b], cond={"c_crossattn": [tensor[a:b]], "c_concat": [image_cond_in[a:b]]}) - x_out[-uncond.shape[0]:] = self.inner_model(x_in[-uncond.shape[0]:], sigma_in[-uncond.shape[0]:], cond=uncond) + x_out[-uncond.shape[0]:] = self.inner_model(x_in[-uncond.shape[0]:], sigma_in[-uncond.shape[0]:], cond={"c_crossattn": [uncond], "c_concat": [image_cond_in[-uncond.shape[0]:]]}) denoised_uncond = x_out[-uncond.shape[0]:] denoised = torch.clone(denoised_uncond) @@ -361,7 +377,7 @@ class KDiffusionSampler: return extra_params_kwargs - def sample_img2img(self, p, x, noise, conditioning, unconditional_conditioning, steps=None): + def sample_img2img(self, p, x, noise, conditioning, unconditional_conditioning, steps=None, image_conditioning=None): steps, t_enc = setup_img2img_steps(p, steps) if p.sampler_noise_scheduler_override: @@ -389,11 +405,16 @@ class KDiffusionSampler: self.model_wrap_cfg.init_latent = x - samples = self.launch_sampling(steps, lambda: self.func(self.model_wrap_cfg, xi, extra_args={'cond': conditioning, 'uncond': unconditional_conditioning, 'cond_scale': p.cfg_scale}, disable=False, callback=self.callback_state, **extra_params_kwargs)) + samples = self.launch_sampling(steps, lambda: self.func(self.model_wrap_cfg, xi, extra_args={ + 'cond': conditioning, + 'image_cond': image_conditioning, + 'uncond': unconditional_conditioning, + 'cond_scale': p.cfg_scale + }, disable=False, callback=self.callback_state, **extra_params_kwargs)) return samples - def sample(self, p, x, conditioning, unconditional_conditioning, steps=None): + def sample(self, p, x, conditioning, unconditional_conditioning, steps=None, image_conditioning = None): steps = steps or p.steps if p.sampler_noise_scheduler_override: @@ -414,7 +435,12 @@ class KDiffusionSampler: else: extra_params_kwargs['sigmas'] = sigmas - samples = self.launch_sampling(steps, lambda: self.func(self.model_wrap_cfg, x, extra_args={'cond': conditioning, 'uncond': unconditional_conditioning, 'cond_scale': p.cfg_scale}, disable=False, callback=self.callback_state, **extra_params_kwargs)) + samples = self.launch_sampling(steps, lambda: self.func(self.model_wrap_cfg, x, extra_args={ + 'cond': conditioning, + 'image_cond': image_conditioning, + 'uncond': unconditional_conditioning, + 'cond_scale': p.cfg_scale + }, disable=False, callback=self.callback_state, **extra_params_kwargs)) return samples -- cgit v1.2.3 From 708c3a7bd8ce68cbe1aa7c268e5a4b1980affc9f Mon Sep 17 00:00:00 2001 From: random_thoughtss Date: Thu, 20 Oct 2022 13:28:43 -0700 Subject: Added PLMS hijack and made sure to always replace methods --- modules/sd_hijack_inpainting.py | 163 ++++++++++++++++++++++++++++++++++++++-- modules/sd_models.py | 3 +- 2 files changed, 157 insertions(+), 9 deletions(-) (limited to 'modules/sd_models.py') diff --git a/modules/sd_hijack_inpainting.py b/modules/sd_hijack_inpainting.py index d4d28d2e..43938071 100644 --- a/modules/sd_hijack_inpainting.py +++ b/modules/sd_hijack_inpainting.py @@ -1,16 +1,14 @@ import torch -import numpy as np -from tqdm import tqdm -from einops import rearrange, repeat +from einops import repeat from omegaconf import ListConfig -from types import MethodType - import ldm.models.diffusion.ddpm import ldm.models.diffusion.ddim +import ldm.models.diffusion.plms from ldm.models.diffusion.ddpm import LatentDiffusion +from ldm.models.diffusion.plms import PLMSSampler from ldm.models.diffusion.ddim import DDIMSampler, noise_like # ================================================================================================= @@ -19,7 +17,7 @@ from ldm.models.diffusion.ddim import DDIMSampler, noise_like # https://github.com/runwayml/stable-diffusion/blob/main/ldm/models/diffusion/ddim.py # ================================================================================================= @torch.no_grad() -def sample(self, +def sample_ddim(self, S, batch_size, shape, @@ -132,6 +130,153 @@ def p_sample_ddim(self, x, c, t, index, repeat_noise=False, use_original_steps=F return x_prev, pred_x0 +# ================================================================================================= +# Monkey patch PLMSSampler methods. +# This one was not actually patched correctly in the RunwayML repo, but we can replicate the changes. +# Adapted from: +# https://github.com/CompVis/stable-diffusion/blob/main/ldm/models/diffusion/plms.py +# ================================================================================================= +@torch.no_grad() +def sample_plms(self, + S, + batch_size, + shape, + conditioning=None, + callback=None, + normals_sequence=None, + img_callback=None, + quantize_x0=False, + eta=0., + mask=None, + x0=None, + temperature=1., + noise_dropout=0., + score_corrector=None, + corrector_kwargs=None, + verbose=True, + x_T=None, + log_every_t=100, + unconditional_guidance_scale=1., + unconditional_conditioning=None, + # this has to come in the same format as the conditioning, # e.g. as encoded tokens, ... + **kwargs + ): + if conditioning is not None: + if isinstance(conditioning, dict): + ctmp = conditioning[list(conditioning.keys())[0]] + while isinstance(ctmp, list): + ctmp = ctmp[0] + cbs = ctmp.shape[0] + if cbs != batch_size: + print(f"Warning: Got {cbs} conditionings but batch-size is {batch_size}") + else: + if conditioning.shape[0] != batch_size: + print(f"Warning: Got {conditioning.shape[0]} conditionings but batch-size is {batch_size}") + + self.make_schedule(ddim_num_steps=S, ddim_eta=eta, verbose=verbose) + # sampling + C, H, W = shape + size = (batch_size, C, H, W) + print(f'Data shape for PLMS sampling is {size}') + + samples, intermediates = self.plms_sampling(conditioning, size, + callback=callback, + img_callback=img_callback, + quantize_denoised=quantize_x0, + mask=mask, x0=x0, + ddim_use_original_steps=False, + noise_dropout=noise_dropout, + temperature=temperature, + score_corrector=score_corrector, + corrector_kwargs=corrector_kwargs, + x_T=x_T, + log_every_t=log_every_t, + unconditional_guidance_scale=unconditional_guidance_scale, + unconditional_conditioning=unconditional_conditioning, + ) + return samples, intermediates + + +@torch.no_grad() +def p_sample_plms(self, x, c, t, index, repeat_noise=False, use_original_steps=False, quantize_denoised=False, + temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None, + unconditional_guidance_scale=1., unconditional_conditioning=None, old_eps=None, t_next=None): + b, *_, device = *x.shape, x.device + + def get_model_output(x, t): + if unconditional_conditioning is None or unconditional_guidance_scale == 1.: + e_t = self.model.apply_model(x, t, c) + else: + x_in = torch.cat([x] * 2) + t_in = torch.cat([t] * 2) + + if isinstance(c, dict): + assert isinstance(unconditional_conditioning, dict) + c_in = dict() + for k in c: + if isinstance(c[k], list): + c_in[k] = [ + torch.cat([unconditional_conditioning[k][i], c[k][i]]) + for i in range(len(c[k])) + ] + else: + c_in[k] = torch.cat([unconditional_conditioning[k], c[k]]) + else: + c_in = torch.cat([unconditional_conditioning, c]) + + e_t_uncond, e_t = self.model.apply_model(x_in, t_in, c_in).chunk(2) + e_t = e_t_uncond + unconditional_guidance_scale * (e_t - e_t_uncond) + + if score_corrector is not None: + assert self.model.parameterization == "eps" + e_t = score_corrector.modify_score(self.model, e_t, x, t, c, **corrector_kwargs) + + return e_t + + alphas = self.model.alphas_cumprod if use_original_steps else self.ddim_alphas + alphas_prev = self.model.alphas_cumprod_prev if use_original_steps else self.ddim_alphas_prev + sqrt_one_minus_alphas = self.model.sqrt_one_minus_alphas_cumprod if use_original_steps else self.ddim_sqrt_one_minus_alphas + sigmas = self.model.ddim_sigmas_for_original_num_steps if use_original_steps else self.ddim_sigmas + + def get_x_prev_and_pred_x0(e_t, index): + # select parameters corresponding to the currently considered timestep + a_t = torch.full((b, 1, 1, 1), alphas[index], device=device) + a_prev = torch.full((b, 1, 1, 1), alphas_prev[index], device=device) + sigma_t = torch.full((b, 1, 1, 1), sigmas[index], device=device) + sqrt_one_minus_at = torch.full((b, 1, 1, 1), sqrt_one_minus_alphas[index],device=device) + + # current prediction for x_0 + pred_x0 = (x - sqrt_one_minus_at * e_t) / a_t.sqrt() + if quantize_denoised: + pred_x0, _, *_ = self.model.first_stage_model.quantize(pred_x0) + # direction pointing to x_t + dir_xt = (1. - a_prev - sigma_t**2).sqrt() * e_t + noise = sigma_t * noise_like(x.shape, device, repeat_noise) * temperature + if noise_dropout > 0.: + noise = torch.nn.functional.dropout(noise, p=noise_dropout) + x_prev = a_prev.sqrt() * pred_x0 + dir_xt + noise + return x_prev, pred_x0 + + e_t = get_model_output(x, t) + if len(old_eps) == 0: + # Pseudo Improved Euler (2nd order) + x_prev, pred_x0 = get_x_prev_and_pred_x0(e_t, index) + e_t_next = get_model_output(x_prev, t_next) + e_t_prime = (e_t + e_t_next) / 2 + elif len(old_eps) == 1: + # 2nd order Pseudo Linear Multistep (Adams-Bashforth) + e_t_prime = (3 * e_t - old_eps[-1]) / 2 + elif len(old_eps) == 2: + # 3nd order Pseudo Linear Multistep (Adams-Bashforth) + e_t_prime = (23 * e_t - 16 * old_eps[-1] + 5 * old_eps[-2]) / 12 + elif len(old_eps) >= 3: + # 4nd order Pseudo Linear Multistep (Adams-Bashforth) + e_t_prime = (55 * e_t - 59 * old_eps[-1] + 37 * old_eps[-2] - 9 * old_eps[-3]) / 24 + + x_prev, pred_x0 = get_x_prev_and_pred_x0(e_t_prime, index) + + return x_prev, pred_x0, e_t + # ================================================================================================= # Monkey patch LatentInpaintDiffusion to load the checkpoint with a proper config. # Adapted from: @@ -175,5 +320,9 @@ def should_hijack_inpainting(checkpoint_info): def do_inpainting_hijack(): ldm.models.diffusion.ddpm.get_unconditional_conditioning = get_unconditional_conditioning ldm.models.diffusion.ddpm.LatentInpaintDiffusion = LatentInpaintDiffusion + ldm.models.diffusion.ddim.DDIMSampler.p_sample_ddim = p_sample_ddim - ldm.models.diffusion.ddim.DDIMSampler.sample = sample \ No newline at end of file + ldm.models.diffusion.ddim.DDIMSampler.sample = sample_ddim + + ldm.models.diffusion.plms.PLMSSampler.p_sample_plms = p_sample_plms + ldm.models.diffusion.plms.PLMSSampler.sample = sample_plms \ No newline at end of file diff --git a/modules/sd_models.py b/modules/sd_models.py index 47836d25..7072db08 100644 --- a/modules/sd_models.py +++ b/modules/sd_models.py @@ -214,8 +214,6 @@ def load_model(): sd_config = OmegaConf.load(checkpoint_info.config) if should_hijack_inpainting(checkpoint_info): - do_inpainting_hijack() - # Hardcoded config for now... sd_config.model.target = "ldm.models.diffusion.ddpm.LatentInpaintDiffusion" sd_config.model.params.use_ema = False @@ -225,6 +223,7 @@ def load_model(): # Create a "fake" config with a different name so that we know to unload it when switching models. checkpoint_info = checkpoint_info._replace(config=checkpoint_info.config.replace(".yaml", "-inpainting.yaml")) + do_inpainting_hijack() sd_model = instantiate_from_config(sd_config.model) load_model_weights(sd_model, checkpoint_info) -- cgit v1.2.3 From 49533eed9e3aad19e9868ee140708baec4fd44be Mon Sep 17 00:00:00 2001 From: random_thoughtss Date: Thu, 20 Oct 2022 16:01:27 -0700 Subject: XY grid correctly re-assignes model when config changes --- modules/sd_models.py | 6 +++--- scripts/xy_grid.py | 1 + 2 files changed, 4 insertions(+), 3 deletions(-) (limited to 'modules/sd_models.py') diff --git a/modules/sd_models.py b/modules/sd_models.py index 7072db08..fea84630 100644 --- a/modules/sd_models.py +++ b/modules/sd_models.py @@ -204,9 +204,9 @@ def load_model_weights(model, checkpoint_info): model.sd_checkpoint_info = checkpoint_info -def load_model(): +def load_model(checkpoint_info=None): from modules import lowvram, sd_hijack - checkpoint_info = select_checkpoint() + checkpoint_info = checkpoint_info or select_checkpoint() if checkpoint_info.config != shared.cmd_opts.config: print(f"Loading config from: {checkpoint_info.config}") @@ -249,7 +249,7 @@ def reload_model_weights(sd_model, info=None): if sd_model.sd_checkpoint_info.config != checkpoint_info.config or should_hijack_inpainting(checkpoint_info) != should_hijack_inpainting(sd_model.sd_checkpoint_info): checkpoints_loaded.clear() - shared.sd_model = load_model() + shared.sd_model = load_model(checkpoint_info) return shared.sd_model if shared.cmd_opts.lowvram or shared.cmd_opts.medvram: diff --git a/scripts/xy_grid.py b/scripts/xy_grid.py index 5cca168a..eff0c942 100644 --- a/scripts/xy_grid.py +++ b/scripts/xy_grid.py @@ -89,6 +89,7 @@ def apply_checkpoint(p, x, xs): if info is None: raise RuntimeError(f"Unknown checkpoint: {x}") modules.sd_models.reload_model_weights(shared.sd_model, info) + p.sd_model = shared.sd_model def confirm_checkpoints(p, xs): -- cgit v1.2.3 From df5706409386cc2e88718bd9101045587c39f8bb Mon Sep 17 00:00:00 2001 From: AUTOMATIC <16777216c@gmail.com> Date: Fri, 21 Oct 2022 16:10:51 +0300 Subject: do not load aesthetic clip model until it's needed add refresh button for aesthetic embeddings add aesthetic params to images' infotext --- modules/aesthetic_clip.py | 40 +++++++++++++++++++---- modules/generation_parameters_copypaste.py | 18 +++++++++-- modules/img2img.py | 5 +-- modules/processing.py | 4 +-- modules/sd_models.py | 3 -- modules/txt2img.py | 4 +-- modules/ui.py | 52 ++++++++++++++++++++---------- style.css | 2 +- 8 files changed, 89 insertions(+), 39 deletions(-) (limited to 'modules/sd_models.py') diff --git a/modules/aesthetic_clip.py b/modules/aesthetic_clip.py index 34efa931..8c828541 100644 --- a/modules/aesthetic_clip.py +++ b/modules/aesthetic_clip.py @@ -40,6 +40,8 @@ def iter_to_batched(iterable, n=1): def create_ui(): + import modules.ui + with gr.Group(): with gr.Accordion("Open for Clip Aesthetic!", open=False): with gr.Row(): @@ -55,6 +57,8 @@ def create_ui(): label="Aesthetic imgs embedding", value="None") + modules.ui.create_refresh_button(aesthetic_imgs, shared.update_aesthetic_embeddings, lambda: {"choices": sorted(shared.aesthetic_embeddings.keys())}, "refresh_aesthetic_embeddings") + with gr.Row(): aesthetic_imgs_text = gr.Textbox(label='Aesthetic text for imgs', placeholder="This text is used to rotate the feature space of the imgs embs", @@ -66,11 +70,21 @@ def create_ui(): return aesthetic_weight, aesthetic_steps, aesthetic_lr, aesthetic_slerp, aesthetic_imgs, aesthetic_imgs_text, aesthetic_slerp_angle, aesthetic_text_negative +aesthetic_clip_model = None + + +def aesthetic_clip(): + global aesthetic_clip_model + + if aesthetic_clip_model is None or aesthetic_clip_model.name_or_path != shared.sd_model.cond_stage_model.wrapped.transformer.name_or_path: + aesthetic_clip_model = CLIPModel.from_pretrained(shared.sd_model.cond_stage_model.wrapped.transformer.name_or_path) + aesthetic_clip_model.cpu() + + return aesthetic_clip_model + + def generate_imgs_embd(name, folder, batch_size): - # clipModel = CLIPModel.from_pretrained( - # shared.sd_model.cond_stage_model.clipModel.name_or_path - # ) - model = shared.clip_model.to(device) + model = aesthetic_clip().to(device) processor = CLIPProcessor.from_pretrained(model.name_or_path) with torch.no_grad(): @@ -91,7 +105,7 @@ def generate_imgs_embd(name, folder, batch_size): path = str(Path(shared.cmd_opts.aesthetic_embeddings_dir) / f"{name}.pt") torch.save(embs, path) - model = model.cpu() + model.cpu() del processor del embs gc.collect() @@ -132,7 +146,7 @@ class AestheticCLIP: self.image_embs = None self.load_image_embs(None) - def set_aesthetic_params(self, aesthetic_lr=0, aesthetic_weight=0, aesthetic_steps=0, image_embs_name=None, + def set_aesthetic_params(self, p, aesthetic_lr=0, aesthetic_weight=0, aesthetic_steps=0, image_embs_name=None, aesthetic_slerp=True, aesthetic_imgs_text="", aesthetic_slerp_angle=0.15, aesthetic_text_negative=False): @@ -145,6 +159,18 @@ class AestheticCLIP: self.aesthetic_steps = aesthetic_steps self.load_image_embs(image_embs_name) + if self.image_embs_name is not None: + p.extra_generation_params.update({ + "Aesthetic LR": aesthetic_lr, + "Aesthetic weight": aesthetic_weight, + "Aesthetic steps": aesthetic_steps, + "Aesthetic embedding": self.image_embs_name, + "Aesthetic slerp": aesthetic_slerp, + "Aesthetic text": aesthetic_imgs_text, + "Aesthetic text negative": aesthetic_text_negative, + "Aesthetic slerp angle": aesthetic_slerp_angle, + }) + def set_skip(self, skip): self.skip = skip @@ -168,7 +194,7 @@ class AestheticCLIP: tokens = torch.asarray(remade_batch_tokens).to(device) - model = copy.deepcopy(shared.clip_model).to(device) + model = copy.deepcopy(aesthetic_clip()).to(device) model.requires_grad_(True) if self.aesthetic_imgs_text is not None and len(self.aesthetic_imgs_text) > 0: text_embs_2 = model.get_text_features( diff --git a/modules/generation_parameters_copypaste.py b/modules/generation_parameters_copypaste.py index 0f041449..f73647da 100644 --- a/modules/generation_parameters_copypaste.py +++ b/modules/generation_parameters_copypaste.py @@ -4,13 +4,22 @@ import gradio as gr from modules.shared import script_path from modules import shared -re_param_code = r"\s*([\w ]+):\s*([^,]+)(?:,|$)" +re_param_code = r'\s*([\w ]+):\s*("(?:\\|\"|[^\"])+"|[^,]*)(?:,|$)' re_param = re.compile(re_param_code) re_params = re.compile(r"^(?:" + re_param_code + "){3,}$") re_imagesize = re.compile(r"^(\d+)x(\d+)$") type_of_gr_update = type(gr.update()) +def quote(text): + if ',' not in str(text): + return text + + text = str(text) + text = text.replace('\\', '\\\\') + text = text.replace('"', '\\"') + return f'"{text}"' + def parse_generation_parameters(x: str): """parses generation parameters string, the one you see in text field under the picture in UI: ``` @@ -83,7 +92,12 @@ def connect_paste(button, paste_fields, input_comp, js=None): else: try: valtype = type(output.value) - val = valtype(v) + + if valtype == bool and v == "False": + val = False + else: + val = valtype(v) + res.append(gr.update(value=val)) except Exception: res.append(gr.update()) diff --git a/modules/img2img.py b/modules/img2img.py index bc7c66bc..eea5199b 100644 --- a/modules/img2img.py +++ b/modules/img2img.py @@ -109,10 +109,7 @@ def img2img(mode: int, prompt: str, negative_prompt: str, prompt_style: str, pro inpainting_mask_invert=inpainting_mask_invert, ) - shared.aesthetic_clip.set_aesthetic_params(float(aesthetic_lr), float(aesthetic_weight), int(aesthetic_steps), - aesthetic_imgs, aesthetic_slerp, aesthetic_imgs_text, - aesthetic_slerp_angle, - aesthetic_text_negative) + shared.aesthetic_clip.set_aesthetic_params(p, float(aesthetic_lr), float(aesthetic_weight), int(aesthetic_steps), aesthetic_imgs, aesthetic_slerp, aesthetic_imgs_text, aesthetic_slerp_angle, aesthetic_text_negative) if shared.cmd_opts.enable_console_prompts: print(f"\nimg2img: {prompt}", file=shared.progress_print_out) diff --git a/modules/processing.py b/modules/processing.py index d1deffa9..f0852cd5 100644 --- a/modules/processing.py +++ b/modules/processing.py @@ -12,7 +12,7 @@ from skimage import exposure from typing import Any, Dict, List, Optional import modules.sd_hijack -from modules import devices, prompt_parser, masking, sd_samplers, lowvram +from modules import devices, prompt_parser, masking, sd_samplers, lowvram, generation_parameters_copypaste from modules.sd_hijack import model_hijack from modules.shared import opts, cmd_opts, state import modules.shared as shared @@ -318,7 +318,7 @@ def create_infotext(p, all_prompts, all_seeds, all_subseeds, comments, iteration generation_params.update(p.extra_generation_params) - generation_params_text = ", ".join([k if k == v else f'{k}: {v}' for k, v in generation_params.items() if v is not None]) + generation_params_text = ", ".join([k if k == v else f'{k}: {generation_parameters_copypaste.quote(v)}' for k, v in generation_params.items() if v is not None]) negative_prompt_text = "\nNegative prompt: " + p.negative_prompt if p.negative_prompt else "" diff --git a/modules/sd_models.py b/modules/sd_models.py index 05a1df28..b1c91b0d 100644 --- a/modules/sd_models.py +++ b/modules/sd_models.py @@ -234,9 +234,6 @@ def load_model(checkpoint_info=None): sd_hijack.model_hijack.hijack(sd_model) - if shared.clip_model is None or shared.clip_model.transformer.name_or_path != sd_model.cond_stage_model.wrapped.transformer.name_or_path: - shared.clip_model = CLIPModel.from_pretrained(sd_model.cond_stage_model.wrapped.transformer.name_or_path) - sd_model.eval() print(f"Model loaded.") diff --git a/modules/txt2img.py b/modules/txt2img.py index 32ed1d8d..1761cfa2 100644 --- a/modules/txt2img.py +++ b/modules/txt2img.py @@ -36,9 +36,7 @@ def txt2img(prompt: str, negative_prompt: str, prompt_style: str, prompt_style2: firstphase_height=firstphase_height if enable_hr else None, ) - shared.aesthetic_clip.set_aesthetic_params(float(aesthetic_lr), float(aesthetic_weight), int(aesthetic_steps), - aesthetic_imgs, aesthetic_slerp, aesthetic_imgs_text, aesthetic_slerp_angle, - aesthetic_text_negative) + shared.aesthetic_clip.set_aesthetic_params(p, float(aesthetic_lr), float(aesthetic_weight), int(aesthetic_steps), aesthetic_imgs, aesthetic_slerp, aesthetic_imgs_text, aesthetic_slerp_angle, aesthetic_text_negative) if cmd_opts.enable_console_prompts: print(f"\ntxt2img: {prompt}", file=shared.progress_print_out) diff --git a/modules/ui.py b/modules/ui.py index 381ca925..0d020de6 100644 --- a/modules/ui.py +++ b/modules/ui.py @@ -597,27 +597,29 @@ def apply_setting(key, value): return value -def create_ui(wrap_gradio_gpu_call): - import modules.img2img - import modules.txt2img +def create_refresh_button(refresh_component, refresh_method, refreshed_args, elem_id): + def refresh(): + refresh_method() + args = refreshed_args() if callable(refreshed_args) else refreshed_args - def create_refresh_button(refresh_component, refresh_method, refreshed_args, elem_id): - def refresh(): - refresh_method() - args = refreshed_args() if callable(refreshed_args) else refreshed_args + for k, v in args.items(): + setattr(refresh_component, k, v) - for k, v in args.items(): - setattr(refresh_component, k, v) + return gr.update(**(args or {})) - return gr.update(**(args or {})) + refresh_button = gr.Button(value=refresh_symbol, elem_id=elem_id) + refresh_button.click( + fn=refresh, + inputs=[], + outputs=[refresh_component] + ) + return refresh_button + + +def create_ui(wrap_gradio_gpu_call): + import modules.img2img + import modules.txt2img - refresh_button = gr.Button(value=refresh_symbol, elem_id=elem_id) - refresh_button.click( - fn = refresh, - inputs = [], - outputs = [refresh_component] - ) - return refresh_button with gr.Blocks(analytics_enabled=False) as txt2img_interface: txt2img_prompt, roll, txt2img_prompt_style, txt2img_negative_prompt, txt2img_prompt_style2, submit, _, _, txt2img_prompt_style_apply, txt2img_save_style, txt2img_paste, token_counter, token_button = create_toprow(is_img2img=False) @@ -802,6 +804,14 @@ def create_ui(wrap_gradio_gpu_call): (hr_options, lambda d: gr.Row.update(visible="Denoising strength" in d)), (firstphase_width, "First pass size-1"), (firstphase_height, "First pass size-2"), + (aesthetic_lr, "Aesthetic LR"), + (aesthetic_weight, "Aesthetic weight"), + (aesthetic_steps, "Aesthetic steps"), + (aesthetic_imgs, "Aesthetic embedding"), + (aesthetic_slerp, "Aesthetic slerp"), + (aesthetic_imgs_text, "Aesthetic text"), + (aesthetic_text_negative, "Aesthetic text negative"), + (aesthetic_slerp_angle, "Aesthetic slerp angle"), ] txt2img_preview_params = [ @@ -1077,6 +1087,14 @@ def create_ui(wrap_gradio_gpu_call): (seed_resize_from_w, "Seed resize from-1"), (seed_resize_from_h, "Seed resize from-2"), (denoising_strength, "Denoising strength"), + (aesthetic_lr_im, "Aesthetic LR"), + (aesthetic_weight_im, "Aesthetic weight"), + (aesthetic_steps_im, "Aesthetic steps"), + (aesthetic_imgs_im, "Aesthetic embedding"), + (aesthetic_slerp_im, "Aesthetic slerp"), + (aesthetic_imgs_text_im, "Aesthetic text"), + (aesthetic_text_negative_im, "Aesthetic text negative"), + (aesthetic_slerp_angle_im, "Aesthetic slerp angle"), ] token_button.click(fn=update_token_counter, inputs=[img2img_prompt, steps], outputs=[token_counter]) diff --git a/style.css b/style.css index 26ae36a5..5d2bacc9 100644 --- a/style.css +++ b/style.css @@ -477,7 +477,7 @@ input[type="range"]{ padding: 0; } -#refresh_sd_model_checkpoint, #refresh_sd_hypernetwork, #refresh_train_hypernetwork_name, #refresh_train_embedding_name, #refresh_localization{ +#refresh_sd_model_checkpoint, #refresh_sd_hypernetwork, #refresh_train_hypernetwork_name, #refresh_train_embedding_name, #refresh_localization, #refresh_aesthetic_embeddings{ max-width: 2.5em; min-width: 2.5em; height: 2.4em; -- cgit v1.2.3 From ac0aa2b18efeeb9220a5994c8dd54c7cdda7cc40 Mon Sep 17 00:00:00 2001 From: AUTOMATIC <16777216c@gmail.com> Date: Fri, 21 Oct 2022 17:35:51 +0300 Subject: loading SD VAE, see PR #3303 --- modules/sd_models.py | 5 ++++- 1 file changed, 4 insertions(+), 1 deletion(-) (limited to 'modules/sd_models.py') diff --git a/modules/sd_models.py b/modules/sd_models.py index b1c91b0d..d99dbce8 100644 --- a/modules/sd_models.py +++ b/modules/sd_models.py @@ -155,6 +155,9 @@ def get_state_dict_from_checkpoint(pl_sd): return pl_sd +vae_ignore_keys = {"model_ema.decay", "model_ema.num_updates"} + + def load_model_weights(model, checkpoint_info): checkpoint_file = checkpoint_info.filename sd_model_hash = checkpoint_info.hash @@ -186,7 +189,7 @@ def load_model_weights(model, checkpoint_info): if os.path.exists(vae_file): print(f"Loading VAE weights from: {vae_file}") vae_ckpt = torch.load(vae_file, map_location=shared.weight_load_location) - vae_dict = {k: v for k, v in vae_ckpt["state_dict"].items() if k[0:4] != "loss"} + vae_dict = {k: v for k, v in vae_ckpt["state_dict"].items() if k[0:4] != "loss" and k not in vae_ignore_keys} model.first_stage_model.load_state_dict(vae_dict) model.first_stage_model.to(devices.dtype_vae) -- cgit v1.2.3 From 2b91251637078e04472c91a06a8d9c4db9c1dcf0 Mon Sep 17 00:00:00 2001 From: AUTOMATIC <16777216c@gmail.com> Date: Sat, 22 Oct 2022 12:23:45 +0300 Subject: removed aesthetic gradients as built-in added support for extensions --- .gitignore | 2 +- extensions/put extension here.txt | 0 modules/aesthetic_clip.py | 241 -------------------------------------- modules/images_history.py | 2 +- modules/img2img.py | 5 +- modules/processing.py | 35 ++++-- modules/script_callbacks.py | 42 +++++++ modules/scripts.py | 210 ++++++++++++++++++++++++--------- modules/sd_hijack.py | 1 - modules/sd_models.py | 7 +- modules/shared.py | 19 --- modules/txt2img.py | 5 +- modules/ui.py | 83 ++----------- webui.py | 7 +- 14 files changed, 249 insertions(+), 410 deletions(-) create mode 100644 extensions/put extension here.txt delete mode 100644 modules/aesthetic_clip.py create mode 100644 modules/script_callbacks.py (limited to 'modules/sd_models.py') diff --git a/.gitignore b/.gitignore index f9c3357c..2f1e08ed 100644 --- a/.gitignore +++ b/.gitignore @@ -27,4 +27,4 @@ __pycache__ notification.mp3 /SwinIR /textual_inversion -.vscode \ No newline at end of file +.vscode diff --git a/extensions/put extension here.txt b/extensions/put extension here.txt new file mode 100644 index 00000000..e69de29b diff --git a/modules/aesthetic_clip.py b/modules/aesthetic_clip.py deleted file mode 100644 index 8c828541..00000000 --- a/modules/aesthetic_clip.py +++ /dev/null @@ -1,241 +0,0 @@ -import copy -import itertools -import os -from pathlib import Path -import html -import gc - -import gradio as gr -import torch -from PIL import Image -from torch import optim - -from modules import shared -from transformers import CLIPModel, CLIPProcessor, CLIPTokenizer -from tqdm.auto import tqdm, trange -from modules.shared import opts, device - - -def get_all_images_in_folder(folder): - return [os.path.join(folder, f) for f in os.listdir(folder) if - os.path.isfile(os.path.join(folder, f)) and check_is_valid_image_file(f)] - - -def check_is_valid_image_file(filename): - return filename.lower().endswith(('.png', '.jpg', '.jpeg', ".gif", ".tiff", ".webp")) - - -def batched(dataset, total, n=1): - for ndx in range(0, total, n): - yield [dataset.__getitem__(i) for i in range(ndx, min(ndx + n, total))] - - -def iter_to_batched(iterable, n=1): - it = iter(iterable) - while True: - chunk = tuple(itertools.islice(it, n)) - if not chunk: - return - yield chunk - - -def create_ui(): - import modules.ui - - with gr.Group(): - with gr.Accordion("Open for Clip Aesthetic!", open=False): - with gr.Row(): - aesthetic_weight = gr.Slider(minimum=0, maximum=1, step=0.01, label="Aesthetic weight", - value=0.9) - aesthetic_steps = gr.Slider(minimum=0, maximum=50, step=1, label="Aesthetic steps", value=5) - - with gr.Row(): - aesthetic_lr = gr.Textbox(label='Aesthetic learning rate', - placeholder="Aesthetic learning rate", value="0.0001") - aesthetic_slerp = gr.Checkbox(label="Slerp interpolation", value=False) - aesthetic_imgs = gr.Dropdown(sorted(shared.aesthetic_embeddings.keys()), - label="Aesthetic imgs embedding", - value="None") - - modules.ui.create_refresh_button(aesthetic_imgs, shared.update_aesthetic_embeddings, lambda: {"choices": sorted(shared.aesthetic_embeddings.keys())}, "refresh_aesthetic_embeddings") - - with gr.Row(): - aesthetic_imgs_text = gr.Textbox(label='Aesthetic text for imgs', - placeholder="This text is used to rotate the feature space of the imgs embs", - value="") - aesthetic_slerp_angle = gr.Slider(label='Slerp angle', minimum=0, maximum=1, step=0.01, - value=0.1) - aesthetic_text_negative = gr.Checkbox(label="Is negative text", value=False) - - return aesthetic_weight, aesthetic_steps, aesthetic_lr, aesthetic_slerp, aesthetic_imgs, aesthetic_imgs_text, aesthetic_slerp_angle, aesthetic_text_negative - - -aesthetic_clip_model = None - - -def aesthetic_clip(): - global aesthetic_clip_model - - if aesthetic_clip_model is None or aesthetic_clip_model.name_or_path != shared.sd_model.cond_stage_model.wrapped.transformer.name_or_path: - aesthetic_clip_model = CLIPModel.from_pretrained(shared.sd_model.cond_stage_model.wrapped.transformer.name_or_path) - aesthetic_clip_model.cpu() - - return aesthetic_clip_model - - -def generate_imgs_embd(name, folder, batch_size): - model = aesthetic_clip().to(device) - processor = CLIPProcessor.from_pretrained(model.name_or_path) - - with torch.no_grad(): - embs = [] - for paths in tqdm(iter_to_batched(get_all_images_in_folder(folder), batch_size), - desc=f"Generating embeddings for {name}"): - if shared.state.interrupted: - break - inputs = processor(images=[Image.open(path) for path in paths], return_tensors="pt").to(device) - outputs = model.get_image_features(**inputs).cpu() - embs.append(torch.clone(outputs)) - inputs.to("cpu") - del inputs, outputs - - embs = torch.cat(embs, dim=0).mean(dim=0, keepdim=True) - - # The generated embedding will be located here - path = str(Path(shared.cmd_opts.aesthetic_embeddings_dir) / f"{name}.pt") - torch.save(embs, path) - - model.cpu() - del processor - del embs - gc.collect() - torch.cuda.empty_cache() - res = f""" - Done generating embedding for {name}! - Aesthetic embedding saved to {html.escape(path)} - """ - shared.update_aesthetic_embeddings() - return gr.Dropdown.update(choices=sorted(shared.aesthetic_embeddings.keys()), label="Imgs embedding", - value="None"), \ - gr.Dropdown.update(choices=sorted(shared.aesthetic_embeddings.keys()), - label="Imgs embedding", - value="None"), res, "" - - -def slerp(low, high, val): - low_norm = low / torch.norm(low, dim=1, keepdim=True) - high_norm = high / torch.norm(high, dim=1, keepdim=True) - omega = torch.acos((low_norm * high_norm).sum(1)) - so = torch.sin(omega) - res = (torch.sin((1.0 - val) * omega) / so).unsqueeze(1) * low + (torch.sin(val * omega) / so).unsqueeze(1) * high - return res - - -class AestheticCLIP: - def __init__(self): - self.skip = False - self.aesthetic_steps = 0 - self.aesthetic_weight = 0 - self.aesthetic_lr = 0 - self.slerp = False - self.aesthetic_text_negative = "" - self.aesthetic_slerp_angle = 0 - self.aesthetic_imgs_text = "" - - self.image_embs_name = None - self.image_embs = None - self.load_image_embs(None) - - def set_aesthetic_params(self, p, aesthetic_lr=0, aesthetic_weight=0, aesthetic_steps=0, image_embs_name=None, - aesthetic_slerp=True, aesthetic_imgs_text="", - aesthetic_slerp_angle=0.15, - aesthetic_text_negative=False): - self.aesthetic_imgs_text = aesthetic_imgs_text - self.aesthetic_slerp_angle = aesthetic_slerp_angle - self.aesthetic_text_negative = aesthetic_text_negative - self.slerp = aesthetic_slerp - self.aesthetic_lr = aesthetic_lr - self.aesthetic_weight = aesthetic_weight - self.aesthetic_steps = aesthetic_steps - self.load_image_embs(image_embs_name) - - if self.image_embs_name is not None: - p.extra_generation_params.update({ - "Aesthetic LR": aesthetic_lr, - "Aesthetic weight": aesthetic_weight, - "Aesthetic steps": aesthetic_steps, - "Aesthetic embedding": self.image_embs_name, - "Aesthetic slerp": aesthetic_slerp, - "Aesthetic text": aesthetic_imgs_text, - "Aesthetic text negative": aesthetic_text_negative, - "Aesthetic slerp angle": aesthetic_slerp_angle, - }) - - def set_skip(self, skip): - self.skip = skip - - def load_image_embs(self, image_embs_name): - if image_embs_name is None or len(image_embs_name) == 0 or image_embs_name == "None": - image_embs_name = None - self.image_embs_name = None - if image_embs_name is not None and self.image_embs_name != image_embs_name: - self.image_embs_name = image_embs_name - self.image_embs = torch.load(shared.aesthetic_embeddings[self.image_embs_name], map_location=device) - self.image_embs /= self.image_embs.norm(dim=-1, keepdim=True) - self.image_embs.requires_grad_(False) - - def __call__(self, z, remade_batch_tokens): - if not self.skip and self.aesthetic_steps != 0 and self.aesthetic_lr != 0 and self.aesthetic_weight != 0 and self.image_embs_name is not None: - tokenizer = shared.sd_model.cond_stage_model.tokenizer - if not opts.use_old_emphasis_implementation: - remade_batch_tokens = [ - [tokenizer.bos_token_id] + x[:75] + [tokenizer.eos_token_id] for x in - remade_batch_tokens] - - tokens = torch.asarray(remade_batch_tokens).to(device) - - model = copy.deepcopy(aesthetic_clip()).to(device) - model.requires_grad_(True) - if self.aesthetic_imgs_text is not None and len(self.aesthetic_imgs_text) > 0: - text_embs_2 = model.get_text_features( - **tokenizer([self.aesthetic_imgs_text], padding=True, return_tensors="pt").to(device)) - if self.aesthetic_text_negative: - text_embs_2 = self.image_embs - text_embs_2 - text_embs_2 /= text_embs_2.norm(dim=-1, keepdim=True) - img_embs = slerp(self.image_embs, text_embs_2, self.aesthetic_slerp_angle) - else: - img_embs = self.image_embs - - with torch.enable_grad(): - - # We optimize the model to maximize the similarity - optimizer = optim.Adam( - model.text_model.parameters(), lr=self.aesthetic_lr - ) - - for _ in trange(self.aesthetic_steps, desc="Aesthetic optimization"): - text_embs = model.get_text_features(input_ids=tokens) - text_embs = text_embs / text_embs.norm(dim=-1, keepdim=True) - sim = text_embs @ img_embs.T - loss = -sim - optimizer.zero_grad() - loss.mean().backward() - optimizer.step() - - zn = model.text_model(input_ids=tokens, output_hidden_states=-opts.CLIP_stop_at_last_layers) - if opts.CLIP_stop_at_last_layers > 1: - zn = zn.hidden_states[-opts.CLIP_stop_at_last_layers] - zn = model.text_model.final_layer_norm(zn) - else: - zn = zn.last_hidden_state - model.cpu() - del model - gc.collect() - torch.cuda.empty_cache() - zn = torch.concat([zn[77 * i:77 * (i + 1)] for i in range(max(z.shape[1] // 77, 1))], 1) - if self.slerp: - z = slerp(z, zn, self.aesthetic_weight) - else: - z = z * (1 - self.aesthetic_weight) + zn * self.aesthetic_weight - - return z diff --git a/modules/images_history.py b/modules/images_history.py index 78fd0543..bc5cf11f 100644 --- a/modules/images_history.py +++ b/modules/images_history.py @@ -310,7 +310,7 @@ def show_images_history(gr, opts, tabname, run_pnginfo, switch_dict): forward = gr.Button('Prev batch') backward = gr.Button('Next batch') with gr.Column(scale=3): - load_info = gr.HTML(visible=not custom_dir) + load_info = gr.HTML(visible=not custom_dir) with gr.Row(visible=False) as warning: warning_box = gr.Textbox("Message", interactive=False) diff --git a/modules/img2img.py b/modules/img2img.py index eea5199b..8d9f7cf9 100644 --- a/modules/img2img.py +++ b/modules/img2img.py @@ -56,7 +56,7 @@ def process_batch(p, input_dir, output_dir, args): processed_image.save(os.path.join(output_dir, filename)) -def img2img(mode: int, prompt: str, negative_prompt: str, prompt_style: str, prompt_style2: str, init_img, init_img_with_mask, init_img_inpaint, init_mask_inpaint, mask_mode, steps: int, sampler_index: int, mask_blur: int, inpainting_fill: int, restore_faces: bool, tiling: bool, n_iter: int, batch_size: int, cfg_scale: float, denoising_strength: float, seed: int, subseed: int, subseed_strength: float, seed_resize_from_h: int, seed_resize_from_w: int, seed_enable_extras: bool, height: int, width: int, resize_mode: int, inpaint_full_res: bool, inpaint_full_res_padding: int, inpainting_mask_invert: int, img2img_batch_input_dir: str, img2img_batch_output_dir: str, aesthetic_lr=0, aesthetic_weight=0, aesthetic_steps=0, aesthetic_imgs=None, aesthetic_slerp=False, aesthetic_imgs_text="", aesthetic_slerp_angle=0.15, aesthetic_text_negative=False, *args): +def img2img(mode: int, prompt: str, negative_prompt: str, prompt_style: str, prompt_style2: str, init_img, init_img_with_mask, init_img_inpaint, init_mask_inpaint, mask_mode, steps: int, sampler_index: int, mask_blur: int, inpainting_fill: int, restore_faces: bool, tiling: bool, n_iter: int, batch_size: int, cfg_scale: float, denoising_strength: float, seed: int, subseed: int, subseed_strength: float, seed_resize_from_h: int, seed_resize_from_w: int, seed_enable_extras: bool, height: int, width: int, resize_mode: int, inpaint_full_res: bool, inpaint_full_res_padding: int, inpainting_mask_invert: int, img2img_batch_input_dir: str, img2img_batch_output_dir: str, *args): is_inpaint = mode == 1 is_batch = mode == 2 @@ -109,7 +109,8 @@ def img2img(mode: int, prompt: str, negative_prompt: str, prompt_style: str, pro inpainting_mask_invert=inpainting_mask_invert, ) - shared.aesthetic_clip.set_aesthetic_params(p, float(aesthetic_lr), float(aesthetic_weight), int(aesthetic_steps), aesthetic_imgs, aesthetic_slerp, aesthetic_imgs_text, aesthetic_slerp_angle, aesthetic_text_negative) + p.scripts = modules.scripts.scripts_txt2img + p.script_args = args if shared.cmd_opts.enable_console_prompts: print(f"\nimg2img: {prompt}", file=shared.progress_print_out) diff --git a/modules/processing.py b/modules/processing.py index ff1ec4c9..372489f7 100644 --- a/modules/processing.py +++ b/modules/processing.py @@ -104,6 +104,12 @@ class StableDiffusionProcessing(): self.seed_resize_from_h = 0 self.seed_resize_from_w = 0 + self.scripts = None + self.script_args = None + self.all_prompts = None + self.all_seeds = None + self.all_subseeds = None + def init(self, all_prompts, all_seeds, all_subseeds): pass @@ -350,32 +356,35 @@ def process_images(p: StableDiffusionProcessing) -> Processed: shared.prompt_styles.apply_styles(p) if type(p.prompt) == list: - all_prompts = p.prompt + p.all_prompts = p.prompt else: - all_prompts = p.batch_size * p.n_iter * [p.prompt] + p.all_prompts = p.batch_size * p.n_iter * [p.prompt] if type(seed) == list: - all_seeds = seed + p.all_seeds = seed else: - all_seeds = [int(seed) + (x if p.subseed_strength == 0 else 0) for x in range(len(all_prompts))] + p.all_seeds = [int(seed) + (x if p.subseed_strength == 0 else 0) for x in range(len(p.all_prompts))] if type(subseed) == list: - all_subseeds = subseed + p.all_subseeds = subseed else: - all_subseeds = [int(subseed) + x for x in range(len(all_prompts))] + p.all_subseeds = [int(subseed) + x for x in range(len(p.all_prompts))] def infotext(iteration=0, position_in_batch=0): - return create_infotext(p, all_prompts, all_seeds, all_subseeds, comments, iteration, position_in_batch) + return create_infotext(p, p.all_prompts, p.all_seeds, p.all_subseeds, comments, iteration, position_in_batch) if os.path.exists(cmd_opts.embeddings_dir) and not p.do_not_reload_embeddings: model_hijack.embedding_db.load_textual_inversion_embeddings() + if p.scripts is not None: + p.scripts.run_alwayson_scripts(p) + infotexts = [] output_images = [] with torch.no_grad(), p.sd_model.ema_scope(): with devices.autocast(): - p.init(all_prompts, all_seeds, all_subseeds) + p.init(p.all_prompts, p.all_seeds, p.all_subseeds) if state.job_count == -1: state.job_count = p.n_iter @@ -387,9 +396,9 @@ def process_images(p: StableDiffusionProcessing) -> Processed: if state.interrupted: break - prompts = all_prompts[n * p.batch_size:(n + 1) * p.batch_size] - seeds = all_seeds[n * p.batch_size:(n + 1) * p.batch_size] - subseeds = all_subseeds[n * p.batch_size:(n + 1) * p.batch_size] + prompts = p.all_prompts[n * p.batch_size:(n + 1) * p.batch_size] + seeds = p.all_seeds[n * p.batch_size:(n + 1) * p.batch_size] + subseeds = p.all_subseeds[n * p.batch_size:(n + 1) * p.batch_size] if (len(prompts) == 0): break @@ -490,10 +499,10 @@ def process_images(p: StableDiffusionProcessing) -> Processed: index_of_first_image = 1 if opts.grid_save: - images.save_image(grid, p.outpath_grids, "grid", all_seeds[0], all_prompts[0], opts.grid_format, info=infotext(), short_filename=not opts.grid_extended_filename, p=p, grid=True) + images.save_image(grid, p.outpath_grids, "grid", p.all_seeds[0], p.all_prompts[0], opts.grid_format, info=infotext(), short_filename=not opts.grid_extended_filename, p=p, grid=True) devices.torch_gc() - return Processed(p, output_images, all_seeds[0], infotext() + "".join(["\n\n" + x for x in comments]), subseed=all_subseeds[0], all_prompts=all_prompts, all_seeds=all_seeds, all_subseeds=all_subseeds, index_of_first_image=index_of_first_image, infotexts=infotexts) + return Processed(p, output_images, p.all_seeds[0], infotext() + "".join(["\n\n" + x for x in comments]), subseed=p.all_subseeds[0], all_prompts=p.all_prompts, all_seeds=p.all_seeds, all_subseeds=p.all_subseeds, index_of_first_image=index_of_first_image, infotexts=infotexts) class StableDiffusionProcessingTxt2Img(StableDiffusionProcessing): diff --git a/modules/script_callbacks.py b/modules/script_callbacks.py new file mode 100644 index 00000000..866b7acd --- /dev/null +++ b/modules/script_callbacks.py @@ -0,0 +1,42 @@ + +callbacks_model_loaded = [] +callbacks_ui_tabs = [] + + +def clear_callbacks(): + callbacks_model_loaded.clear() + callbacks_ui_tabs.clear() + + +def model_loaded_callback(sd_model): + for callback in callbacks_model_loaded: + callback(sd_model) + + +def ui_tabs_callback(): + res = [] + + for callback in callbacks_ui_tabs: + res += callback() or [] + + return res + + +def on_model_loaded(callback): + """register a function to be called when the stable diffusion model is created; the model is + passed as an argument""" + callbacks_model_loaded.append(callback) + + +def on_ui_tabs(callback): + """register a function to be called when the UI is creating new tabs. + The function must either return a None, which means no new tabs to be added, or a list, where + each element is a tuple: + (gradio_component, title, elem_id) + + gradio_component is a gradio component to be used for contents of the tab (usually gr.Blocks) + title is tab text displayed to user in the UI + elem_id is HTML id for the tab + """ + callbacks_ui_tabs.append(callback) + diff --git a/modules/scripts.py b/modules/scripts.py index 1039fa9c..65f25f49 100644 --- a/modules/scripts.py +++ b/modules/scripts.py @@ -1,86 +1,153 @@ import os import sys import traceback +from collections import namedtuple import modules.ui as ui import gradio as gr from modules.processing import StableDiffusionProcessing -from modules import shared +from modules import shared, paths, script_callbacks + +AlwaysVisible = object() + class Script: filename = None args_from = None args_to = None + alwayson = False + + infotext_fields = None + """if set in ui(), this is a list of pairs of gradio component + text; the text will be used when + parsing infotext to set the value for the component; see ui.py's txt2img_paste_fields for an example + """ - # The title of the script. This is what will be displayed in the dropdown menu. def title(self): + """this function should return the title of the script. This is what will be displayed in the dropdown menu.""" + raise NotImplementedError() - # How the script is displayed in the UI. See https://gradio.app/docs/#components - # for the different UI components you can use and how to create them. - # Most UI components can return a value, such as a boolean for a checkbox. - # The returned values are passed to the run method as parameters. def ui(self, is_img2img): + """this function should create gradio UI elements. See https://gradio.app/docs/#components + The return value should be an array of all components that are used in processing. + Values of those returned componenbts will be passed to run() and process() functions. + """ + pass - # Determines when the script should be shown in the dropdown menu via the - # returned value. As an example: - # is_img2img is True if the current tab is img2img, and False if it is txt2img. - # Thus, return is_img2img to only show the script on the img2img tab. def show(self, is_img2img): + """ + is_img2img is True if this function is called for the img2img interface, and Fasle otherwise + + This function should return: + - False if the script should not be shown in UI at all + - True if the script should be shown in UI if it's scelected in the scripts drowpdown + - script.AlwaysVisible if the script should be shown in UI at all times + """ + return True - # This is where the additional processing is implemented. The parameters include - # self, the model object "p" (a StableDiffusionProcessing class, see - # processing.py), and the parameters returned by the ui method. - # Custom functions can be defined here, and additional libraries can be imported - # to be used in processing. The return value should be a Processed object, which is - # what is returned by the process_images method. - def run(self, *args): + def run(self, p, *args): + """ + This function is called if the script has been selected in the script dropdown. + It must do all processing and return the Processed object with results, same as + one returned by processing.process_images. + + Usually the processing is done by calling the processing.process_images function. + + args contains all values returned by components from ui() + """ + raise NotImplementedError() - # The description method is currently unused. - # To add a description that appears when hovering over the title, amend the "titles" - # dict in script.js to include the script title (returned by title) as a key, and - # your description as the value. + def process(self, p, *args): + """ + This function is called before processing begins for AlwaysVisible scripts. + scripts. You can modify the processing object (p) here, inject hooks, etc. + """ + + pass + def describe(self): + """unused""" return "" +current_basedir = paths.script_path + + +def basedir(): + """returns the base directory for the current script. For scripts in the main scripts directory, + this is the main directory (where webui.py resides), and for scripts in extensions directory + (ie extensions/aesthetic/script/aesthetic.py), this is extension's directory (extensions/aesthetic) + """ + return current_basedir + + scripts_data = [] +ScriptFile = namedtuple("ScriptFile", ["basedir", "filename", "path"]) +ScriptClassData = namedtuple("ScriptClassData", ["script_class", "path", "basedir"]) + + +def list_scripts(scriptdirname, extension): + scripts_list = [] + + basedir = os.path.join(paths.script_path, scriptdirname) + if os.path.exists(basedir): + for filename in sorted(os.listdir(basedir)): + scripts_list.append(ScriptFile(paths.script_path, filename, os.path.join(basedir, filename))) + + extdir = os.path.join(paths.script_path, "extensions") + if os.path.exists(extdir): + for dirname in sorted(os.listdir(extdir)): + dirpath = os.path.join(extdir, dirname) + if not os.path.isdir(dirpath): + continue + for filename in sorted(os.listdir(os.path.join(dirpath, scriptdirname))): + scripts_list.append(ScriptFile(dirpath, filename, os.path.join(dirpath, scriptdirname, filename))) -def load_scripts(basedir): - if not os.path.exists(basedir): - return + scripts_list = [x for x in scripts_list if os.path.splitext(x.path)[1].lower() == extension and os.path.isfile(x.path)] - for filename in sorted(os.listdir(basedir)): - path = os.path.join(basedir, filename) + return scripts_list - if os.path.splitext(path)[1].lower() != '.py': - continue - if not os.path.isfile(path): - continue +def load_scripts(): + global current_basedir + scripts_data.clear() + script_callbacks.clear_callbacks() + + scripts_list = list_scripts("scripts", ".py") + + syspath = sys.path + for scriptfile in sorted(scripts_list): try: - with open(path, "r", encoding="utf8") as file: + if scriptfile.basedir != paths.script_path: + sys.path = [scriptfile.basedir] + sys.path + current_basedir = scriptfile.basedir + + with open(scriptfile.path, "r", encoding="utf8") as file: text = file.read() from types import ModuleType - compiled = compile(text, path, 'exec') - module = ModuleType(filename) + compiled = compile(text, scriptfile.path, 'exec') + module = ModuleType(scriptfile.filename) exec(compiled, module.__dict__) for key, script_class in module.__dict__.items(): if type(script_class) == type and issubclass(script_class, Script): - scripts_data.append((script_class, path)) + scripts_data.append(ScriptClassData(script_class, scriptfile.path, scriptfile.basedir)) except Exception: - print(f"Error loading script: {filename}", file=sys.stderr) + print(f"Error loading script: {scriptfile.filename}", file=sys.stderr) print(traceback.format_exc(), file=sys.stderr) + finally: + sys.path = syspath + current_basedir = paths.script_path + def wrap_call(func, filename, funcname, *args, default=None, **kwargs): try: @@ -96,56 +163,80 @@ def wrap_call(func, filename, funcname, *args, default=None, **kwargs): class ScriptRunner: def __init__(self): self.scripts = [] + self.selectable_scripts = [] + self.alwayson_scripts = [] self.titles = [] + self.infotext_fields = [] def setup_ui(self, is_img2img): - for script_class, path in scripts_data: + for script_class, path, basedir in scripts_data: script = script_class() script.filename = path - if not script.show(is_img2img): - continue + visibility = script.show(is_img2img) - self.scripts.append(script) + if visibility == AlwaysVisible: + self.scripts.append(script) + self.alwayson_scripts.append(script) + script.alwayson = True - self.titles = [wrap_call(script.title, script.filename, "title") or f"{script.filename} [error]" for script in self.scripts] + elif visibility: + self.scripts.append(script) + self.selectable_scripts.append(script) - dropdown = gr.Dropdown(label="Script", choices=["None"] + self.titles, value="None", type="index") - dropdown.save_to_config = True - inputs = [dropdown] + self.titles = [wrap_call(script.title, script.filename, "title") or f"{script.filename} [error]" for script in self.selectable_scripts] + + inputs = [None] + inputs_alwayson = [True] - for script in self.scripts: + def create_script_ui(script, inputs, inputs_alwayson): script.args_from = len(inputs) script.args_to = len(inputs) controls = wrap_call(script.ui, script.filename, "ui", is_img2img) if controls is None: - continue + return for control in controls: control.custom_script_source = os.path.basename(script.filename) - control.visible = False + if not script.alwayson: + control.visible = False + + if script.infotext_fields is not None: + self.infotext_fields += script.infotext_fields inputs += controls + inputs_alwayson += [script.alwayson for _ in controls] script.args_to = len(inputs) + for script in self.alwayson_scripts: + with gr.Group(): + create_script_ui(script, inputs, inputs_alwayson) + + dropdown = gr.Dropdown(label="Script", choices=["None"] + self.titles, value="None", type="index") + dropdown.save_to_config = True + inputs[0] = dropdown + + for script in self.selectable_scripts: + create_script_ui(script, inputs, inputs_alwayson) + def select_script(script_index): - if 0 < script_index <= len(self.scripts): - script = self.scripts[script_index-1] + if 0 < script_index <= len(self.selectable_scripts): + script = self.selectable_scripts[script_index-1] args_from = script.args_from args_to = script.args_to else: args_from = 0 args_to = 0 - return [ui.gr_show(True if i == 0 else args_from <= i < args_to) for i in range(len(inputs))] + return [ui.gr_show(True if i == 0 else args_from <= i < args_to or is_alwayson) for i, is_alwayson in enumerate(inputs_alwayson)] def init_field(title): if title == 'None': return script_index = self.titles.index(title) - script = self.scripts[script_index] + script = self.selectable_scripts[script_index] for i in range(script.args_from, script.args_to): inputs[i].visible = True @@ -164,7 +255,7 @@ class ScriptRunner: if script_index == 0: return None - script = self.scripts[script_index-1] + script = self.selectable_scripts[script_index-1] if script is None: return None @@ -176,6 +267,15 @@ class ScriptRunner: return processed + def run_alwayson_scripts(self, p): + for script in self.alwayson_scripts: + try: + script_args = p.script_args[script.args_from:script.args_to] + script.process(p, *script_args) + except Exception: + print(f"Error running alwayson script: {script.filename}", file=sys.stderr) + print(traceback.format_exc(), file=sys.stderr) + def reload_sources(self): for si, script in list(enumerate(self.scripts)): with open(script.filename, "r", encoding="utf8") as file: @@ -197,19 +297,21 @@ class ScriptRunner: self.scripts[si].args_from = args_from self.scripts[si].args_to = args_to + scripts_txt2img = ScriptRunner() scripts_img2img = ScriptRunner() + def reload_script_body_only(): scripts_txt2img.reload_sources() scripts_img2img.reload_sources() -def reload_scripts(basedir): +def reload_scripts(): global scripts_txt2img, scripts_img2img - scripts_data.clear() - load_scripts(basedir) + load_scripts() scripts_txt2img = ScriptRunner() scripts_img2img = ScriptRunner() + diff --git a/modules/sd_hijack.py b/modules/sd_hijack.py index 1f8587d1..0f10828e 100644 --- a/modules/sd_hijack.py +++ b/modules/sd_hijack.py @@ -332,7 +332,6 @@ class FrozenCLIPEmbedderWithCustomWords(torch.nn.Module): multipliers.append([1.0] * 75) z1 = self.process_tokens(tokens, multipliers) - z1 = shared.aesthetic_clip(z1, remade_batch_tokens) z = z1 if z is None else torch.cat((z, z1), axis=-2) remade_batch_tokens = rem_tokens diff --git a/modules/sd_models.py b/modules/sd_models.py index d99dbce8..f9b3063d 100644 --- a/modules/sd_models.py +++ b/modules/sd_models.py @@ -7,7 +7,7 @@ from omegaconf import OmegaConf from ldm.util import instantiate_from_config -from modules import shared, modelloader, devices +from modules import shared, modelloader, devices, script_callbacks from modules.paths import models_path from modules.sd_hijack_inpainting import do_inpainting_hijack, should_hijack_inpainting @@ -238,6 +238,9 @@ def load_model(checkpoint_info=None): sd_hijack.model_hijack.hijack(sd_model) sd_model.eval() + shared.sd_model = sd_model + + script_callbacks.model_loaded_callback(sd_model) print(f"Model loaded.") return sd_model @@ -252,7 +255,7 @@ def reload_model_weights(sd_model, info=None): if sd_model.sd_checkpoint_info.config != checkpoint_info.config or should_hijack_inpainting(checkpoint_info) != should_hijack_inpainting(sd_model.sd_checkpoint_info): checkpoints_loaded.clear() - shared.sd_model = load_model(checkpoint_info) + load_model(checkpoint_info) return shared.sd_model if shared.cmd_opts.lowvram or shared.cmd_opts.medvram: diff --git a/modules/shared.py b/modules/shared.py index 0dbe360d..7d786f07 100644 --- a/modules/shared.py +++ b/modules/shared.py @@ -31,7 +31,6 @@ parser.add_argument("--no-half-vae", action='store_true', help="do not switch th parser.add_argument("--no-progressbar-hiding", action='store_true', help="do not hide progressbar in gradio UI (we hide it because it slows down ML if you have hardware acceleration in browser)") parser.add_argument("--max-batch-count", type=int, default=16, help="maximum batch count value for the UI") parser.add_argument("--embeddings-dir", type=str, default=os.path.join(script_path, 'embeddings'), help="embeddings directory for textual inversion (default: embeddings)") -parser.add_argument("--aesthetic_embeddings-dir", type=str, default=os.path.join(models_path, 'aesthetic_embeddings'), help="aesthetic_embeddings directory(default: aesthetic_embeddings)") parser.add_argument("--hypernetwork-dir", type=str, default=os.path.join(models_path, 'hypernetworks'), help="hypernetwork directory") parser.add_argument("--localizations-dir", type=str, default=os.path.join(script_path, 'localizations'), help="localizations directory") parser.add_argument("--allow-code", action='store_true', help="allow custom script execution from webui") @@ -109,21 +108,6 @@ os.makedirs(cmd_opts.hypernetwork_dir, exist_ok=True) hypernetworks = hypernetwork.list_hypernetworks(cmd_opts.hypernetwork_dir) loaded_hypernetwork = None - -os.makedirs(cmd_opts.aesthetic_embeddings_dir, exist_ok=True) -aesthetic_embeddings = {} - - -def update_aesthetic_embeddings(): - global aesthetic_embeddings - aesthetic_embeddings = {f.replace(".pt", ""): os.path.join(cmd_opts.aesthetic_embeddings_dir, f) for f in - os.listdir(cmd_opts.aesthetic_embeddings_dir) if f.endswith(".pt")} - aesthetic_embeddings = OrderedDict(**{"None": None}, **aesthetic_embeddings) - - -update_aesthetic_embeddings() - - def reload_hypernetworks(): global hypernetworks @@ -415,9 +399,6 @@ sd_model = None clip_model = None -from modules.aesthetic_clip import AestheticCLIP -aesthetic_clip = AestheticCLIP() - progress_print_out = sys.stdout diff --git a/modules/txt2img.py b/modules/txt2img.py index 1761cfa2..c9d5a090 100644 --- a/modules/txt2img.py +++ b/modules/txt2img.py @@ -7,7 +7,7 @@ import modules.processing as processing from modules.ui import plaintext_to_html -def txt2img(prompt: str, negative_prompt: str, prompt_style: str, prompt_style2: str, steps: int, sampler_index: int, restore_faces: bool, tiling: bool, n_iter: int, batch_size: int, cfg_scale: float, seed: int, subseed: int, subseed_strength: float, seed_resize_from_h: int, seed_resize_from_w: int, seed_enable_extras: bool, height: int, width: int, enable_hr: bool, denoising_strength: float, firstphase_width: int, firstphase_height: int, aesthetic_lr=0, aesthetic_weight=0, aesthetic_steps=0, aesthetic_imgs=None, aesthetic_slerp=False, aesthetic_imgs_text="", aesthetic_slerp_angle=0.15, aesthetic_text_negative=False, *args): +def txt2img(prompt: str, negative_prompt: str, prompt_style: str, prompt_style2: str, steps: int, sampler_index: int, restore_faces: bool, tiling: bool, n_iter: int, batch_size: int, cfg_scale: float, seed: int, subseed: int, subseed_strength: float, seed_resize_from_h: int, seed_resize_from_w: int, seed_enable_extras: bool, height: int, width: int, enable_hr: bool, denoising_strength: float, firstphase_width: int, firstphase_height: int, *args): p = StableDiffusionProcessingTxt2Img( sd_model=shared.sd_model, outpath_samples=opts.outdir_samples or opts.outdir_txt2img_samples, @@ -36,7 +36,8 @@ def txt2img(prompt: str, negative_prompt: str, prompt_style: str, prompt_style2: firstphase_height=firstphase_height if enable_hr else None, ) - shared.aesthetic_clip.set_aesthetic_params(p, float(aesthetic_lr), float(aesthetic_weight), int(aesthetic_steps), aesthetic_imgs, aesthetic_slerp, aesthetic_imgs_text, aesthetic_slerp_angle, aesthetic_text_negative) + p.scripts = modules.scripts.scripts_txt2img + p.script_args = args if cmd_opts.enable_console_prompts: print(f"\ntxt2img: {prompt}", file=shared.progress_print_out) diff --git a/modules/ui.py b/modules/ui.py index 70a9cf10..c977482c 100644 --- a/modules/ui.py +++ b/modules/ui.py @@ -23,10 +23,10 @@ import gradio as gr import gradio.utils import gradio.routes -from modules import sd_hijack, sd_models, localization +from modules import sd_hijack, sd_models, localization, script_callbacks from modules.paths import script_path -from modules.shared import opts, cmd_opts, restricted_opts, aesthetic_embeddings +from modules.shared import opts, cmd_opts, restricted_opts if cmd_opts.deepdanbooru: from modules.deepbooru import get_deepbooru_tags @@ -44,7 +44,6 @@ from modules.images import save_image import modules.textual_inversion.ui import modules.hypernetworks.ui -import modules.aesthetic_clip as aesthetic_clip import modules.images_history as img_his @@ -662,8 +661,6 @@ def create_ui(wrap_gradio_gpu_call): seed, reuse_seed, subseed, reuse_subseed, subseed_strength, seed_resize_from_h, seed_resize_from_w, seed_checkbox = create_seed_inputs() - aesthetic_weight, aesthetic_steps, aesthetic_lr, aesthetic_slerp, aesthetic_imgs, aesthetic_imgs_text, aesthetic_slerp_angle, aesthetic_text_negative = aesthetic_clip.create_ui() - with gr.Group(): custom_inputs = modules.scripts.scripts_txt2img.setup_ui(is_img2img=False) @@ -718,14 +715,6 @@ def create_ui(wrap_gradio_gpu_call): denoising_strength, firstphase_width, firstphase_height, - aesthetic_lr, - aesthetic_weight, - aesthetic_steps, - aesthetic_imgs, - aesthetic_slerp, - aesthetic_imgs_text, - aesthetic_slerp_angle, - aesthetic_text_negative ] + custom_inputs, outputs=[ @@ -804,14 +793,7 @@ def create_ui(wrap_gradio_gpu_call): (hr_options, lambda d: gr.Row.update(visible="Denoising strength" in d)), (firstphase_width, "First pass size-1"), (firstphase_height, "First pass size-2"), - (aesthetic_lr, "Aesthetic LR"), - (aesthetic_weight, "Aesthetic weight"), - (aesthetic_steps, "Aesthetic steps"), - (aesthetic_imgs, "Aesthetic embedding"), - (aesthetic_slerp, "Aesthetic slerp"), - (aesthetic_imgs_text, "Aesthetic text"), - (aesthetic_text_negative, "Aesthetic text negative"), - (aesthetic_slerp_angle, "Aesthetic slerp angle"), + *modules.scripts.scripts_txt2img.infotext_fields ] txt2img_preview_params = [ @@ -896,8 +878,6 @@ def create_ui(wrap_gradio_gpu_call): seed, reuse_seed, subseed, reuse_subseed, subseed_strength, seed_resize_from_h, seed_resize_from_w, seed_checkbox = create_seed_inputs() - aesthetic_weight_im, aesthetic_steps_im, aesthetic_lr_im, aesthetic_slerp_im, aesthetic_imgs_im, aesthetic_imgs_text_im, aesthetic_slerp_angle_im, aesthetic_text_negative_im = aesthetic_clip.create_ui() - with gr.Group(): custom_inputs = modules.scripts.scripts_img2img.setup_ui(is_img2img=True) @@ -988,14 +968,6 @@ def create_ui(wrap_gradio_gpu_call): inpainting_mask_invert, img2img_batch_input_dir, img2img_batch_output_dir, - aesthetic_lr_im, - aesthetic_weight_im, - aesthetic_steps_im, - aesthetic_imgs_im, - aesthetic_slerp_im, - aesthetic_imgs_text_im, - aesthetic_slerp_angle_im, - aesthetic_text_negative_im, ] + custom_inputs, outputs=[ img2img_gallery, @@ -1087,14 +1059,7 @@ def create_ui(wrap_gradio_gpu_call): (seed_resize_from_w, "Seed resize from-1"), (seed_resize_from_h, "Seed resize from-2"), (denoising_strength, "Denoising strength"), - (aesthetic_lr_im, "Aesthetic LR"), - (aesthetic_weight_im, "Aesthetic weight"), - (aesthetic_steps_im, "Aesthetic steps"), - (aesthetic_imgs_im, "Aesthetic embedding"), - (aesthetic_slerp_im, "Aesthetic slerp"), - (aesthetic_imgs_text_im, "Aesthetic text"), - (aesthetic_text_negative_im, "Aesthetic text negative"), - (aesthetic_slerp_angle_im, "Aesthetic slerp angle"), + *modules.scripts.scripts_img2img.infotext_fields ] token_button.click(fn=update_token_counter, inputs=[img2img_prompt, steps], outputs=[token_counter]) @@ -1217,9 +1182,9 @@ def create_ui(wrap_gradio_gpu_call): ) #images history images_history_switch_dict = { - "fn":modules.generation_parameters_copypaste.connect_paste, - "t2i":txt2img_paste_fields, - "i2i":img2img_paste_fields + "fn": modules.generation_parameters_copypaste.connect_paste, + "t2i": txt2img_paste_fields, + "i2i": img2img_paste_fields } images_history = img_his.create_history_tabs(gr, opts, cmd_opts, wrap_gradio_call(modules.extras.run_pnginfo), images_history_switch_dict) @@ -1264,18 +1229,6 @@ def create_ui(wrap_gradio_gpu_call): with gr.Column(): create_embedding = gr.Button(value="Create embedding", variant='primary') - with gr.Tab(label="Create aesthetic images embedding"): - - new_embedding_name_ae = gr.Textbox(label="Name") - process_src_ae = gr.Textbox(label='Source directory') - batch_ae = gr.Slider(minimum=1, maximum=1024, step=1, label="Batch size", value=256) - with gr.Row(): - with gr.Column(scale=3): - gr.HTML(value="") - - with gr.Column(): - create_embedding_ae = gr.Button(value="Create images embedding", variant='primary') - with gr.Tab(label="Create hypernetwork"): new_hypernetwork_name = gr.Textbox(label="Name") new_hypernetwork_sizes = gr.CheckboxGroup(label="Modules", value=["768", "320", "640", "1280"], choices=["768", "320", "640", "1280"]) @@ -1375,21 +1328,6 @@ def create_ui(wrap_gradio_gpu_call): ] ) - create_embedding_ae.click( - fn=aesthetic_clip.generate_imgs_embd, - inputs=[ - new_embedding_name_ae, - process_src_ae, - batch_ae - ], - outputs=[ - aesthetic_imgs, - aesthetic_imgs_im, - ti_output, - ti_outcome, - ] - ) - create_hypernetwork.click( fn=modules.hypernetworks.ui.create_hypernetwork, inputs=[ @@ -1580,10 +1518,10 @@ Requested path was: {f} if not opts.same_type(value, opts.data_labels[key].default): return gr.update(visible=True), opts.dumpjson() + oldval = opts.data.get(key, None) if cmd_opts.hide_ui_dir_config and key in restricted_opts: return gr.update(value=oldval), opts.dumpjson() - oldval = opts.data.get(key, None) opts.data[key] = value if oldval != value: @@ -1692,9 +1630,12 @@ Requested path was: {f} (images_history, "Image Browser", "images_history"), (modelmerger_interface, "Checkpoint Merger", "modelmerger"), (train_interface, "Train", "ti"), - (settings_interface, "Settings", "settings"), ] + interfaces += script_callbacks.ui_tabs_callback() + + interfaces += [(settings_interface, "Settings", "settings")] + with open(os.path.join(script_path, "style.css"), "r", encoding="utf8") as file: css = file.read() diff --git a/webui.py b/webui.py index 87589064..b1deca1b 100644 --- a/webui.py +++ b/webui.py @@ -71,6 +71,7 @@ def wrap_gradio_gpu_call(func, extra_outputs=None): return modules.ui.wrap_gradio_call(f, extra_outputs=extra_outputs) + def initialize(): modelloader.cleanup_models() modules.sd_models.setup_model() @@ -79,9 +80,9 @@ def initialize(): shared.face_restorers.append(modules.face_restoration.FaceRestoration()) modelloader.load_upscalers() - modules.scripts.load_scripts(os.path.join(script_path, "scripts")) + modules.scripts.load_scripts() - shared.sd_model = modules.sd_models.load_model() + modules.sd_models.load_model() shared.opts.onchange("sd_model_checkpoint", wrap_queued_call(lambda: modules.sd_models.reload_model_weights(shared.sd_model))) shared.opts.onchange("sd_hypernetwork", wrap_queued_call(lambda: modules.hypernetworks.hypernetwork.load_hypernetwork(shared.opts.sd_hypernetwork))) shared.opts.onchange("sd_hypernetwork_strength", modules.hypernetworks.hypernetwork.apply_strength) @@ -145,7 +146,7 @@ def webui(): sd_samplers.set_samplers() print('Reloading Custom Scripts') - modules.scripts.reload_scripts(os.path.join(script_path, "scripts")) + modules.scripts.reload_scripts() print('Reloading modules: modules.ui') importlib.reload(modules.ui) print('Refreshing Model List') -- cgit v1.2.3 From 0df94d3fcf9d1fc47c4d39039352a3d5b3380c1f Mon Sep 17 00:00:00 2001 From: MrCheeze Date: Sat, 22 Oct 2022 12:59:21 -0400 Subject: fix aesthetic gradients doing nothing after loading a different model --- modules/sd_models.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) (limited to 'modules/sd_models.py') diff --git a/modules/sd_models.py b/modules/sd_models.py index f9b3063d..49dc3238 100644 --- a/modules/sd_models.py +++ b/modules/sd_models.py @@ -236,12 +236,11 @@ def load_model(checkpoint_info=None): sd_model.to(shared.device) sd_hijack.model_hijack.hijack(sd_model) + script_callbacks.model_loaded_callback(sd_model) sd_model.eval() shared.sd_model = sd_model - script_callbacks.model_loaded_callback(sd_model) - print(f"Model loaded.") return sd_model @@ -268,6 +267,7 @@ def reload_model_weights(sd_model, info=None): load_model_weights(sd_model, checkpoint_info) sd_hijack.model_hijack.hijack(sd_model) + script_callbacks.model_loaded_callback(sd_model) if not shared.cmd_opts.lowvram and not shared.cmd_opts.medvram: sd_model.to(devices.device) -- cgit v1.2.3 From 321bacc6a9eaf4a25f31279f288fa752be507a20 Mon Sep 17 00:00:00 2001 From: AUTOMATIC <16777216c@gmail.com> Date: Sat, 22 Oct 2022 20:15:12 +0300 Subject: call model_loaded_callback after setting shared.sd_model in case scripts refer to it using that --- modules/sd_models.py | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) (limited to 'modules/sd_models.py') diff --git a/modules/sd_models.py b/modules/sd_models.py index 49dc3238..e697bb72 100644 --- a/modules/sd_models.py +++ b/modules/sd_models.py @@ -236,11 +236,12 @@ def load_model(checkpoint_info=None): sd_model.to(shared.device) sd_hijack.model_hijack.hijack(sd_model) - script_callbacks.model_loaded_callback(sd_model) sd_model.eval() shared.sd_model = sd_model + script_callbacks.model_loaded_callback(sd_model) + print(f"Model loaded.") return sd_model -- cgit v1.2.3 From b50ff4f4e4d4d6bf31e222832d3fe4cfde4703c9 Mon Sep 17 00:00:00 2001 From: Josh Watzman Date: Thu, 27 Oct 2022 21:59:16 +0100 Subject: Reduce peak memory usage when changing models A few tweaks to reduce peak memory usage, the biggest being that if we aren't using the checkpoint cache, we shouldn't duplicate the model state dict just to immediately throw it away. On my machine with 16GB of RAM, this change means I can typically change models, whereas before it would typically OOM. --- modules/sd_models.py | 11 +++++++---- 1 file changed, 7 insertions(+), 4 deletions(-) (limited to 'modules/sd_models.py') diff --git a/modules/sd_models.py b/modules/sd_models.py index e697bb72..203e99a8 100644 --- a/modules/sd_models.py +++ b/modules/sd_models.py @@ -170,7 +170,9 @@ def load_model_weights(model, checkpoint_info): print(f"Global Step: {pl_sd['global_step']}") sd = get_state_dict_from_checkpoint(pl_sd) - missing, extra = model.load_state_dict(sd, strict=False) + del pl_sd + model.load_state_dict(sd, strict=False) + del sd if shared.cmd_opts.opt_channelslast: model.to(memory_format=torch.channels_last) @@ -194,9 +196,10 @@ def load_model_weights(model, checkpoint_info): model.first_stage_model.to(devices.dtype_vae) - checkpoints_loaded[checkpoint_info] = model.state_dict().copy() - while len(checkpoints_loaded) > shared.opts.sd_checkpoint_cache: - checkpoints_loaded.popitem(last=False) # LRU + if shared.opts.sd_checkpoint_cache > 0: + checkpoints_loaded[checkpoint_info] = model.state_dict().copy() + while len(checkpoints_loaded) > shared.opts.sd_checkpoint_cache: + checkpoints_loaded.popitem(last=False) # LRU else: print(f"Loading weights [{sd_model_hash}] from cache") checkpoints_loaded.move_to_end(checkpoint_info) -- cgit v1.2.3 From 5d5dc64064d8ca399a76fe44dbb62bdef6c4b7c4 Mon Sep 17 00:00:00 2001 From: Antonio Date: Fri, 28 Oct 2022 05:49:39 +0200 Subject: Natural sorting for dropdown checkpoint list Example: Before After 11.ckpt 11.ckpt ab.ckpt ab.ckpt ade_pablo_step_1000.ckpt ade_pablo_step_500.ckpt ade_pablo_step_500.ckpt ade_pablo_step_1000.ckpt ade_step_1000.ckpt ade_step_500.ckpt ade_step_1500.ckpt ade_step_1000.ckpt ade_step_2000.ckpt ade_step_1500.ckpt ade_step_2500.ckpt ade_step_2000.ckpt ade_step_3000.ckpt ade_step_2500.ckpt ade_step_500.ckpt ade_step_3000.ckpt atp_step_5500.ckpt atp_step_5500.ckpt model1.ckpt model1.ckpt model10.ckpt model10.ckpt model1000.ckpt model33.ckpt model33.ckpt model50.ckpt model400.ckpt model400.ckpt model50.ckpt model1000.ckpt moo44.ckpt moo44.ckpt v1-4-pruned-emaonly.ckpt v1-4-pruned-emaonly.ckpt v1-5-pruned-emaonly.ckpt v1-5-pruned-emaonly.ckpt v1-5-pruned.ckpt v1-5-pruned.ckpt v1-5-vae.ckpt v1-5-vae.ckpt --- modules/sd_models.py | 7 +++++-- 1 file changed, 5 insertions(+), 2 deletions(-) (limited to 'modules/sd_models.py') diff --git a/modules/sd_models.py b/modules/sd_models.py index e697bb72..64d5ee0d 100644 --- a/modules/sd_models.py +++ b/modules/sd_models.py @@ -3,6 +3,7 @@ import os.path import sys from collections import namedtuple import torch +import re from omegaconf import OmegaConf from ldm.util import instantiate_from_config @@ -35,8 +36,10 @@ def setup_model(): list_models() -def checkpoint_tiles(): - return sorted([x.title for x in checkpoints_list.values()]) +def checkpoint_tiles(): + convert = lambda name: int(name) if name.isdigit() else name.lower() + alphanumeric_key = lambda key: [convert(c) for c in re.split('([0-9]+)', key)] + return sorted([x.title for x in checkpoints_list.values()], key = alphanumeric_key) def list_models(): -- cgit v1.2.3 From cb31abcf58ea1f64266e6d821937eed058c35f4d Mon Sep 17 00:00:00 2001 From: Muhammad Rizqi Nur Date: Sun, 30 Oct 2022 21:54:31 +0700 Subject: Settings to select VAE --- modules/sd_models.py | 31 +++++-------- modules/sd_vae.py | 121 +++++++++++++++++++++++++++++++++++++++++++++++++++ modules/shared.py | 8 ++-- webui.py | 5 +++ 4 files changed, 141 insertions(+), 24 deletions(-) create mode 100644 modules/sd_vae.py (limited to 'modules/sd_models.py') diff --git a/modules/sd_models.py b/modules/sd_models.py index f86dc3ed..91ad4b5e 100644 --- a/modules/sd_models.py +++ b/modules/sd_models.py @@ -8,7 +8,7 @@ from omegaconf import OmegaConf from ldm.util import instantiate_from_config -from modules import shared, modelloader, devices, script_callbacks +from modules import shared, modelloader, devices, script_callbacks, sd_vae from modules.paths import models_path from modules.sd_hijack_inpainting import do_inpainting_hijack, should_hijack_inpainting @@ -160,12 +160,11 @@ def get_state_dict_from_checkpoint(pl_sd): vae_ignore_keys = {"model_ema.decay", "model_ema.num_updates"} - -def load_model_weights(model, checkpoint_info): +def load_model_weights(model, checkpoint_info, force=False): checkpoint_file = checkpoint_info.filename sd_model_hash = checkpoint_info.hash - if checkpoint_info not in checkpoints_loaded: + if force or checkpoint_info not in checkpoints_loaded: print(f"Loading weights [{sd_model_hash}] from {checkpoint_file}") pl_sd = torch.load(checkpoint_file, map_location=shared.weight_load_location) @@ -186,17 +185,7 @@ def load_model_weights(model, checkpoint_info): devices.dtype = torch.float32 if shared.cmd_opts.no_half else torch.float16 devices.dtype_vae = torch.float32 if shared.cmd_opts.no_half or shared.cmd_opts.no_half_vae else torch.float16 - vae_file = os.path.splitext(checkpoint_file)[0] + ".vae.pt" - - if not os.path.exists(vae_file) and shared.cmd_opts.vae_path is not None: - vae_file = shared.cmd_opts.vae_path - - if os.path.exists(vae_file): - print(f"Loading VAE weights from: {vae_file}") - vae_ckpt = torch.load(vae_file, map_location=shared.weight_load_location) - vae_dict = {k: v for k, v in vae_ckpt["state_dict"].items() if k[0:4] != "loss" and k not in vae_ignore_keys} - model.first_stage_model.load_state_dict(vae_dict) - + sd_vae.load_vae(model, checkpoint_file) model.first_stage_model.to(devices.dtype_vae) if shared.opts.sd_checkpoint_cache > 0: @@ -213,7 +202,7 @@ def load_model_weights(model, checkpoint_info): model.sd_checkpoint_info = checkpoint_info -def load_model(checkpoint_info=None): +def load_model(checkpoint_info=None, force=False): from modules import lowvram, sd_hijack checkpoint_info = checkpoint_info or select_checkpoint() @@ -234,7 +223,7 @@ def load_model(checkpoint_info=None): do_inpainting_hijack() sd_model = instantiate_from_config(sd_config.model) - load_model_weights(sd_model, checkpoint_info) + load_model_weights(sd_model, checkpoint_info, force=force) if shared.cmd_opts.lowvram or shared.cmd_opts.medvram: lowvram.setup_for_low_vram(sd_model, shared.cmd_opts.medvram) @@ -252,16 +241,16 @@ def load_model(checkpoint_info=None): return sd_model -def reload_model_weights(sd_model, info=None): +def reload_model_weights(sd_model, info=None, force=False): from modules import lowvram, devices, sd_hijack checkpoint_info = info or select_checkpoint() - if sd_model.sd_model_checkpoint == checkpoint_info.filename: + if sd_model.sd_model_checkpoint == checkpoint_info.filename and not force: return if sd_model.sd_checkpoint_info.config != checkpoint_info.config or should_hijack_inpainting(checkpoint_info) != should_hijack_inpainting(sd_model.sd_checkpoint_info): checkpoints_loaded.clear() - load_model(checkpoint_info) + load_model(checkpoint_info, force=force) return shared.sd_model if shared.cmd_opts.lowvram or shared.cmd_opts.medvram: @@ -271,7 +260,7 @@ def reload_model_weights(sd_model, info=None): sd_hijack.model_hijack.undo_hijack(sd_model) - load_model_weights(sd_model, checkpoint_info) + load_model_weights(sd_model, checkpoint_info, force=force) sd_hijack.model_hijack.hijack(sd_model) script_callbacks.model_loaded_callback(sd_model) diff --git a/modules/sd_vae.py b/modules/sd_vae.py new file mode 100644 index 00000000..82764e55 --- /dev/null +++ b/modules/sd_vae.py @@ -0,0 +1,121 @@ +import torch +import os +from collections import namedtuple +from modules import shared, devices +from modules.paths import models_path +import glob + +model_dir = "Stable-diffusion" +model_path = os.path.abspath(os.path.join(models_path, model_dir)) +vae_dir = "VAE" +vae_path = os.path.abspath(os.path.join(models_path, vae_dir)) + +vae_ignore_keys = {"model_ema.decay", "model_ema.num_updates"} +default_vae_dict = {"auto": "auto", "None": "None"} +default_vae_list = ["auto", "None"] +default_vae_values = [default_vae_dict[x] for x in default_vae_list] +vae_dict = dict(default_vae_dict) +vae_list = list(default_vae_list) +first_load = True + +def get_filename(filepath): + return os.path.splitext(os.path.basename(filepath))[0] + +def refresh_vae_list(vae_path=vae_path, model_path=model_path): + global vae_dict, vae_list + res = {} + candidates = [ + *glob.iglob(os.path.join(model_path, '**/*.vae.pt'), recursive=True), + *glob.iglob(os.path.join(model_path, '**/*.vae.ckpt'), recursive=True), + *glob.iglob(os.path.join(vae_path, '**/*.pt'), recursive=True), + *glob.iglob(os.path.join(vae_path, '**/*.ckpt'), recursive=True) + ] + if shared.cmd_opts.vae_path is not None and os.path.isfile(shared.cmd_opts.vae_path): + candidates.append(shared.cmd_opts.vae_path) + for filepath in candidates: + name = get_filename(filepath) + res[name] = filepath + vae_list.clear() + vae_list.extend(default_vae_list) + vae_list.extend(list(res.keys())) + vae_dict.clear() + vae_dict.update(default_vae_dict) + vae_dict.update(res) + return vae_list + +def load_vae(model, checkpoint_file, vae_file="auto"): + global first_load, vae_dict, vae_list + # save_settings = False + + # if vae_file argument is provided, it takes priority + if vae_file and vae_file not in default_vae_list: + if not os.path.isfile(vae_file): + vae_file = "auto" + # save_settings = True + print("VAE provided as function argument doesn't exist") + # for the first load, if vae-path is provided, it takes priority and failure is reported + if first_load and shared.cmd_opts.vae_path is not None: + if os.path.isfile(shared.cmd_opts.vae_path): + vae_file = shared.cmd_opts.vae_path + # save_settings = True + # print("Using VAE provided as command line argument") + else: + print("VAE provided as command line argument doesn't exist") + # else, we load from settings + if vae_file == "auto" and shared.opts.sd_vae is not None: + # if saved VAE settings isn't recognized, fallback to auto + vae_file = vae_dict.get(shared.opts.sd_vae, "auto") + # if VAE selected but not found, fallback to auto + if vae_file not in default_vae_values and not os.path.isfile(vae_file): + vae_file = "auto" + print("Selected VAE doesn't exist") + # vae-path cmd arg takes priority for auto + if vae_file == "auto" and shared.cmd_opts.vae_path is not None: + if os.path.isfile(shared.cmd_opts.vae_path): + vae_file = shared.cmd_opts.vae_path + print("Using VAE provided as command line argument") + # if still not found, try look for ".vae.pt" beside model + model_path = os.path.splitext(checkpoint_file)[0] + if vae_file == "auto": + vae_file_try = model_path + ".vae.pt" + if os.path.isfile(vae_file_try): + vae_file = vae_file_try + print("Using VAE found beside selected model") + # if still not found, try look for ".vae.ckpt" beside model + if vae_file == "auto": + vae_file_try = model_path + ".vae.ckpt" + if os.path.isfile(vae_file_try): + vae_file = vae_file_try + print("Using VAE found beside selected model") + # No more fallbacks for auto + if vae_file == "auto": + vae_file = None + # Last check, just because + if vae_file and not os.path.exists(vae_file): + vae_file = None + + if vae_file: + print(f"Loading VAE weights from: {vae_file}") + vae_ckpt = torch.load(vae_file, map_location=shared.weight_load_location) + vae_dict_1 = {k: v for k, v in vae_ckpt["state_dict"].items() if k[0:4] != "loss" and k not in vae_ignore_keys} + model.first_stage_model.load_state_dict(vae_dict_1) + + # If vae used is not in dict, update it + # It will be removed on refresh though + if vae_file is not None: + vae_opt = get_filename(vae_file) + if vae_opt not in vae_dict: + vae_dict[vae_opt] = vae_file + vae_list.append(vae_opt) + + """ + # Save current VAE to VAE settings, maybe? will it work? + if save_settings: + if vae_file is None: + vae_opt = "None" + + # shared.opts.sd_vae = vae_opt + """ + + first_load = False + model.first_stage_model.to(devices.dtype_vae) diff --git a/modules/shared.py b/modules/shared.py index e4f163c1..06440ac4 100644 --- a/modules/shared.py +++ b/modules/shared.py @@ -14,7 +14,7 @@ import modules.memmon import modules.sd_models import modules.styles import modules.devices as devices -from modules import sd_samplers, sd_models, localization +from modules import sd_samplers, sd_models, localization, sd_vae from modules.hypernetworks import hypernetwork from modules.paths import models_path, script_path, sd_path @@ -295,6 +295,7 @@ options_templates.update(options_section(('training', "Training"), { options_templates.update(options_section(('sd', "Stable Diffusion"), { "sd_model_checkpoint": OptionInfo(None, "Stable Diffusion checkpoint", gr.Dropdown, lambda: {"choices": modules.sd_models.checkpoint_tiles()}, refresh=sd_models.list_models), "sd_checkpoint_cache": OptionInfo(0, "Checkpoints to cache in RAM", gr.Slider, {"minimum": 0, "maximum": 10, "step": 1}), + "sd_vae": OptionInfo("auto", "SD VAE", gr.Dropdown, lambda: {"choices": list(sd_vae.vae_list)}, refresh=sd_vae.refresh_vae_list), "sd_hypernetwork": OptionInfo("None", "Hypernetwork", gr.Dropdown, lambda: {"choices": ["None"] + [x for x in hypernetworks.keys()]}, refresh=reload_hypernetworks), "sd_hypernetwork_strength": OptionInfo(1.0, "Hypernetwork strength", gr.Slider, {"minimum": 0.0, "maximum": 1.0, "step": 0.001}), "inpainting_mask_weight": OptionInfo(1.0, "Inpainting conditioning mask strength", gr.Slider, {"minimum": 0.0, "maximum": 1.0, "step": 0.01}), @@ -407,11 +408,12 @@ class Options: if bad_settings > 0: print(f"The program is likely to not work with bad settings.\nSettings file: {filename}\nEither fix the file, or delete it and restart.", file=sys.stderr) - def onchange(self, key, func): + def onchange(self, key, func, call=True): item = self.data_labels.get(key) item.onchange = func - func() + if call: + func() def dumpjson(self): d = {k: self.data.get(k, self.data_labels.get(k).default) for k in self.data_labels.keys()} diff --git a/webui.py b/webui.py index 29530872..27949f3d 100644 --- a/webui.py +++ b/webui.py @@ -21,6 +21,7 @@ import modules.paths import modules.scripts import modules.sd_hijack import modules.sd_models +import modules.sd_vae import modules.shared as shared import modules.txt2img @@ -74,8 +75,12 @@ def initialize(): modules.scripts.load_scripts() + modules.sd_vae.refresh_vae_list() modules.sd_models.load_model() shared.opts.onchange("sd_model_checkpoint", wrap_queued_call(lambda: modules.sd_models.reload_model_weights(shared.sd_model))) + # I don't know what needs to be done to only reload VAE, with all those hijacks callbacks, and lowvram, + # so for now this reloads the whole model too, and no cache + shared.opts.onchange("sd_vae", wrap_queued_call(lambda: modules.sd_models.reload_model_weights(shared.sd_model, force=True)), call=False) shared.opts.onchange("sd_hypernetwork", wrap_queued_call(lambda: modules.hypernetworks.hypernetwork.load_hypernetwork(shared.opts.sd_hypernetwork))) shared.opts.onchange("sd_hypernetwork_strength", modules.hypernetworks.hypernetwork.apply_strength) -- cgit v1.2.3 From 726769da35970f4c100fa7edf11850f9dc059c41 Mon Sep 17 00:00:00 2001 From: Muhammad Rizqi Nur Date: Mon, 31 Oct 2022 15:19:34 +0700 Subject: Checkpoint cache by combination key of checkpoint and vae --- modules/sd_models.py | 27 ++++++++++++++++----------- modules/sd_vae.py | 8 +++++++- 2 files changed, 23 insertions(+), 12 deletions(-) (limited to 'modules/sd_models.py') diff --git a/modules/sd_models.py b/modules/sd_models.py index 91ad4b5e..850f7b7b 100644 --- a/modules/sd_models.py +++ b/modules/sd_models.py @@ -160,11 +160,15 @@ def get_state_dict_from_checkpoint(pl_sd): vae_ignore_keys = {"model_ema.decay", "model_ema.num_updates"} -def load_model_weights(model, checkpoint_info, force=False): +def load_model_weights(model, checkpoint_info, vae_file="auto"): checkpoint_file = checkpoint_info.filename sd_model_hash = checkpoint_info.hash - if force or checkpoint_info not in checkpoints_loaded: + vae_file = sd_vae.resolve_vae(checkpoint_file, vae_file=vae_file) + + checkpoint_key = (checkpoint_info, vae_file) + + if checkpoint_key not in checkpoints_loaded: print(f"Loading weights [{sd_model_hash}] from {checkpoint_file}") pl_sd = torch.load(checkpoint_file, map_location=shared.weight_load_location) @@ -185,24 +189,25 @@ def load_model_weights(model, checkpoint_info, force=False): devices.dtype = torch.float32 if shared.cmd_opts.no_half else torch.float16 devices.dtype_vae = torch.float32 if shared.cmd_opts.no_half or shared.cmd_opts.no_half_vae else torch.float16 - sd_vae.load_vae(model, checkpoint_file) + sd_vae.load_vae(model, vae_file) model.first_stage_model.to(devices.dtype_vae) if shared.opts.sd_checkpoint_cache > 0: - checkpoints_loaded[checkpoint_info] = model.state_dict().copy() + checkpoints_loaded[checkpoint_key] = model.state_dict().copy() while len(checkpoints_loaded) > shared.opts.sd_checkpoint_cache: checkpoints_loaded.popitem(last=False) # LRU else: - print(f"Loading weights [{sd_model_hash}] from cache") - checkpoints_loaded.move_to_end(checkpoint_info) - model.load_state_dict(checkpoints_loaded[checkpoint_info]) + vae_name = sd_vae.get_filename(vae_file) + print(f"Loading weights [{sd_model_hash}] with {vae_name} VAE from cache") + checkpoints_loaded.move_to_end(checkpoint_key) + model.load_state_dict(checkpoints_loaded[checkpoint_key]) model.sd_model_hash = sd_model_hash model.sd_model_checkpoint = checkpoint_file model.sd_checkpoint_info = checkpoint_info -def load_model(checkpoint_info=None, force=False): +def load_model(checkpoint_info=None): from modules import lowvram, sd_hijack checkpoint_info = checkpoint_info or select_checkpoint() @@ -223,7 +228,7 @@ def load_model(checkpoint_info=None, force=False): do_inpainting_hijack() sd_model = instantiate_from_config(sd_config.model) - load_model_weights(sd_model, checkpoint_info, force=force) + load_model_weights(sd_model, checkpoint_info) if shared.cmd_opts.lowvram or shared.cmd_opts.medvram: lowvram.setup_for_low_vram(sd_model, shared.cmd_opts.medvram) @@ -250,7 +255,7 @@ def reload_model_weights(sd_model, info=None, force=False): if sd_model.sd_checkpoint_info.config != checkpoint_info.config or should_hijack_inpainting(checkpoint_info) != should_hijack_inpainting(sd_model.sd_checkpoint_info): checkpoints_loaded.clear() - load_model(checkpoint_info, force=force) + load_model(checkpoint_info) return shared.sd_model if shared.cmd_opts.lowvram or shared.cmd_opts.medvram: @@ -260,7 +265,7 @@ def reload_model_weights(sd_model, info=None, force=False): sd_hijack.model_hijack.undo_hijack(sd_model) - load_model_weights(sd_model, checkpoint_info, force=force) + load_model_weights(sd_model, checkpoint_info) sd_hijack.model_hijack.hijack(sd_model) script_callbacks.model_loaded_callback(sd_model) diff --git a/modules/sd_vae.py b/modules/sd_vae.py index 2ce44d5f..e9239326 100644 --- a/modules/sd_vae.py +++ b/modules/sd_vae.py @@ -43,7 +43,7 @@ def refresh_vae_list(vae_path=vae_path, model_path=model_path): vae_dict.update(res) return vae_list -def load_vae(model, checkpoint_file, vae_file="auto"): +def resolve_vae(checkpoint_file, vae_file="auto"): global first_load, vae_dict, vae_list # save_settings = False @@ -94,6 +94,12 @@ def load_vae(model, checkpoint_file, vae_file="auto"): if vae_file and not os.path.exists(vae_file): vae_file = None + return vae_file + +def load_vae(model, vae_file): + global first_load, vae_dict, vae_list + # save_settings = False + if vae_file: print(f"Loading VAE weights from: {vae_file}") vae_ckpt = torch.load(vae_file, map_location=shared.weight_load_location) -- cgit v1.2.3 From 36966e3200943dbf890b5338cfa939df552d3c47 Mon Sep 17 00:00:00 2001 From: Muhammad Rizqi Nur Date: Mon, 31 Oct 2022 15:38:58 +0700 Subject: Fix #4035 --- modules/sd_models.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) (limited to 'modules/sd_models.py') diff --git a/modules/sd_models.py b/modules/sd_models.py index f86dc3ed..a29c8c1a 100644 --- a/modules/sd_models.py +++ b/modules/sd_models.py @@ -201,7 +201,7 @@ def load_model_weights(model, checkpoint_info): if shared.opts.sd_checkpoint_cache > 0: checkpoints_loaded[checkpoint_info] = model.state_dict().copy() - while len(checkpoints_loaded) > shared.opts.sd_checkpoint_cache: + while len(checkpoints_loaded) > shared.opts.sd_checkpoint_cache + 1: checkpoints_loaded.popitem(last=False) # LRU else: print(f"Loading weights [{sd_model_hash}] from cache") -- cgit v1.2.3 From bf7a699845675eefdabb9cfa40c55398976274ae Mon Sep 17 00:00:00 2001 From: Muhammad Rizqi Nur Date: Mon, 31 Oct 2022 16:27:27 +0700 Subject: Fix #4035 for real now --- modules/sd_models.py | 13 +++++++------ 1 file changed, 7 insertions(+), 6 deletions(-) (limited to 'modules/sd_models.py') diff --git a/modules/sd_models.py b/modules/sd_models.py index a29c8c1a..b2dd005a 100644 --- a/modules/sd_models.py +++ b/modules/sd_models.py @@ -165,6 +165,9 @@ def load_model_weights(model, checkpoint_info): checkpoint_file = checkpoint_info.filename sd_model_hash = checkpoint_info.hash + if shared.opts.sd_checkpoint_cache > 0 and hasattr(model, "sd_checkpoint_info"): + checkpoints_loaded[model.sd_checkpoint_info] = model.state_dict().copy() + if checkpoint_info not in checkpoints_loaded: print(f"Loading weights [{sd_model_hash}] from {checkpoint_file}") @@ -198,16 +201,14 @@ def load_model_weights(model, checkpoint_info): model.first_stage_model.load_state_dict(vae_dict) model.first_stage_model.to(devices.dtype_vae) - - if shared.opts.sd_checkpoint_cache > 0: - checkpoints_loaded[checkpoint_info] = model.state_dict().copy() - while len(checkpoints_loaded) > shared.opts.sd_checkpoint_cache + 1: - checkpoints_loaded.popitem(last=False) # LRU else: print(f"Loading weights [{sd_model_hash}] from cache") - checkpoints_loaded.move_to_end(checkpoint_info) model.load_state_dict(checkpoints_loaded[checkpoint_info]) + if shared.opts.sd_checkpoint_cache > 0: + while len(checkpoints_loaded) > shared.opts.sd_checkpoint_cache: + checkpoints_loaded.popitem(last=False) # LRU + model.sd_model_hash = sd_model_hash model.sd_model_checkpoint = checkpoint_file model.sd_checkpoint_info = checkpoint_info -- cgit v1.2.3 From af758e97fa2c4c853042f121af4e974be01e6696 Mon Sep 17 00:00:00 2001 From: Jairo Correa Date: Tue, 1 Nov 2022 04:01:49 -0300 Subject: Unload sd_model before loading the other --- modules/lowvram.py | 21 +++++++++++++-------- modules/processing.py | 3 +++ modules/sd_hijack.py | 4 ++++ modules/sd_models.py | 14 +++++++++++++- webui.py | 2 +- 5 files changed, 34 insertions(+), 10 deletions(-) (limited to 'modules/sd_models.py') diff --git a/modules/lowvram.py b/modules/lowvram.py index f327c3df..a4652cb1 100644 --- a/modules/lowvram.py +++ b/modules/lowvram.py @@ -38,13 +38,18 @@ def setup_for_low_vram(sd_model, use_medvram): # see below for register_forward_pre_hook; # first_stage_model does not use forward(), it uses encode/decode, so register_forward_pre_hook is # useless here, and we just replace those methods - def first_stage_model_encode_wrap(self, encoder, x): - send_me_to_gpu(self, None) - return encoder(x) - def first_stage_model_decode_wrap(self, decoder, z): - send_me_to_gpu(self, None) - return decoder(z) + first_stage_model = sd_model.first_stage_model + first_stage_model_encode = sd_model.first_stage_model.encode + first_stage_model_decode = sd_model.first_stage_model.decode + + def first_stage_model_encode_wrap(x): + send_me_to_gpu(first_stage_model, None) + return first_stage_model_encode(x) + + def first_stage_model_decode_wrap(z): + send_me_to_gpu(first_stage_model, None) + return first_stage_model_decode(z) # remove three big modules, cond, first_stage, and unet from the model and then # send the model to GPU. Then put modules back. the modules will be in CPU. @@ -56,8 +61,8 @@ def setup_for_low_vram(sd_model, use_medvram): # register hooks for those the first two models sd_model.cond_stage_model.transformer.register_forward_pre_hook(send_me_to_gpu) sd_model.first_stage_model.register_forward_pre_hook(send_me_to_gpu) - sd_model.first_stage_model.encode = lambda x, en=sd_model.first_stage_model.encode: first_stage_model_encode_wrap(sd_model.first_stage_model, en, x) - sd_model.first_stage_model.decode = lambda z, de=sd_model.first_stage_model.decode: first_stage_model_decode_wrap(sd_model.first_stage_model, de, z) + sd_model.first_stage_model.encode = first_stage_model_encode_wrap + sd_model.first_stage_model.decode = first_stage_model_decode_wrap parents[sd_model.cond_stage_model.transformer] = sd_model.cond_stage_model if use_medvram: diff --git a/modules/processing.py b/modules/processing.py index b1df4918..57d3a523 100644 --- a/modules/processing.py +++ b/modules/processing.py @@ -597,6 +597,9 @@ def process_images_inner(p: StableDiffusionProcessing) -> Processed: if p.scripts is not None: p.scripts.postprocess(p, res) + p.sd_model = None + p.sampler = None + return res diff --git a/modules/sd_hijack.py b/modules/sd_hijack.py index 0f10828e..bc49d235 100644 --- a/modules/sd_hijack.py +++ b/modules/sd_hijack.py @@ -94,6 +94,10 @@ class StableDiffusionModelHijack: if type(model_embeddings.token_embedding) == EmbeddingsWithFixes: model_embeddings.token_embedding = model_embeddings.token_embedding.wrapped + self.layers = None + self.circular_enabled = False + self.clip = None + def apply_circular(self, enable): if self.circular_enabled == enable: return diff --git a/modules/sd_models.py b/modules/sd_models.py index f86dc3ed..90007da3 100644 --- a/modules/sd_models.py +++ b/modules/sd_models.py @@ -1,6 +1,7 @@ import collections import os.path import sys +import gc from collections import namedtuple import torch import re @@ -220,6 +221,12 @@ def load_model(checkpoint_info=None): if checkpoint_info.config != shared.cmd_opts.config: print(f"Loading config from: {checkpoint_info.config}") + if shared.sd_model: + sd_hijack.model_hijack.undo_hijack(shared.sd_model) + shared.sd_model = None + gc.collect() + devices.torch_gc() + sd_config = OmegaConf.load(checkpoint_info.config) if should_hijack_inpainting(checkpoint_info): @@ -233,6 +240,7 @@ def load_model(checkpoint_info=None): checkpoint_info = checkpoint_info._replace(config=checkpoint_info.config.replace(".yaml", "-inpainting.yaml")) do_inpainting_hijack() + sd_model = instantiate_from_config(sd_config.model) load_model_weights(sd_model, checkpoint_info) @@ -252,14 +260,18 @@ def load_model(checkpoint_info=None): return sd_model -def reload_model_weights(sd_model, info=None): +def reload_model_weights(sd_model=None, info=None): from modules import lowvram, devices, sd_hijack checkpoint_info = info or select_checkpoint() + if not sd_model: + sd_model = shared.sd_model + if sd_model.sd_model_checkpoint == checkpoint_info.filename: return if sd_model.sd_checkpoint_info.config != checkpoint_info.config or should_hijack_inpainting(checkpoint_info) != should_hijack_inpainting(sd_model.sd_checkpoint_info): + del sd_model checkpoints_loaded.clear() load_model(checkpoint_info) return shared.sd_model diff --git a/webui.py b/webui.py index 6ff95dc4..9c393e55 100644 --- a/webui.py +++ b/webui.py @@ -77,7 +77,7 @@ def initialize(): modules.scripts.load_scripts() modules.sd_models.load_model() - shared.opts.onchange("sd_model_checkpoint", wrap_queued_call(lambda: modules.sd_models.reload_model_weights(shared.sd_model))) + shared.opts.onchange("sd_model_checkpoint", wrap_queued_call(lambda: modules.sd_models.reload_model_weights())) shared.opts.onchange("sd_hypernetwork", wrap_queued_call(lambda: modules.hypernetworks.hypernetwork.load_hypernetwork(shared.opts.sd_hypernetwork))) shared.opts.onchange("sd_hypernetwork_strength", modules.hypernetworks.hypernetwork.apply_strength) -- cgit v1.2.3 From 056f06d3738c267b1014e6e8e1ef5bd97af1fb45 Mon Sep 17 00:00:00 2001 From: Muhammad Rizqi Nur Date: Wed, 2 Nov 2022 12:51:46 +0700 Subject: Reload VAE without reloading sd checkpoint --- modules/sd_models.py | 15 ++++---- modules/sd_vae.py | 97 ++++++++++++++++++++++++++++++++++++++++++++++++---- webui.py | 4 +-- 3 files changed, 98 insertions(+), 18 deletions(-) (limited to 'modules/sd_models.py') diff --git a/modules/sd_models.py b/modules/sd_models.py index 6ab85b65..883639d1 100644 --- a/modules/sd_models.py +++ b/modules/sd_models.py @@ -159,15 +159,13 @@ def get_state_dict_from_checkpoint(pl_sd): return pl_sd -vae_ignore_keys = {"model_ema.decay", "model_ema.num_updates"} - def load_model_weights(model, checkpoint_info, vae_file="auto"): checkpoint_file = checkpoint_info.filename sd_model_hash = checkpoint_info.hash vae_file = sd_vae.resolve_vae(checkpoint_file, vae_file=vae_file) - checkpoint_key = (checkpoint_info, vae_file) + checkpoint_key = checkpoint_info if checkpoint_key not in checkpoints_loaded: print(f"Loading weights [{sd_model_hash}] from {checkpoint_file}") @@ -190,13 +188,12 @@ def load_model_weights(model, checkpoint_info, vae_file="auto"): devices.dtype = torch.float32 if shared.cmd_opts.no_half else torch.float16 devices.dtype_vae = torch.float32 if shared.cmd_opts.no_half or shared.cmd_opts.no_half_vae else torch.float16 - sd_vae.load_vae(model, vae_file) - model.first_stage_model.to(devices.dtype_vae) - if shared.opts.sd_checkpoint_cache > 0: + # if PR #4035 were to get merged, restore base VAE first before caching checkpoints_loaded[checkpoint_key] = model.state_dict().copy() while len(checkpoints_loaded) > shared.opts.sd_checkpoint_cache: checkpoints_loaded.popitem(last=False) # LRU + else: vae_name = sd_vae.get_filename(vae_file) print(f"Loading weights [{sd_model_hash}] with {vae_name} VAE from cache") @@ -207,6 +204,8 @@ def load_model_weights(model, checkpoint_info, vae_file="auto"): model.sd_model_checkpoint = checkpoint_file model.sd_checkpoint_info = checkpoint_info + sd_vae.load_vae(model, vae_file) + def load_model(checkpoint_info=None): from modules import lowvram, sd_hijack @@ -254,14 +253,14 @@ def load_model(checkpoint_info=None): return sd_model -def reload_model_weights(sd_model=None, info=None, force=False): +def reload_model_weights(sd_model=None, info=None): from modules import lowvram, devices, sd_hijack checkpoint_info = info or select_checkpoint() if not sd_model: sd_model = shared.sd_model - if sd_model.sd_model_checkpoint == checkpoint_info.filename and not force: + if sd_model.sd_model_checkpoint == checkpoint_info.filename: return if sd_model.sd_checkpoint_info.config != checkpoint_info.config or should_hijack_inpainting(checkpoint_info) != should_hijack_inpainting(sd_model.sd_checkpoint_info): diff --git a/modules/sd_vae.py b/modules/sd_vae.py index e9239326..78e14e8a 100644 --- a/modules/sd_vae.py +++ b/modules/sd_vae.py @@ -1,26 +1,65 @@ import torch import os from collections import namedtuple -from modules import shared, devices +from modules import shared, devices, script_callbacks from modules.paths import models_path import glob + model_dir = "Stable-diffusion" model_path = os.path.abspath(os.path.join(models_path, model_dir)) vae_dir = "VAE" vae_path = os.path.abspath(os.path.join(models_path, vae_dir)) + vae_ignore_keys = {"model_ema.decay", "model_ema.num_updates"} + + default_vae_dict = {"auto": "auto", "None": "None"} default_vae_list = ["auto", "None"] + + default_vae_values = [default_vae_dict[x] for x in default_vae_list] vae_dict = dict(default_vae_dict) vae_list = list(default_vae_list) first_load = True + +base_vae = None +loaded_vae_file = None +checkpoint_info = None + + +def get_base_vae(model): + if base_vae is not None and checkpoint_info == model.sd_checkpoint_info and model: + return base_vae + return None + + +def store_base_vae(model): + global base_vae, checkpoint_info + if checkpoint_info != model.sd_checkpoint_info: + base_vae = model.first_stage_model.state_dict().copy() + checkpoint_info = model.sd_checkpoint_info + + +def delete_base_vae(): + global base_vae, checkpoint_info + base_vae = None + checkpoint_info = None + + +def restore_base_vae(model): + global base_vae, checkpoint_info + if base_vae is not None and checkpoint_info == model.sd_checkpoint_info: + load_vae_dict(model, base_vae) + delete_base_vae() + + def get_filename(filepath): return os.path.splitext(os.path.basename(filepath))[0] + def refresh_vae_list(vae_path=vae_path, model_path=model_path): global vae_dict, vae_list res = {} @@ -43,6 +82,7 @@ def refresh_vae_list(vae_path=vae_path, model_path=model_path): vae_dict.update(res) return vae_list + def resolve_vae(checkpoint_file, vae_file="auto"): global first_load, vae_dict, vae_list # save_settings = False @@ -96,24 +136,26 @@ def resolve_vae(checkpoint_file, vae_file="auto"): return vae_file -def load_vae(model, vae_file): - global first_load, vae_dict, vae_list + +def load_vae(model, vae_file=None): + global first_load, vae_dict, vae_list, loaded_vae_file # save_settings = False if vae_file: print(f"Loading VAE weights from: {vae_file}") vae_ckpt = torch.load(vae_file, map_location=shared.weight_load_location) vae_dict_1 = {k: v for k, v in vae_ckpt["state_dict"].items() if k[0:4] != "loss" and k not in vae_ignore_keys} - model.first_stage_model.load_state_dict(vae_dict_1) + load_vae_dict(model, vae_dict_1) - # If vae used is not in dict, update it - # It will be removed on refresh though - if vae_file is not None: + # If vae used is not in dict, update it + # It will be removed on refresh though vae_opt = get_filename(vae_file) if vae_opt not in vae_dict: vae_dict[vae_opt] = vae_file vae_list.append(vae_opt) + loaded_vae_file = vae_file + """ # Save current VAE to VAE settings, maybe? will it work? if save_settings: @@ -124,4 +166,45 @@ def load_vae(model, vae_file): """ first_load = False + + +# don't call this from outside +def load_vae_dict(model, vae_dict_1=None): + if vae_dict_1: + store_base_vae(model) + model.first_stage_model.load_state_dict(vae_dict_1) + else: + restore_base_vae() model.first_stage_model.to(devices.dtype_vae) + + +def reload_vae_weights(sd_model=None, vae_file="auto"): + from modules import lowvram, devices, sd_hijack + + if not sd_model: + sd_model = shared.sd_model + + checkpoint_info = sd_model.sd_checkpoint_info + checkpoint_file = checkpoint_info.filename + vae_file = resolve_vae(checkpoint_file, vae_file=vae_file) + + if loaded_vae_file == vae_file: + return + + if shared.cmd_opts.lowvram or shared.cmd_opts.medvram: + lowvram.send_everything_to_cpu() + else: + sd_model.to(devices.cpu) + + sd_hijack.model_hijack.undo_hijack(sd_model) + + load_vae(sd_model, vae_file) + + sd_hijack.model_hijack.hijack(sd_model) + script_callbacks.model_loaded_callback(sd_model) + + if not shared.cmd_opts.lowvram and not shared.cmd_opts.medvram: + sd_model.to(devices.device) + + print(f"VAE Weights loaded.") + return sd_model diff --git a/webui.py b/webui.py index 7cb4691b..034777a2 100644 --- a/webui.py +++ b/webui.py @@ -81,9 +81,7 @@ def initialize(): modules.sd_vae.refresh_vae_list() modules.sd_models.load_model() shared.opts.onchange("sd_model_checkpoint", wrap_queued_call(lambda: modules.sd_models.reload_model_weights())) - # I don't know what needs to be done to only reload VAE, with all those hijacks callbacks, and lowvram, - # so for now this reloads the whole model too - shared.opts.onchange("sd_vae", wrap_queued_call(lambda: modules.sd_models.reload_model_weights(force=True)), call=False) + shared.opts.onchange("sd_vae", wrap_queued_call(lambda: modules.sd_vae.reload_vae_weights()), call=False) shared.opts.onchange("sd_hypernetwork", wrap_queued_call(lambda: modules.hypernetworks.hypernetwork.load_hypernetwork(shared.opts.sd_hypernetwork))) shared.opts.onchange("sd_hypernetwork_strength", modules.hypernetworks.hypernetwork.apply_strength) -- cgit v1.2.3 From f2a5cbe6f55592c4c5527b8e0bf99ea8d658f057 Mon Sep 17 00:00:00 2001 From: AUTOMATIC <16777216c@gmail.com> Date: Wed, 2 Nov 2022 14:41:29 +0300 Subject: fix #3986 breaking --no-half-vae --- modules/sd_models.py | 9 +++++++++ 1 file changed, 9 insertions(+) (limited to 'modules/sd_models.py') diff --git a/modules/sd_models.py b/modules/sd_models.py index 883639d1..5075fadb 100644 --- a/modules/sd_models.py +++ b/modules/sd_models.py @@ -183,11 +183,20 @@ def load_model_weights(model, checkpoint_info, vae_file="auto"): model.to(memory_format=torch.channels_last) if not shared.cmd_opts.no_half: + vae = model.first_stage_model + + # with --no-half-vae, remove VAE from model when doing half() to prevent its weights from being converted to float16 + if shared.cmd_opts.no_half_vae: + model.first_stage_model = None + model.half() + model.first_stage_model = vae devices.dtype = torch.float32 if shared.cmd_opts.no_half else torch.float16 devices.dtype_vae = torch.float32 if shared.cmd_opts.no_half or shared.cmd_opts.no_half_vae else torch.float16 + model.first_stage_model.to(devices.dtype_vae) + if shared.opts.sd_checkpoint_cache > 0: # if PR #4035 were to get merged, restore base VAE first before caching checkpoints_loaded[checkpoint_key] = model.state_dict().copy() -- cgit v1.2.3 From 3780ad3ad837dd406da39eebd5d91009b5a58445 Mon Sep 17 00:00:00 2001 From: digburn Date: Fri, 4 Nov 2022 00:40:21 +0000 Subject: fix: loading models without vae from cache --- modules/sd_models.py | 5 +++-- 1 file changed, 3 insertions(+), 2 deletions(-) (limited to 'modules/sd_models.py') diff --git a/modules/sd_models.py b/modules/sd_models.py index 5075fadb..ae427a5c 100644 --- a/modules/sd_models.py +++ b/modules/sd_models.py @@ -204,8 +204,9 @@ def load_model_weights(model, checkpoint_info, vae_file="auto"): checkpoints_loaded.popitem(last=False) # LRU else: - vae_name = sd_vae.get_filename(vae_file) - print(f"Loading weights [{sd_model_hash}] with {vae_name} VAE from cache") + vae_name = sd_vae.get_filename(vae_file) if vae_file else None + vae_message = f" with {vae_name} VAE" if vae_name else "" + print(f"Loading weights [{sd_model_hash}]{vae_message} from cache") checkpoints_loaded.move_to_end(checkpoint_key) model.load_state_dict(checkpoints_loaded[checkpoint_key]) -- cgit v1.2.3 From 99043f33606d3057f83ea52a403e10cd29d1f7e7 Mon Sep 17 00:00:00 2001 From: AUTOMATIC <16777216c@gmail.com> Date: Fri, 4 Nov 2022 11:20:42 +0300 Subject: fix one of previous merges breaking the program --- modules/sd_models.py | 2 ++ 1 file changed, 2 insertions(+) (limited to 'modules/sd_models.py') diff --git a/modules/sd_models.py b/modules/sd_models.py index 63e07a12..34c57bfa 100644 --- a/modules/sd_models.py +++ b/modules/sd_models.py @@ -167,6 +167,8 @@ def load_model_weights(model, checkpoint_info, vae_file="auto"): sd_vae.restore_base_vae(model) checkpoints_loaded[model.sd_checkpoint_info] = model.state_dict().copy() + vae_file = sd_vae.resolve_vae(checkpoint_file, vae_file=vae_file) + if checkpoint_info not in checkpoints_loaded: print(f"Loading weights [{sd_model_hash}] from {checkpoint_file}") -- cgit v1.2.3 From 3b51d239ac9201228c6032fc109111e347e8e6b0 Mon Sep 17 00:00:00 2001 From: cluder <1590330+cluder@users.noreply.github.com> Date: Wed, 9 Nov 2022 04:54:21 +0100 Subject: - do not use ckpt cache, if disabled - cache model after is has been loaded from file --- modules/sd_models.py | 27 +++++++++++++++++---------- 1 file changed, 17 insertions(+), 10 deletions(-) (limited to 'modules/sd_models.py') diff --git a/modules/sd_models.py b/modules/sd_models.py index 34c57bfa..720c2a96 100644 --- a/modules/sd_models.py +++ b/modules/sd_models.py @@ -163,13 +163,21 @@ def load_model_weights(model, checkpoint_info, vae_file="auto"): checkpoint_file = checkpoint_info.filename sd_model_hash = checkpoint_info.hash - if shared.opts.sd_checkpoint_cache > 0 and hasattr(model, "sd_checkpoint_info"): + cache_enabled = shared.opts.sd_checkpoint_cache > 0 + + if cache_enabled: sd_vae.restore_base_vae(model) - checkpoints_loaded[model.sd_checkpoint_info] = model.state_dict().copy() vae_file = sd_vae.resolve_vae(checkpoint_file, vae_file=vae_file) - if checkpoint_info not in checkpoints_loaded: + if cache_enabled and checkpoint_info in checkpoints_loaded: + # use checkpoint cache + vae_name = sd_vae.get_filename(vae_file) if vae_file else None + vae_message = f" with {vae_name} VAE" if vae_name else "" + print(f"Loading weights [{sd_model_hash}]{vae_message} from cache") + model.load_state_dict(checkpoints_loaded[checkpoint_info]) + else: + # load from file print(f"Loading weights [{sd_model_hash}] from {checkpoint_file}") pl_sd = torch.load(checkpoint_file, map_location=shared.weight_load_location) @@ -180,6 +188,10 @@ def load_model_weights(model, checkpoint_info, vae_file="auto"): del pl_sd model.load_state_dict(sd, strict=False) del sd + + if cache_enabled: + # cache newly loaded model + checkpoints_loaded[checkpoint_info] = model.state_dict().copy() if shared.cmd_opts.opt_channelslast: model.to(memory_format=torch.channels_last) @@ -199,13 +211,8 @@ def load_model_weights(model, checkpoint_info, vae_file="auto"): model.first_stage_model.to(devices.dtype_vae) - else: - vae_name = sd_vae.get_filename(vae_file) if vae_file else None - vae_message = f" with {vae_name} VAE" if vae_name else "" - print(f"Loading weights [{sd_model_hash}]{vae_message} from cache") - model.load_state_dict(checkpoints_loaded[checkpoint_info]) - - if shared.opts.sd_checkpoint_cache > 0: + # clean up cache if limit is reached + if cache_enabled: while len(checkpoints_loaded) > shared.opts.sd_checkpoint_cache: checkpoints_loaded.popitem(last=False) # LRU -- cgit v1.2.3 From eebf49592ad2c0933e58b06a098b92e48d47e4fe Mon Sep 17 00:00:00 2001 From: cluder <1590330+cluder@users.noreply.github.com> Date: Wed, 9 Nov 2022 07:17:09 +0100 Subject: restore #4035 behavior - if checkpoint cache is set to 1, keep 2 models in cache (current +1 more) --- modules/sd_models.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) (limited to 'modules/sd_models.py') diff --git a/modules/sd_models.py b/modules/sd_models.py index 720c2a96..80addf03 100644 --- a/modules/sd_models.py +++ b/modules/sd_models.py @@ -213,7 +213,7 @@ def load_model_weights(model, checkpoint_info, vae_file="auto"): # clean up cache if limit is reached if cache_enabled: - while len(checkpoints_loaded) > shared.opts.sd_checkpoint_cache: + while len(checkpoints_loaded) > shared.opts.sd_checkpoint_cache + 1: # we need to count the current model checkpoints_loaded.popitem(last=False) # LRU model.sd_model_hash = sd_model_hash -- cgit v1.2.3 From abc1e79a5da24a1ea0f4bceedcdf225f32010aa8 Mon Sep 17 00:00:00 2001 From: Muhammad Rizqi Nur Date: Thu, 3 Nov 2022 11:10:53 +0700 Subject: Fix base VAE caching was done after loading VAE, also add safeguard --- modules/sd_models.py | 1 + modules/sd_vae.py | 19 ++++++++----------- 2 files changed, 9 insertions(+), 11 deletions(-) (limited to 'modules/sd_models.py') diff --git a/modules/sd_models.py b/modules/sd_models.py index 80addf03..e4dba62c 100644 --- a/modules/sd_models.py +++ b/modules/sd_models.py @@ -220,6 +220,7 @@ def load_model_weights(model, checkpoint_info, vae_file="auto"): model.sd_model_checkpoint = checkpoint_file model.sd_checkpoint_info = checkpoint_info + sd_vae.clear_loaded_vae() sd_vae.load_vae(model, vae_file) diff --git a/modules/sd_vae.py b/modules/sd_vae.py index 7a79239f..dd69a5e6 100644 --- a/modules/sd_vae.py +++ b/modules/sd_vae.py @@ -15,7 +15,7 @@ vae_path = os.path.abspath(os.path.join(models_path, vae_dir)) vae_ignore_keys = {"model_ema.decay", "model_ema.num_updates"} -default_vae_dict = {"auto": "auto", "None": "None"} +default_vae_dict = {"auto": "auto", "None": None, None: None} default_vae_list = ["auto", "None"] @@ -39,6 +39,7 @@ def get_base_vae(model): def store_base_vae(model): global base_vae, checkpoint_info if checkpoint_info != model.sd_checkpoint_info: + assert not loaded_vae_file, "Trying to store non-base VAE!" base_vae = model.first_stage_model.state_dict().copy() checkpoint_info = model.sd_checkpoint_info @@ -50,9 +51,11 @@ def delete_base_vae(): def restore_base_vae(model): + global loaded_vae_file if base_vae is not None and checkpoint_info == model.sd_checkpoint_info: print("Restoring base VAE") load_vae_dict(model, base_vae) + loaded_vae_file = None delete_base_vae() @@ -140,10 +143,10 @@ def load_vae(model, vae_file=None): if vae_file: print(f"Loading VAE weights from: {vae_file}") + store_base_vae(model) vae_ckpt = torch.load(vae_file, map_location=shared.weight_load_location) vae_dict_1 = {k: v for k, v in vae_ckpt["state_dict"].items() if k[0:4] != "loss" and k not in vae_ignore_keys} load_vae_dict(model, vae_dict_1) - store_base_vae(model) # If vae used is not in dict, update it # It will be removed on refresh though @@ -157,15 +160,6 @@ def load_vae(model, vae_file=None): loaded_vae_file = vae_file - """ - # Save current VAE to VAE settings, maybe? will it work? - if save_settings: - if vae_file is None: - vae_opt = "None" - - # shared.opts.sd_vae = vae_opt - """ - first_load = False @@ -174,6 +168,9 @@ def load_vae_dict(model, vae_dict_1): model.first_stage_model.load_state_dict(vae_dict_1) model.first_stage_model.to(devices.dtype_vae) +def clear_loaded_vae(): + global loaded_vae_file + loaded_vae_file = None def reload_vae_weights(sd_model=None, vae_file="auto"): from modules import lowvram, devices, sd_hijack -- cgit v1.2.3 From c7be83bf0240498d9382e2afeaa3f0677d26c7f6 Mon Sep 17 00:00:00 2001 From: Muhammad Rizqi Nur Date: Sun, 13 Nov 2022 11:11:14 +0700 Subject: Misc Misc --- modules/sd_models.py | 1 + modules/sd_vae.py | 3 +-- modules/shared.py | 2 +- 3 files changed, 3 insertions(+), 3 deletions(-) (limited to 'modules/sd_models.py') diff --git a/modules/sd_models.py b/modules/sd_models.py index e4dba62c..cd7fe37a 100644 --- a/modules/sd_models.py +++ b/modules/sd_models.py @@ -220,6 +220,7 @@ def load_model_weights(model, checkpoint_info, vae_file="auto"): model.sd_model_checkpoint = checkpoint_file model.sd_checkpoint_info = checkpoint_info + sd_vae.delete_base_vae() sd_vae.clear_loaded_vae() sd_vae.load_vae(model, vae_file) diff --git a/modules/sd_vae.py b/modules/sd_vae.py index dd69a5e6..13bf3d31 100644 --- a/modules/sd_vae.py +++ b/modules/sd_vae.py @@ -154,8 +154,7 @@ def load_vae(model, vae_file=None): if vae_opt not in vae_dict: vae_dict[vae_opt] = vae_file vae_list.append(vae_opt) - # shared.opts.data['sd_vae'] = vae_opt - else: + elif loaded_vae_file: restore_base_vae(model) loaded_vae_file = vae_file diff --git a/modules/shared.py b/modules/shared.py index 17132e42..a9daf800 100644 --- a/modules/shared.py +++ b/modules/shared.py @@ -335,7 +335,7 @@ options_templates.update(options_section(('training', "Training"), { options_templates.update(options_section(('sd', "Stable Diffusion"), { "sd_model_checkpoint": OptionInfo(None, "Stable Diffusion checkpoint", gr.Dropdown, lambda: {"choices": modules.sd_models.checkpoint_tiles()}, refresh=sd_models.list_models), "sd_checkpoint_cache": OptionInfo(0, "Checkpoints to cache in RAM", gr.Slider, {"minimum": 0, "maximum": 10, "step": 1}), - "sd_vae": OptionInfo("auto", "SD VAE", gr.Dropdown, lambda: {"choices": list(sd_vae.vae_list)}, refresh=sd_vae.refresh_vae_list), + "sd_vae": OptionInfo("auto", "SD VAE", gr.Dropdown, lambda: {"choices": sd_vae.vae_list}, refresh=sd_vae.refresh_vae_list), "sd_hypernetwork": OptionInfo("None", "Hypernetwork", gr.Dropdown, lambda: {"choices": ["None"] + [x for x in hypernetworks.keys()]}, refresh=reload_hypernetworks), "sd_hypernetwork_strength": OptionInfo(1.0, "Hypernetwork strength", gr.Slider, {"minimum": 0.0, "maximum": 1.0, "step": 0.001}), "inpainting_mask_weight": OptionInfo(1.0, "Inpainting conditioning mask strength", gr.Slider, {"minimum": 0.0, "maximum": 1.0, "step": 0.01}), -- cgit v1.2.3 From 2c5ca706a7e624d268545ba3318ba230b7b33477 Mon Sep 17 00:00:00 2001 From: Muhammad Rizqi Nur Date: Sun, 13 Nov 2022 10:55:47 +0700 Subject: Remove no longer necessary parts and add vae_file safeguard --- modules/sd_models.py | 10 ++-------- modules/sd_vae.py | 1 + 2 files changed, 3 insertions(+), 8 deletions(-) (limited to 'modules/sd_models.py') diff --git a/modules/sd_models.py b/modules/sd_models.py index 80addf03..c59151e0 100644 --- a/modules/sd_models.py +++ b/modules/sd_models.py @@ -165,16 +165,9 @@ def load_model_weights(model, checkpoint_info, vae_file="auto"): cache_enabled = shared.opts.sd_checkpoint_cache > 0 - if cache_enabled: - sd_vae.restore_base_vae(model) - - vae_file = sd_vae.resolve_vae(checkpoint_file, vae_file=vae_file) - if cache_enabled and checkpoint_info in checkpoints_loaded: # use checkpoint cache - vae_name = sd_vae.get_filename(vae_file) if vae_file else None - vae_message = f" with {vae_name} VAE" if vae_name else "" - print(f"Loading weights [{sd_model_hash}]{vae_message} from cache") + print(f"Loading weights [{sd_model_hash}] from cache") model.load_state_dict(checkpoints_loaded[checkpoint_info]) else: # load from file @@ -220,6 +213,7 @@ def load_model_weights(model, checkpoint_info, vae_file="auto"): model.sd_model_checkpoint = checkpoint_file model.sd_checkpoint_info = checkpoint_info + vae_file = sd_vae.resolve_vae(checkpoint_file, vae_file=vae_file) sd_vae.load_vae(model, vae_file) diff --git a/modules/sd_vae.py b/modules/sd_vae.py index 71e7a6e6..8bdb2c17 100644 --- a/modules/sd_vae.py +++ b/modules/sd_vae.py @@ -139,6 +139,7 @@ def load_vae(model, vae_file=None): # save_settings = False if vae_file: + assert os.path.isfile(vae_file), f"VAE file doesn't exist: {vae_file}" print(f"Loading VAE weights from: {vae_file}") vae_ckpt = torch.load(vae_file, map_location=shared.weight_load_location) vae_dict_1 = {k: v for k, v in vae_ckpt["state_dict"].items() if k[0:4] != "loss" and k not in vae_ignore_keys} -- cgit v1.2.3 From 0efffbb407a9d07eae6850374099775385ce176c Mon Sep 17 00:00:00 2001 From: Nicolas Patry Date: Mon, 21 Nov 2022 14:04:25 +0100 Subject: Supporting `*.safetensors` format. If a model file exists with extension `.safetensors` then we can load it more safely than with PyTorch weights. --- modules/sd_models.py | 11 +++++++++-- 1 file changed, 9 insertions(+), 2 deletions(-) (limited to 'modules/sd_models.py') diff --git a/modules/sd_models.py b/modules/sd_models.py index 80addf03..0164cc1b 100644 --- a/modules/sd_models.py +++ b/modules/sd_models.py @@ -45,7 +45,7 @@ def checkpoint_tiles(): def list_models(): checkpoints_list.clear() - model_list = modelloader.load_models(model_path=model_path, command_path=shared.cmd_opts.ckpt_dir, ext_filter=[".ckpt"]) + model_list = modelloader.load_models(model_path=model_path, command_path=shared.cmd_opts.ckpt_dir, ext_filter=[".ckpt", ".safetensors"]) def modeltitle(path, shorthash): abspath = os.path.abspath(path) @@ -180,7 +180,14 @@ def load_model_weights(model, checkpoint_info, vae_file="auto"): # load from file print(f"Loading weights [{sd_model_hash}] from {checkpoint_file}") - pl_sd = torch.load(checkpoint_file, map_location=shared.weight_load_location) + if checkpoint_file.endswith(".safetensors"): + try: + from safetensors.torch import load_file + except ImportError as e: + raise ImportError(f"The model is in safetensors format and it is not installed, use `pip install safetensors`: {e}") + pl_sd = load_file(checkpoint_file, device=shared.weight_load_location) + else: + pl_sd = torch.load(checkpoint_file, map_location=shared.weight_load_location) if "global_step" in pl_sd: print(f"Global Step: {pl_sd['global_step']}") -- cgit v1.2.3 From 1e506657e1cb732a5f0e567ba2585fba2bbb1327 Mon Sep 17 00:00:00 2001 From: MrCheeze Date: Sat, 26 Nov 2022 13:28:44 -0500 Subject: no-half support for SD 2.0 --- modules/sd_models.py | 3 +++ 1 file changed, 3 insertions(+) (limited to 'modules/sd_models.py') diff --git a/modules/sd_models.py b/modules/sd_models.py index c59151e0..0e0bd79e 100644 --- a/modules/sd_models.py +++ b/modules/sd_models.py @@ -244,6 +244,9 @@ def load_model(checkpoint_info=None): do_inpainting_hijack() + if shared.cmd_opts.no_half: + sd_config.model.params.unet_config.params.use_fp16 = False + sd_model = instantiate_from_config(sd_config.model) load_model_weights(sd_model, checkpoint_info) -- cgit v1.2.3 From 6074175faa751dde933aa8e15cd687ca4e4b4a23 Mon Sep 17 00:00:00 2001 From: AUTOMATIC <16777216c@gmail.com> Date: Sun, 27 Nov 2022 14:46:40 +0300 Subject: add safetensors to requirements --- modules/sd_models.py | 11 +++++------ requirements.txt | 1 + requirements_versions.txt | 1 + 3 files changed, 7 insertions(+), 6 deletions(-) (limited to 'modules/sd_models.py') diff --git a/modules/sd_models.py b/modules/sd_models.py index ae36841a..77236480 100644 --- a/modules/sd_models.py +++ b/modules/sd_models.py @@ -5,6 +5,7 @@ import gc from collections import namedtuple import torch import re +import safetensors.torch from omegaconf import OmegaConf from ldm.util import instantiate_from_config @@ -173,14 +174,12 @@ def load_model_weights(model, checkpoint_info, vae_file="auto"): # load from file print(f"Loading weights [{sd_model_hash}] from {checkpoint_file}") - if checkpoint_file.endswith(".safetensors"): - try: - from safetensors.torch import load_file - except ImportError as e: - raise ImportError(f"The model is in safetensors format and it is not installed, use `pip install safetensors`: {e}") - pl_sd = load_file(checkpoint_file, device=shared.weight_load_location) + _, extension = os.path.splitext(checkpoint_file) + if extension.lower() == ".safetensors": + pl_sd = safetensors.torch.load_file(checkpoint_file, device=shared.weight_load_location) else: pl_sd = torch.load(checkpoint_file, map_location=shared.weight_load_location) + if "global_step" in pl_sd: print(f"Global Step: {pl_sd['global_step']}") diff --git a/requirements.txt b/requirements.txt index e4e5ec64..5f3d9623 100644 --- a/requirements.txt +++ b/requirements.txt @@ -29,3 +29,4 @@ lark inflection GitPython torchsde +safetensors diff --git a/requirements_versions.txt b/requirements_versions.txt index 8d557fe3..035fa82f 100644 --- a/requirements_versions.txt +++ b/requirements_versions.txt @@ -26,3 +26,4 @@ lark==1.1.2 inflection==0.5.1 GitPython==3.1.27 torchsde==0.2.5 +safetensors==0.2.5 -- cgit v1.2.3 From dac9b6f15de5e675053d9490a20e0457dcd1a23e Mon Sep 17 00:00:00 2001 From: AUTOMATIC <16777216c@gmail.com> Date: Sun, 27 Nov 2022 15:51:29 +0300 Subject: add safetensors support for model merging #4869 --- modules/extras.py | 26 ++++++++++++++------------ modules/sd_models.py | 26 +++++++++++++++----------- modules/ui.py | 7 ++++++- 3 files changed, 35 insertions(+), 24 deletions(-) (limited to 'modules/sd_models.py') diff --git a/modules/extras.py b/modules/extras.py index 71b93a06..3d65d90a 100644 --- a/modules/extras.py +++ b/modules/extras.py @@ -20,6 +20,7 @@ import modules.codeformer_model import piexif import piexif.helper import gradio as gr +import safetensors.torch class LruCache(OrderedDict): @@ -249,7 +250,7 @@ def run_pnginfo(image): return '', geninfo, info -def run_modelmerger(primary_model_name, secondary_model_name, teritary_model_name, interp_method, multiplier, save_as_half, custom_name): +def run_modelmerger(primary_model_name, secondary_model_name, teritary_model_name, interp_method, multiplier, save_as_half, custom_name, checkpoint_format): def weighted_sum(theta0, theta1, alpha): return ((1 - alpha) * theta0) + (alpha * theta1) @@ -264,19 +265,15 @@ def run_modelmerger(primary_model_name, secondary_model_name, teritary_model_nam teritary_model_info = sd_models.checkpoints_list.get(teritary_model_name, None) print(f"Loading {primary_model_info.filename}...") - primary_model = torch.load(primary_model_info.filename, map_location='cpu') - theta_0 = sd_models.get_state_dict_from_checkpoint(primary_model) + theta_0 = sd_models.read_state_dict(primary_model_info.filename, map_location='cpu') print(f"Loading {secondary_model_info.filename}...") - secondary_model = torch.load(secondary_model_info.filename, map_location='cpu') - theta_1 = sd_models.get_state_dict_from_checkpoint(secondary_model) + theta_1 = sd_models.read_state_dict(secondary_model_info.filename, map_location='cpu') if teritary_model_info is not None: print(f"Loading {teritary_model_info.filename}...") - teritary_model = torch.load(teritary_model_info.filename, map_location='cpu') - theta_2 = sd_models.get_state_dict_from_checkpoint(teritary_model) + theta_2 = sd_models.read_state_dict(teritary_model_info.filename, map_location='cpu') else: - teritary_model = None theta_2 = None theta_funcs = { @@ -295,7 +292,7 @@ def run_modelmerger(primary_model_name, secondary_model_name, teritary_model_nam theta_1[key] = theta_func1(theta_1[key], t2) else: theta_1[key] = torch.zeros_like(theta_1[key]) - del theta_2, teritary_model + del theta_2 for key in tqdm.tqdm(theta_0.keys()): if 'model' in key and key in theta_1: @@ -314,12 +311,17 @@ def run_modelmerger(primary_model_name, secondary_model_name, teritary_model_nam ckpt_dir = shared.cmd_opts.ckpt_dir or sd_models.model_path - filename = primary_model_info.model_name + '_' + str(round(1-multiplier, 2)) + '-' + secondary_model_info.model_name + '_' + str(round(multiplier, 2)) + '-' + interp_method.replace(" ", "_") + '-merged.ckpt' - filename = filename if custom_name == '' else (custom_name + '.ckpt') + filename = primary_model_info.model_name + '_' + str(round(1-multiplier, 2)) + '-' + secondary_model_info.model_name + '_' + str(round(multiplier, 2)) + '-' + interp_method.replace(" ", "_") + '-merged.' + checkpoint_format + filename = filename if custom_name == '' else (custom_name + '.' + checkpoint_format) output_modelname = os.path.join(ckpt_dir, filename) print(f"Saving to {output_modelname}...") - torch.save(primary_model, output_modelname) + + _, extension = os.path.splitext(output_modelname) + if extension.lower() == ".safetensors": + safetensors.torch.save_file(theta_0, output_modelname, metadata={"format": "pt"}) + else: + torch.save(theta_0, output_modelname) sd_models.list_models() diff --git a/modules/sd_models.py b/modules/sd_models.py index 77236480..a1ea5611 100644 --- a/modules/sd_models.py +++ b/modules/sd_models.py @@ -160,6 +160,20 @@ def get_state_dict_from_checkpoint(pl_sd): return pl_sd +def read_state_dict(checkpoint_file, print_global_state=False, map_location=None): + _, extension = os.path.splitext(checkpoint_file) + if extension.lower() == ".safetensors": + pl_sd = safetensors.torch.load_file(checkpoint_file, device=map_location or shared.weight_load_location) + else: + pl_sd = torch.load(checkpoint_file, map_location=map_location or shared.weight_load_location) + + if print_global_state and "global_step" in pl_sd: + print(f"Global Step: {pl_sd['global_step']}") + + sd = get_state_dict_from_checkpoint(pl_sd) + return sd + + def load_model_weights(model, checkpoint_info, vae_file="auto"): checkpoint_file = checkpoint_info.filename sd_model_hash = checkpoint_info.hash @@ -174,17 +188,7 @@ def load_model_weights(model, checkpoint_info, vae_file="auto"): # load from file print(f"Loading weights [{sd_model_hash}] from {checkpoint_file}") - _, extension = os.path.splitext(checkpoint_file) - if extension.lower() == ".safetensors": - pl_sd = safetensors.torch.load_file(checkpoint_file, device=shared.weight_load_location) - else: - pl_sd = torch.load(checkpoint_file, map_location=shared.weight_load_location) - - if "global_step" in pl_sd: - print(f"Global Step: {pl_sd['global_step']}") - - sd = get_state_dict_from_checkpoint(pl_sd) - del pl_sd + sd = read_state_dict(checkpoint_file) model.load_state_dict(sd, strict=False) del sd diff --git a/modules/ui.py b/modules/ui.py index de2b5544..aa13978d 100644 --- a/modules/ui.py +++ b/modules/ui.py @@ -1164,7 +1164,11 @@ def create_ui(wrap_gradio_gpu_call): custom_name = gr.Textbox(label="Custom Name (Optional)") interp_amount = gr.Slider(minimum=0.0, maximum=1.0, step=0.05, label='Multiplier (M) - set to 0 to get model A', value=0.3) interp_method = gr.Radio(choices=["Weighted sum", "Add difference"], value="Weighted sum", label="Interpolation Method") - save_as_half = gr.Checkbox(value=False, label="Save as float16") + + with gr.Row(): + checkpoint_format = gr.Radio(choices=["ckpt", "safetensors"], value="ckpt", label="Checkpoint format") + save_as_half = gr.Checkbox(value=False, label="Save as float16") + modelmerger_merge = gr.Button(elem_id="modelmerger_merge", label="Merge", variant='primary') with gr.Column(variant='panel'): @@ -1692,6 +1696,7 @@ def create_ui(wrap_gradio_gpu_call): interp_amount, save_as_half, custom_name, + checkpoint_format, ], outputs=[ submit_result, -- cgit v1.2.3 From 0376da180c81a11880a2587903d69d85541051e7 Mon Sep 17 00:00:00 2001 From: AUTOMATIC <16777216c@gmail.com> Date: Mon, 28 Nov 2022 08:39:59 +0300 Subject: make it possible to save nai model using safetensors --- modules/sd_models.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) (limited to 'modules/sd_models.py') diff --git a/modules/sd_models.py b/modules/sd_models.py index a1ea5611..283cf1cd 100644 --- a/modules/sd_models.py +++ b/modules/sd_models.py @@ -144,8 +144,8 @@ def transform_checkpoint_dict_key(k): def get_state_dict_from_checkpoint(pl_sd): - if "state_dict" in pl_sd: - pl_sd = pl_sd["state_dict"] + pl_sd = pl_sd.pop("state_dict", pl_sd) + pl_sd.pop("state_dict", None) sd = {} for k, v in pl_sd.items(): -- cgit v1.2.3 From 1ed4f0e22807f3afef925210182cbbee51f0cb2c Mon Sep 17 00:00:00 2001 From: Jay Smith Date: Thu, 8 Dec 2022 18:14:35 -0600 Subject: Depth2img model support --- README.md | 1 + modules/processing.py | 38 ++++++++++++++++++++++++++++++++++---- modules/sd_models.py | 46 ++++++++++++++++++++++++++++++++++++++++++++++ 3 files changed, 81 insertions(+), 4 deletions(-) (limited to 'modules/sd_models.py') diff --git a/README.md b/README.md index 8a4ffade..55990581 100644 --- a/README.md +++ b/README.md @@ -135,6 +135,7 @@ The documentation was moved from this README over to the project's [wiki](https: - SwinIR - https://github.com/JingyunLiang/SwinIR - Swin2SR - https://github.com/mv-lab/swin2sr - LDSR - https://github.com/Hafiidz/latent-diffusion +- MiDaS - https://github.com/isl-org/MiDaS - Ideas for optimizations - https://github.com/basujindal/stable-diffusion - Cross Attention layer optimization - Doggettx - https://github.com/Doggettx/stable-diffusion, original idea for prompt editing. - Cross Attention layer optimization - InvokeAI, lstein - https://github.com/invoke-ai/InvokeAI (originally http://github.com/lstein/stable-diffusion) diff --git a/modules/processing.py b/modules/processing.py index 3d2c4dc9..0417ffc5 100644 --- a/modules/processing.py +++ b/modules/processing.py @@ -21,7 +21,10 @@ import modules.face_restoration import modules.images as images import modules.styles import logging +from ldm.data.util import AddMiDaS +from ldm.models.diffusion.ddpm import LatentDepth2ImageDiffusion +from einops import repeat, rearrange # some of those options should not be changed at all because they would break the model, so I removed them from options. opt_C = 4 @@ -150,11 +153,26 @@ class StableDiffusionProcessing(): return image_conditioning - def img2img_image_conditioning(self, source_image, latent_image, image_mask = None): - if self.sampler.conditioning_key not in {'hybrid', 'concat'}: - # Dummy zero conditioning if we're not using inpainting model. - return latent_image.new_zeros(latent_image.shape[0], 5, 1, 1) + def depth2img_image_conditioning(self, source_image): + # Use the AddMiDaS helper to Format our source image to suit the MiDaS model + transformer = AddMiDaS(model_type="dpt_hybrid") + transformed = transformer({"jpg": rearrange(source_image[0], "c h w -> h w c")}) + midas_in = torch.from_numpy(transformed["midas_in"][None, ...]).to(device=shared.device) + midas_in = repeat(midas_in, "1 ... -> n ...", n=self.batch_size) + + conditioning_image = self.sd_model.get_first_stage_encoding(self.sd_model.encode_first_stage(source_image)) + conditioning = torch.nn.functional.interpolate( + self.sd_model.depth_model(midas_in), + size=conditioning_image.shape[2:], + mode="bicubic", + align_corners=False, + ) + + (depth_min, depth_max) = torch.aminmax(conditioning) + conditioning = 2. * (conditioning - depth_min) / (depth_max - depth_min) - 1. + return conditioning + def inpainting_image_conditioning(self, source_image, latent_image, image_mask = None): self.is_using_inpainting_conditioning = True # Handle the different mask inputs @@ -191,6 +209,18 @@ class StableDiffusionProcessing(): return image_conditioning + def img2img_image_conditioning(self, source_image, latent_image, image_mask=None): + # HACK: Using introspection as the Depth2Image model doesn't appear to uniquely + # identify itself with a field common to all models. The conditioning_key is also hybrid. + if isinstance(self.sd_model, LatentDepth2ImageDiffusion): + return self.depth2img_image_conditioning(source_image) + + if self.sampler.conditioning_key in {'hybrid', 'concat'}: + return self.inpainting_image_conditioning(source_image, latent_image, image_mask=image_mask) + + # Dummy zero conditioning if we're not using inpainting or depth model. + return latent_image.new_zeros(latent_image.shape[0], 5, 1, 1) + def init(self, all_prompts, all_seeds, all_subseeds): pass diff --git a/modules/sd_models.py b/modules/sd_models.py index 283cf1cd..139952ba 100644 --- a/modules/sd_models.py +++ b/modules/sd_models.py @@ -7,6 +7,9 @@ import torch import re import safetensors.torch from omegaconf import OmegaConf +from os import mkdir +from urllib import request +import ldm.modules.midas as midas from ldm.util import instantiate_from_config @@ -36,6 +39,7 @@ def setup_model(): os.makedirs(model_path) list_models() + enable_midas_autodownload() def checkpoint_tiles(): @@ -227,6 +231,48 @@ def load_model_weights(model, checkpoint_info, vae_file="auto"): sd_vae.load_vae(model, vae_file) +def enable_midas_autodownload(): + """ + Gives the ldm.modules.midas.api.load_model function automatic downloading. + + When the 512-depth-ema model, and other future models like it, is loaded, + it calls midas.api.load_model to load the associated midas depth model. + This function applies a wrapper to download the model to the correct + location automatically. + """ + + midas_path = os.path.join(models_path, 'midas') + + # stable-diffusion-stability-ai hard-codes the midas model path to + # a location that differs from where other scripts using this model look. + # HACK: Overriding the path here. + for k, v in midas.api.ISL_PATHS.items(): + file_name = os.path.basename(v) + midas.api.ISL_PATHS[k] = os.path.join(midas_path, file_name) + + midas_urls = { + "dpt_large": "https://github.com/intel-isl/DPT/releases/download/1_0/dpt_large-midas-2f21e586.pt", + "dpt_hybrid": "https://github.com/intel-isl/DPT/releases/download/1_0/dpt_hybrid-midas-501f0c75.pt", + "midas_v21": "https://github.com/AlexeyAB/MiDaS/releases/download/midas_dpt/midas_v21-f6b98070.pt", + "midas_v21_small": "https://github.com/AlexeyAB/MiDaS/releases/download/midas_dpt/midas_v21_small-70d6b9c8.pt", + } + + midas.api.load_model_inner = midas.api.load_model + + def load_model_wrapper(model_type): + path = midas.api.ISL_PATHS[model_type] + if not os.path.exists(path): + if not os.path.exists(midas_path): + mkdir(midas_path) + + print(f"Downloading midas model weights for {model_type} to {path}") + request.urlretrieve(midas_urls[model_type], path) + print(f"{model_type} downloaded") + + return midas.api.load_model_inner(model_type) + + midas.api.load_model = load_model_wrapper + def load_model(checkpoint_info=None): from modules import lowvram, sd_hijack checkpoint_info = checkpoint_info or select_checkpoint() -- cgit v1.2.3 From bd81a09eacf02dad095b98094ab936f276d0343f Mon Sep 17 00:00:00 2001 From: MrCheeze Date: Sat, 10 Dec 2022 11:29:26 -0500 Subject: fix support for 2.0 inpainting model while maintaining support for 1.5 inpainting model --- modules/sd_hijack_inpainting.py | 3 +-- modules/sd_models.py | 1 + 2 files changed, 2 insertions(+), 2 deletions(-) (limited to 'modules/sd_models.py') diff --git a/modules/sd_hijack_inpainting.py b/modules/sd_hijack_inpainting.py index 938f9a58..5018b047 100644 --- a/modules/sd_hijack_inpainting.py +++ b/modules/sd_hijack_inpainting.py @@ -324,12 +324,11 @@ def should_hijack_inpainting(checkpoint_info): def do_inpainting_hijack(): # most of this stuff seems to no longer be needed because it is already included into SD2.0 - # LatentInpaintDiffusion remains because SD2.0's LatentInpaintDiffusion can't be loaded without specifying a checkpoint # p_sample_plms is needed because PLMS can't work with dicts as conditionings # this file should be cleaned up later if weverything tuens out to work fine # ldm.models.diffusion.ddpm.get_unconditional_conditioning = get_unconditional_conditioning - ldm.models.diffusion.ddpm.LatentInpaintDiffusion = LatentInpaintDiffusion + # ldm.models.diffusion.ddpm.LatentInpaintDiffusion = LatentInpaintDiffusion # ldm.models.diffusion.ddim.DDIMSampler.p_sample_ddim = p_sample_ddim # ldm.models.diffusion.ddim.DDIMSampler.sample = sample_ddim diff --git a/modules/sd_models.py b/modules/sd_models.py index 5b37f3fe..b64f573f 100644 --- a/modules/sd_models.py +++ b/modules/sd_models.py @@ -296,6 +296,7 @@ def load_model(checkpoint_info=None): sd_config.model.params.use_ema = False sd_config.model.params.conditioning_key = "hybrid" sd_config.model.params.unet_config.params.in_channels = 9 + sd_config.model.params.finetune_keys = None # Create a "fake" config with a different name so that we know to unload it when switching models. checkpoint_info = checkpoint_info._replace(config=checkpoint_info.config.replace(".yaml", "-inpainting.yaml")) -- cgit v1.2.3 From 59c6511494c55a578eecdf71fb4590b6bd5d04a7 Mon Sep 17 00:00:00 2001 From: Dean van Dugteren <31391056+deanpress@users.noreply.github.com> Date: Sun, 11 Dec 2022 17:08:51 +0100 Subject: fix: fallback model_checkpoint if it's empty This fixes the following error when SD attempts to start with a deleted checkpoint: ``` Traceback (most recent call last): File "D:\Web\stable-diffusion-webui\launch.py", line 295, in start() File "D:\Web\stable-diffusion-webui\launch.py", line 290, in start webui.webui() File "D:\Web\stable-diffusion-webui\webui.py", line 132, in webui initialize() File "D:\Web\stable-diffusion-webui\webui.py", line 62, in initialize modules.sd_models.load_model() File "D:\Web\stable-diffusion-webui\modules\sd_models.py", line 283, in load_model checkpoint_info = checkpoint_info or select_checkpoint() File "D:\Web\stable-diffusion-webui\modules\sd_models.py", line 117, in select_checkpoint checkpoint_info = checkpoints_list.get(model_checkpoint, None) TypeError: unhashable type: 'list' ``` --- modules/sd_models.py | 4 ++++ 1 file changed, 4 insertions(+) (limited to 'modules/sd_models.py') diff --git a/modules/sd_models.py b/modules/sd_models.py index 5b37f3fe..b6d75db7 100644 --- a/modules/sd_models.py +++ b/modules/sd_models.py @@ -111,6 +111,10 @@ def model_hash(filename): def select_checkpoint(): model_checkpoint = shared.opts.sd_model_checkpoint + + if len(model_checkpoint) == 0: + model_checkpoint = shared.default_sd_model_file + checkpoint_info = checkpoints_list.get(model_checkpoint, None) if checkpoint_info is not None: return checkpoint_info -- cgit v1.2.3 From ec0a48826fb41c1b1baab45a9030f7eb55568fd0 Mon Sep 17 00:00:00 2001 From: MrCheeze Date: Sun, 11 Dec 2022 10:19:46 -0500 Subject: unconditionally set use_ema=False if value not specified (True never worked, and all configs except v1-inpainting-inference.yaml already correctly set it to False) --- modules/sd_models.py | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-) (limited to 'modules/sd_models.py') diff --git a/modules/sd_models.py b/modules/sd_models.py index b64f573f..f36b299f 100644 --- a/modules/sd_models.py +++ b/modules/sd_models.py @@ -293,7 +293,6 @@ def load_model(checkpoint_info=None): if should_hijack_inpainting(checkpoint_info): # Hardcoded config for now... sd_config.model.target = "ldm.models.diffusion.ddpm.LatentInpaintDiffusion" - sd_config.model.params.use_ema = False sd_config.model.params.conditioning_key = "hybrid" sd_config.model.params.unet_config.params.in_channels = 9 sd_config.model.params.finetune_keys = None @@ -301,6 +300,9 @@ def load_model(checkpoint_info=None): # Create a "fake" config with a different name so that we know to unload it when switching models. checkpoint_info = checkpoint_info._replace(config=checkpoint_info.config.replace(".yaml", "-inpainting.yaml")) + if not hasattr(sd_config.model.params, "use_ema"): + sd_config.model.params.use_ema = False + do_inpainting_hijack() if shared.cmd_opts.no_half: -- cgit v1.2.3 From 5a650055de3792223a91925aba8130ebdee29e35 Mon Sep 17 00:00:00 2001 From: "linuxmobile ( リナックス )" Date: Sat, 24 Dec 2022 09:25:35 -0300 Subject: Removed lenght in sd_model at line 115 Commit eba60a4 is what is causing this error, delete the length check in sd_model starting at line 115 and it's fine. https://github.com/AUTOMATIC1111/stable-diffusion-webui/issues/5971#issuecomment-1364507379 --- modules/sd_models.py | 3 --- 1 file changed, 3 deletions(-) (limited to 'modules/sd_models.py') diff --git a/modules/sd_models.py b/modules/sd_models.py index 1254e5ae..6ca06211 100644 --- a/modules/sd_models.py +++ b/modules/sd_models.py @@ -111,9 +111,6 @@ def model_hash(filename): def select_checkpoint(): model_checkpoint = shared.opts.sd_model_checkpoint - - if len(model_checkpoint) == 0: - model_checkpoint = shared.default_sd_model_file checkpoint_info = checkpoints_list.get(model_checkpoint, None) if checkpoint_info is not None: -- cgit v1.2.3 From 3bf5591efe9a9f219c6088be322a87adc4f48f95 Mon Sep 17 00:00:00 2001 From: Yuval Aboulafia Date: Sat, 24 Dec 2022 21:35:29 +0200 Subject: fix F541 f-string without any placeholders --- extensions-builtin/LDSR/ldsr_model_arch.py | 2 +- modules/codeformer/vqgan_arch.py | 4 ++-- modules/hypernetworks/hypernetwork.py | 4 ++-- modules/images.py | 2 +- modules/interrogate.py | 2 +- modules/safe.py | 8 ++++---- modules/sd_models.py | 8 ++++---- modules/sd_vae.py | 2 +- modules/textual_inversion/textual_inversion.py | 2 +- scripts/prompts_from_file.py | 2 +- 10 files changed, 18 insertions(+), 18 deletions(-) (limited to 'modules/sd_models.py') diff --git a/extensions-builtin/LDSR/ldsr_model_arch.py b/extensions-builtin/LDSR/ldsr_model_arch.py index f5bd8ae4..0ad49f4e 100644 --- a/extensions-builtin/LDSR/ldsr_model_arch.py +++ b/extensions-builtin/LDSR/ldsr_model_arch.py @@ -26,7 +26,7 @@ class LDSR: global cached_ldsr_model if shared.opts.ldsr_cached and cached_ldsr_model is not None: - print(f"Loading model from cache") + print("Loading model from cache") model: torch.nn.Module = cached_ldsr_model else: print(f"Loading model from {self.modelPath}") diff --git a/modules/codeformer/vqgan_arch.py b/modules/codeformer/vqgan_arch.py index c06c590c..e7293683 100644 --- a/modules/codeformer/vqgan_arch.py +++ b/modules/codeformer/vqgan_arch.py @@ -382,7 +382,7 @@ class VQAutoEncoder(nn.Module): self.load_state_dict(torch.load(model_path, map_location='cpu')['params']) logger.info(f'vqgan is loaded from: {model_path} [params]') else: - raise ValueError(f'Wrong params!') + raise ValueError('Wrong params!') def forward(self, x): @@ -431,7 +431,7 @@ class VQGANDiscriminator(nn.Module): elif 'params' in chkpt: self.load_state_dict(torch.load(model_path, map_location='cpu')['params']) else: - raise ValueError(f'Wrong params!') + raise ValueError('Wrong params!') def forward(self, x): return self.main(x) \ No newline at end of file diff --git a/modules/hypernetworks/hypernetwork.py b/modules/hypernetworks/hypernetwork.py index c406ffb3..9d3034ae 100644 --- a/modules/hypernetworks/hypernetwork.py +++ b/modules/hypernetworks/hypernetwork.py @@ -277,7 +277,7 @@ def load_hypernetwork(filename): print(traceback.format_exc(), file=sys.stderr) else: if shared.loaded_hypernetwork is not None: - print(f"Unloading hypernetwork") + print("Unloading hypernetwork") shared.loaded_hypernetwork = None @@ -417,7 +417,7 @@ def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, gradient_step, initial_step = hypernetwork.step or 0 if initial_step >= steps: - shared.state.textinfo = f"Model has already been trained beyond specified max steps" + shared.state.textinfo = "Model has already been trained beyond specified max steps" return hypernetwork, filename scheduler = LearnRateScheduler(learn_rate, steps, initial_step) diff --git a/modules/images.py b/modules/images.py index 809ad9f7..31d4528d 100644 --- a/modules/images.py +++ b/modules/images.py @@ -599,7 +599,7 @@ def read_info_from_image(image): Negative prompt: {json_info["uc"]} Steps: {json_info["steps"]}, Sampler: {sampler}, CFG scale: {json_info["scale"]}, Seed: {json_info["seed"]}, Size: {image.width}x{image.height}, Clip skip: 2, ENSD: 31337""" except Exception: - print(f"Error parsing NovelAI image generation parameters:", file=sys.stderr) + print("Error parsing NovelAI image generation parameters:", file=sys.stderr) print(traceback.format_exc(), file=sys.stderr) return geninfo, items diff --git a/modules/interrogate.py b/modules/interrogate.py index 0068b81c..46935210 100644 --- a/modules/interrogate.py +++ b/modules/interrogate.py @@ -172,7 +172,7 @@ class InterrogateModels: res += ", " + match except Exception: - print(f"Error interrogating", file=sys.stderr) + print("Error interrogating", file=sys.stderr) print(traceback.format_exc(), file=sys.stderr) res += "" diff --git a/modules/safe.py b/modules/safe.py index 479c8b86..1d4c20b9 100644 --- a/modules/safe.py +++ b/modules/safe.py @@ -137,15 +137,15 @@ def load_with_extra(filename, extra_handler=None, *args, **kwargs): except pickle.UnpicklingError: print(f"Error verifying pickled file from {filename}:", file=sys.stderr) print(traceback.format_exc(), file=sys.stderr) - print(f"-----> !!!! The file is most likely corrupted !!!! <-----", file=sys.stderr) - print(f"You can skip this check with --disable-safe-unpickle commandline argument, but that is not going to help you.\n\n", file=sys.stderr) + print("-----> !!!! The file is most likely corrupted !!!! <-----", file=sys.stderr) + print("You can skip this check with --disable-safe-unpickle commandline argument, but that is not going to help you.\n\n", file=sys.stderr) return None except Exception: print(f"Error verifying pickled file from {filename}:", file=sys.stderr) print(traceback.format_exc(), file=sys.stderr) - print(f"\nThe file may be malicious, so the program is not going to read it.", file=sys.stderr) - print(f"You can skip this check with --disable-safe-unpickle commandline argument.\n\n", file=sys.stderr) + print("\nThe file may be malicious, so the program is not going to read it.", file=sys.stderr) + print("You can skip this check with --disable-safe-unpickle commandline argument.\n\n", file=sys.stderr) return None return unsafe_torch_load(filename, *args, **kwargs) diff --git a/modules/sd_models.py b/modules/sd_models.py index 6ca06211..ecdd91c5 100644 --- a/modules/sd_models.py +++ b/modules/sd_models.py @@ -117,13 +117,13 @@ def select_checkpoint(): return checkpoint_info if len(checkpoints_list) == 0: - print(f"No checkpoints found. When searching for checkpoints, looked at:", file=sys.stderr) + print("No checkpoints found. When searching for checkpoints, looked at:", file=sys.stderr) if shared.cmd_opts.ckpt is not None: print(f" - file {os.path.abspath(shared.cmd_opts.ckpt)}", file=sys.stderr) print(f" - directory {model_path}", file=sys.stderr) if shared.cmd_opts.ckpt_dir is not None: print(f" - directory {os.path.abspath(shared.cmd_opts.ckpt_dir)}", file=sys.stderr) - print(f"Can't run without a checkpoint. Find and place a .ckpt file into any of those locations. The program will exit.", file=sys.stderr) + print("Can't run without a checkpoint. Find and place a .ckpt file into any of those locations. The program will exit.", file=sys.stderr) exit(1) checkpoint_info = next(iter(checkpoints_list.values())) @@ -324,7 +324,7 @@ def load_model(checkpoint_info=None): script_callbacks.model_loaded_callback(sd_model) - print(f"Model loaded.") + print("Model loaded.") return sd_model @@ -359,5 +359,5 @@ def reload_model_weights(sd_model=None, info=None): if not shared.cmd_opts.lowvram and not shared.cmd_opts.medvram: sd_model.to(devices.device) - print(f"Weights loaded.") + print("Weights loaded.") return sd_model diff --git a/modules/sd_vae.py b/modules/sd_vae.py index 25638a83..3856418e 100644 --- a/modules/sd_vae.py +++ b/modules/sd_vae.py @@ -208,5 +208,5 @@ def reload_vae_weights(sd_model=None, vae_file="auto"): if not shared.cmd_opts.lowvram and not shared.cmd_opts.medvram: sd_model.to(devices.device) - print(f"VAE Weights loaded.") + print("VAE Weights loaded.") return sd_model diff --git a/modules/textual_inversion/textual_inversion.py b/modules/textual_inversion/textual_inversion.py index daf3997b..f6112578 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -263,7 +263,7 @@ def train_embedding(embedding_name, learn_rate, batch_size, gradient_step, data_ initial_step = embedding.step or 0 if initial_step >= steps: - shared.state.textinfo = f"Model has already been trained beyond specified max steps" + shared.state.textinfo = "Model has already been trained beyond specified max steps" return embedding, filename scheduler = LearnRateScheduler(learn_rate, steps, initial_step) diff --git a/scripts/prompts_from_file.py b/scripts/prompts_from_file.py index 6e118ddb..e8386ed2 100644 --- a/scripts/prompts_from_file.py +++ b/scripts/prompts_from_file.py @@ -140,7 +140,7 @@ class Script(scripts.Script): try: args = cmdargs(line) except Exception: - print(f"Error parsing line [line] as commandline:", file=sys.stderr) + print(f"Error parsing line {line} as commandline:", file=sys.stderr) print(traceback.format_exc(), file=sys.stderr) args = {"prompt": line} else: -- cgit v1.2.3 From 5ba04f9ec050a66e918571f07e8863f157f05b44 Mon Sep 17 00:00:00 2001 From: Nicolas Patry Date: Wed, 21 Dec 2022 13:45:58 +0100 Subject: Attempting to solve slow loads for `safetensors`. Fixes #5893 --- modules/sd_models.py | 5 ++++- 1 file changed, 4 insertions(+), 1 deletion(-) (limited to 'modules/sd_models.py') diff --git a/modules/sd_models.py b/modules/sd_models.py index ecdd91c5..cd938656 100644 --- a/modules/sd_models.py +++ b/modules/sd_models.py @@ -168,7 +168,10 @@ def get_state_dict_from_checkpoint(pl_sd): def read_state_dict(checkpoint_file, print_global_state=False, map_location=None): _, extension = os.path.splitext(checkpoint_file) if extension.lower() == ".safetensors": - pl_sd = safetensors.torch.load_file(checkpoint_file, device=map_location or shared.weight_load_location) + device = map_location or shared.weight_load_location + if device is None: + device = "cuda:0" if torch.cuda.is_available() else "cpu" + pl_sd = safetensors.torch.load_file(checkpoint_file, device=device) else: pl_sd = torch.load(checkpoint_file, map_location=map_location or shared.weight_load_location) -- cgit v1.2.3 From f55ac33d446185680604e872ceda2ae858821d5c Mon Sep 17 00:00:00 2001 From: Vladimir Mandic Date: Sat, 31 Dec 2022 11:27:02 -0500 Subject: validate textual inversion embeddings --- modules/sd_models.py | 3 ++ modules/textual_inversion/textual_inversion.py | 43 +++++++++++++++++++++++--- modules/ui.py | 2 -- 3 files changed, 41 insertions(+), 7 deletions(-) (limited to 'modules/sd_models.py') diff --git a/modules/sd_models.py b/modules/sd_models.py index ecdd91c5..ebd4dff7 100644 --- a/modules/sd_models.py +++ b/modules/sd_models.py @@ -325,6 +325,9 @@ def load_model(checkpoint_info=None): script_callbacks.model_loaded_callback(sd_model) print("Model loaded.") + + sd_hijack.model_hijack.embedding_db.load_textual_inversion_embeddings(force_reload = True) # Reload embeddings after model load as they may or may not fit the model + return sd_model diff --git a/modules/textual_inversion/textual_inversion.py b/modules/textual_inversion/textual_inversion.py index f6112578..103ace60 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -23,6 +23,8 @@ class Embedding: self.vec = vec self.name = name self.step = step + self.shape = None + self.vectors = 0 self.cached_checksum = None self.sd_checkpoint = None self.sd_checkpoint_name = None @@ -57,8 +59,10 @@ class EmbeddingDatabase: def __init__(self, embeddings_dir): self.ids_lookup = {} self.word_embeddings = {} + self.skipped_embeddings = [] self.dir_mtime = None self.embeddings_dir = embeddings_dir + self.expected_shape = -1 def register_embedding(self, embedding, model): @@ -75,14 +79,35 @@ class EmbeddingDatabase: return embedding - def load_textual_inversion_embeddings(self): + def get_expected_shape(self): + expected_shape = -1 # initialize with unknown + idx = torch.tensor(0).to(shared.device) + if expected_shape == -1: + try: # matches sd15 signature + first_embedding = shared.sd_model.cond_stage_model.wrapped.transformer.text_model.embeddings.token_embedding.wrapped(idx) + expected_shape = first_embedding.shape[0] + except: + pass + if expected_shape == -1: + try: # matches sd20 signature + first_embedding = shared.sd_model.cond_stage_model.wrapped.model.token_embedding.wrapped(idx) + expected_shape = first_embedding.shape[0] + except: + pass + if expected_shape == -1: + print('Could not determine expected embeddings shape from model') + return expected_shape + + def load_textual_inversion_embeddings(self, force_reload = False): mt = os.path.getmtime(self.embeddings_dir) - if self.dir_mtime is not None and mt <= self.dir_mtime: + if not force_reload and self.dir_mtime is not None and mt <= self.dir_mtime: return self.dir_mtime = mt self.ids_lookup.clear() self.word_embeddings.clear() + self.skipped_embeddings = [] + self.expected_shape = self.get_expected_shape() def process_file(path, filename): name = os.path.splitext(filename)[0] @@ -122,7 +147,14 @@ class EmbeddingDatabase: embedding.step = data.get('step', None) embedding.sd_checkpoint = data.get('sd_checkpoint', None) embedding.sd_checkpoint_name = data.get('sd_checkpoint_name', None) - self.register_embedding(embedding, shared.sd_model) + embedding.vectors = vec.shape[0] + embedding.shape = vec.shape[-1] + + if (self.expected_shape == -1) or (self.expected_shape == embedding.shape): + self.register_embedding(embedding, shared.sd_model) + else: + self.skipped_embeddings.append(name) + # print('Skipping embedding {name}: shape was {shape} expected {expected}'.format(name = name, shape = embedding.shape, expected = self.expected_shape)) for fn in os.listdir(self.embeddings_dir): try: @@ -137,8 +169,9 @@ class EmbeddingDatabase: print(traceback.format_exc(), file=sys.stderr) continue - print(f"Loaded a total of {len(self.word_embeddings)} textual inversion embeddings.") - print("Embeddings:", ', '.join(self.word_embeddings.keys())) + print("Textual inversion embeddings {num} loaded: {val}".format(num = len(self.word_embeddings), val = ', '.join(self.word_embeddings.keys()))) + if (len(self.skipped_embeddings) > 0): + print("Textual inversion embeddings {num} skipped: {val}".format(num = len(self.skipped_embeddings), val = ', '.join(self.skipped_embeddings))) def find_embedding_at_position(self, tokens, offset): token = tokens[offset] diff --git a/modules/ui.py b/modules/ui.py index 57ee0465..397dd804 100644 --- a/modules/ui.py +++ b/modules/ui.py @@ -1157,8 +1157,6 @@ def create_ui(): with gr.Column(variant='panel'): submit_result = gr.Textbox(elem_id="modelmerger_result", show_label=False) - sd_hijack.model_hijack.embedding_db.load_textual_inversion_embeddings() - with gr.Blocks(analytics_enabled=False) as train_interface: with gr.Row().style(equal_height=False): gr.HTML(value="

See wiki for detailed explanation.

") -- cgit v1.2.3 From 311354c0bb8930ea939d6aa6b3edd50c69301320 Mon Sep 17 00:00:00 2001 From: AUTOMATIC <16777216c@gmail.com> Date: Mon, 2 Jan 2023 00:38:09 +0300 Subject: fix the issue with training on SD2.0 --- modules/sd_models.py | 2 ++ modules/textual_inversion/textual_inversion.py | 3 +-- 2 files changed, 3 insertions(+), 2 deletions(-) (limited to 'modules/sd_models.py') diff --git a/modules/sd_models.py b/modules/sd_models.py index ebd4dff7..bff8d6c9 100644 --- a/modules/sd_models.py +++ b/modules/sd_models.py @@ -228,6 +228,8 @@ def load_model_weights(model, checkpoint_info, vae_file="auto"): model.sd_model_checkpoint = checkpoint_file model.sd_checkpoint_info = checkpoint_info + model.logvar = model.logvar.to(devices.device) # fix for training + sd_vae.delete_base_vae() sd_vae.clear_loaded_vae() vae_file = sd_vae.resolve_vae(checkpoint_file, vae_file=vae_file) diff --git a/modules/textual_inversion/textual_inversion.py b/modules/textual_inversion/textual_inversion.py index 66f40367..1e5722e7 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -282,7 +282,7 @@ def train_embedding(embedding_name, learn_rate, batch_size, gradient_step, data_ return embedding, filename scheduler = LearnRateScheduler(learn_rate, steps, initial_step) - # dataset loading may take a while, so input validations and early returns should be done before this + # dataset loading may take a while, so input validations and early returns should be done before this shared.state.textinfo = f"Preparing dataset from {html.escape(data_root)}..." old_parallel_processing_allowed = shared.parallel_processing_allowed @@ -310,7 +310,6 @@ def train_embedding(embedding_name, learn_rate, batch_size, gradient_step, data_ loss_step = 0 _loss_step = 0 #internal - last_saved_file = "" last_saved_image = "" forced_filename = "" -- cgit v1.2.3 From 8f96f9289981a66741ba770d14f3d27ce335a0fb Mon Sep 17 00:00:00 2001 From: AUTOMATIC <16777216c@gmail.com> Date: Tue, 3 Jan 2023 18:39:14 +0300 Subject: call script callbacks for reloaded model after loading embeddings --- modules/sd_models.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) (limited to 'modules/sd_models.py') diff --git a/modules/sd_models.py b/modules/sd_models.py index bff8d6c9..b98b05fc 100644 --- a/modules/sd_models.py +++ b/modules/sd_models.py @@ -324,12 +324,12 @@ def load_model(checkpoint_info=None): sd_model.eval() shared.sd_model = sd_model + sd_hijack.model_hijack.embedding_db.load_textual_inversion_embeddings(force_reload=True) # Reload embeddings after model load as they may or may not fit the model + script_callbacks.model_loaded_callback(sd_model) print("Model loaded.") - sd_hijack.model_hijack.embedding_db.load_textual_inversion_embeddings(force_reload = True) # Reload embeddings after model load as they may or may not fit the model - return sd_model -- cgit v1.2.3 From 02d7abf5141431b9a3a8a189bb3136c71abd5e79 Mon Sep 17 00:00:00 2001 From: AUTOMATIC <16777216c@gmail.com> Date: Wed, 4 Jan 2023 12:35:07 +0300 Subject: helpful error message when trying to load 2.0 without config failing to load model weights from settings won't break generation for currently loaded model anymore --- modules/errors.py | 25 +++++++++++++++++++++++-- modules/sd_models.py | 26 ++++++++++++++++++-------- modules/shared.py | 9 +++++++-- webui.py | 12 ++++++++++-- 4 files changed, 58 insertions(+), 14 deletions(-) (limited to 'modules/sd_models.py') diff --git a/modules/errors.py b/modules/errors.py index 372dc51a..a668c014 100644 --- a/modules/errors.py +++ b/modules/errors.py @@ -2,9 +2,30 @@ import sys import traceback +def print_error_explanation(message): + lines = message.strip().split("\n") + max_len = max([len(x) for x in lines]) + + print('=' * max_len, file=sys.stderr) + for line in lines: + print(line, file=sys.stderr) + print('=' * max_len, file=sys.stderr) + + +def display(e: Exception, task): + print(f"{task or 'error'}: {type(e).__name__}", file=sys.stderr) + print(traceback.format_exc(), file=sys.stderr) + + message = str(e) + if "copying a param with shape torch.Size([640, 1024]) from checkpoint, the shape in current model is torch.Size([640, 768])" in message: + print_error_explanation(""" +The most likely cause of this is you are trying to load Stable Diffusion 2.0 model without specifying its connfig file. +See https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features#stable-diffusion-20 for how to solve this. + """) + + def run(code, task): try: code() except Exception as e: - print(f"{task}: {type(e).__name__}", file=sys.stderr) - print(traceback.format_exc(), file=sys.stderr) + display(task, e) diff --git a/modules/sd_models.py b/modules/sd_models.py index b98b05fc..6846b74a 100644 --- a/modules/sd_models.py +++ b/modules/sd_models.py @@ -278,6 +278,7 @@ def enable_midas_autodownload(): midas.api.load_model = load_model_wrapper + def load_model(checkpoint_info=None): from modules import lowvram, sd_hijack checkpoint_info = checkpoint_info or select_checkpoint() @@ -312,6 +313,7 @@ def load_model(checkpoint_info=None): sd_config.model.params.unet_config.params.use_fp16 = False sd_model = instantiate_from_config(sd_config.model) + load_model_weights(sd_model, checkpoint_info) if shared.cmd_opts.lowvram or shared.cmd_opts.medvram: @@ -336,10 +338,12 @@ def load_model(checkpoint_info=None): def reload_model_weights(sd_model=None, info=None): from modules import lowvram, devices, sd_hijack checkpoint_info = info or select_checkpoint() - + if not sd_model: sd_model = shared.sd_model + current_checkpoint_info = sd_model.sd_checkpoint_info + if sd_model.sd_model_checkpoint == checkpoint_info.filename: return @@ -356,13 +360,19 @@ def reload_model_weights(sd_model=None, info=None): sd_hijack.model_hijack.undo_hijack(sd_model) - load_model_weights(sd_model, checkpoint_info) - - sd_hijack.model_hijack.hijack(sd_model) - script_callbacks.model_loaded_callback(sd_model) - - if not shared.cmd_opts.lowvram and not shared.cmd_opts.medvram: - sd_model.to(devices.device) + try: + load_model_weights(sd_model, checkpoint_info) + except Exception as e: + print("Failed to load checkpoint, restoring previous") + load_model_weights(sd_model, current_checkpoint_info) + raise + finally: + sd_hijack.model_hijack.hijack(sd_model) + script_callbacks.model_loaded_callback(sd_model) + + if not shared.cmd_opts.lowvram and not shared.cmd_opts.medvram: + sd_model.to(devices.device) print("Weights loaded.") + return sd_model diff --git a/modules/shared.py b/modules/shared.py index 23657a93..7588c47b 100644 --- a/modules/shared.py +++ b/modules/shared.py @@ -14,7 +14,7 @@ import modules.interrogate import modules.memmon import modules.styles import modules.devices as devices -from modules import localization, sd_vae, extensions, script_loading +from modules import localization, sd_vae, extensions, script_loading, errors from modules.paths import models_path, script_path, sd_path @@ -494,7 +494,12 @@ class Options: return False if self.data_labels[key].onchange is not None: - self.data_labels[key].onchange() + try: + self.data_labels[key].onchange() + except Exception as e: + errors.display(e, f"changing setting {key} to {value}") + setattr(self, key, oldval) + return False return True diff --git a/webui.py b/webui.py index c7d55a97..13375e71 100644 --- a/webui.py +++ b/webui.py @@ -9,7 +9,7 @@ from fastapi import FastAPI from fastapi.middleware.cors import CORSMiddleware from fastapi.middleware.gzip import GZipMiddleware -from modules import import_hook +from modules import import_hook, errors from modules.call_queue import wrap_queued_call, queue_lock, wrap_gradio_gpu_call from modules.paths import script_path @@ -61,7 +61,15 @@ def initialize(): modelloader.load_upscalers() modules.sd_vae.refresh_vae_list() - modules.sd_models.load_model() + + try: + modules.sd_models.load_model() + except Exception as e: + errors.display(e, "loading stable diffusion model") + print("", file=sys.stderr) + print("Stable diffusion model failed to load, exiting", file=sys.stderr) + exit(1) + shared.opts.onchange("sd_model_checkpoint", wrap_queued_call(lambda: modules.sd_models.reload_model_weights())) shared.opts.onchange("sd_vae", wrap_queued_call(lambda: modules.sd_vae.reload_vae_weights()), call=False) shared.opts.onchange("sd_vae_as_default", wrap_queued_call(lambda: modules.sd_vae.reload_vae_weights()), call=False) -- cgit v1.2.3 From 8d8a05a3bbb50fdfeab51679a919d2487bd97976 Mon Sep 17 00:00:00 2001 From: AUTOMATIC <16777216c@gmail.com> Date: Wed, 4 Jan 2023 12:47:42 +0300 Subject: find configs for models at runtime rather than when starting --- modules/sd_hijack_inpainting.py | 5 ++++- modules/sd_models.py | 31 ++++++++++++++++++------------- 2 files changed, 22 insertions(+), 14 deletions(-) (limited to 'modules/sd_models.py') diff --git a/modules/sd_hijack_inpainting.py b/modules/sd_hijack_inpainting.py index 3c214a35..31d2c898 100644 --- a/modules/sd_hijack_inpainting.py +++ b/modules/sd_hijack_inpainting.py @@ -97,8 +97,11 @@ def p_sample_plms(self, x, c, t, index, repeat_noise=False, use_original_steps=F def should_hijack_inpainting(checkpoint_info): + from modules import sd_models + ckpt_basename = os.path.basename(checkpoint_info.filename).lower() - cfg_basename = os.path.basename(checkpoint_info.config).lower() + cfg_basename = os.path.basename(sd_models.find_checkpoint_config(checkpoint_info)).lower() + return "inpainting" in ckpt_basename and not "inpainting" in cfg_basename diff --git a/modules/sd_models.py b/modules/sd_models.py index 6846b74a..6dca4ddf 100644 --- a/modules/sd_models.py +++ b/modules/sd_models.py @@ -20,7 +20,7 @@ from modules.sd_hijack_inpainting import do_inpainting_hijack, should_hijack_inp model_dir = "Stable-diffusion" model_path = os.path.abspath(os.path.join(models_path, model_dir)) -CheckpointInfo = namedtuple("CheckpointInfo", ['filename', 'title', 'hash', 'model_name', 'config']) +CheckpointInfo = namedtuple("CheckpointInfo", ['filename', 'title', 'hash', 'model_name']) checkpoints_list = {} checkpoints_loaded = collections.OrderedDict() @@ -48,6 +48,14 @@ def checkpoint_tiles(): return sorted([x.title for x in checkpoints_list.values()], key = alphanumeric_key) +def find_checkpoint_config(info): + config = os.path.splitext(info.filename)[0] + ".yaml" + if os.path.exists(config): + return config + + return shared.cmd_opts.config + + def list_models(): checkpoints_list.clear() model_list = modelloader.load_models(model_path=model_path, command_path=shared.cmd_opts.ckpt_dir, ext_filter=[".ckpt", ".safetensors"]) @@ -73,7 +81,7 @@ def list_models(): if os.path.exists(cmd_ckpt): h = model_hash(cmd_ckpt) title, short_model_name = modeltitle(cmd_ckpt, h) - checkpoints_list[title] = CheckpointInfo(cmd_ckpt, title, h, short_model_name, shared.cmd_opts.config) + checkpoints_list[title] = CheckpointInfo(cmd_ckpt, title, h, short_model_name) shared.opts.data['sd_model_checkpoint'] = title elif cmd_ckpt is not None and cmd_ckpt != shared.default_sd_model_file: print(f"Checkpoint in --ckpt argument not found (Possible it was moved to {model_path}: {cmd_ckpt}", file=sys.stderr) @@ -81,12 +89,7 @@ def list_models(): h = model_hash(filename) title, short_model_name = modeltitle(filename, h) - basename, _ = os.path.splitext(filename) - config = basename + ".yaml" - if not os.path.exists(config): - config = shared.cmd_opts.config - - checkpoints_list[title] = CheckpointInfo(filename, title, h, short_model_name, config) + checkpoints_list[title] = CheckpointInfo(filename, title, h, short_model_name) def get_closet_checkpoint_match(searchString): @@ -282,9 +285,10 @@ def enable_midas_autodownload(): def load_model(checkpoint_info=None): from modules import lowvram, sd_hijack checkpoint_info = checkpoint_info or select_checkpoint() + checkpoint_config = find_checkpoint_config(checkpoint_info) - if checkpoint_info.config != shared.cmd_opts.config: - print(f"Loading config from: {checkpoint_info.config}") + if checkpoint_config != shared.cmd_opts.config: + print(f"Loading config from: {checkpoint_config}") if shared.sd_model: sd_hijack.model_hijack.undo_hijack(shared.sd_model) @@ -292,7 +296,7 @@ def load_model(checkpoint_info=None): gc.collect() devices.torch_gc() - sd_config = OmegaConf.load(checkpoint_info.config) + sd_config = OmegaConf.load(checkpoint_config) if should_hijack_inpainting(checkpoint_info): # Hardcoded config for now... @@ -302,7 +306,7 @@ def load_model(checkpoint_info=None): sd_config.model.params.finetune_keys = None # Create a "fake" config with a different name so that we know to unload it when switching models. - checkpoint_info = checkpoint_info._replace(config=checkpoint_info.config.replace(".yaml", "-inpainting.yaml")) + checkpoint_info = checkpoint_info._replace(config=checkpoint_config.replace(".yaml", "-inpainting.yaml")) if not hasattr(sd_config.model.params, "use_ema"): sd_config.model.params.use_ema = False @@ -343,11 +347,12 @@ def reload_model_weights(sd_model=None, info=None): sd_model = shared.sd_model current_checkpoint_info = sd_model.sd_checkpoint_info + checkpoint_config = find_checkpoint_config(current_checkpoint_info) if sd_model.sd_model_checkpoint == checkpoint_info.filename: return - if sd_model.sd_checkpoint_info.config != checkpoint_info.config or should_hijack_inpainting(checkpoint_info) != should_hijack_inpainting(sd_model.sd_checkpoint_info): + if checkpoint_config != find_checkpoint_config(checkpoint_info) or should_hijack_inpainting(checkpoint_info) != should_hijack_inpainting(sd_model.sd_checkpoint_info): del sd_model checkpoints_loaded.clear() load_model(checkpoint_info) -- cgit v1.2.3 From 0cd6399b8b1699b8b7acad6f0ad2988111fe618e Mon Sep 17 00:00:00 2001 From: AUTOMATIC <16777216c@gmail.com> Date: Wed, 4 Jan 2023 14:29:13 +0300 Subject: fix broken inpainting model --- modules/sd_models.py | 3 --- 1 file changed, 3 deletions(-) (limited to 'modules/sd_models.py') diff --git a/modules/sd_models.py b/modules/sd_models.py index 6dca4ddf..a568823d 100644 --- a/modules/sd_models.py +++ b/modules/sd_models.py @@ -305,9 +305,6 @@ def load_model(checkpoint_info=None): sd_config.model.params.unet_config.params.in_channels = 9 sd_config.model.params.finetune_keys = None - # Create a "fake" config with a different name so that we know to unload it when switching models. - checkpoint_info = checkpoint_info._replace(config=checkpoint_config.replace(".yaml", "-inpainting.yaml")) - if not hasattr(sd_config.model.params, "use_ema"): sd_config.model.params.use_ema = False -- cgit v1.2.3 From 642142556d8ecdea9beb86d7618b628b1803ab98 Mon Sep 17 00:00:00 2001 From: AUTOMATIC <16777216c@gmail.com> Date: Wed, 4 Jan 2023 15:09:53 +0300 Subject: use commandline-supplied cuda device name instead of cuda:0 for safetensors PR that doesn't fix anything --- modules/sd_models.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) (limited to 'modules/sd_models.py') diff --git a/modules/sd_models.py b/modules/sd_models.py index ee918f24..76a89e88 100644 --- a/modules/sd_models.py +++ b/modules/sd_models.py @@ -173,7 +173,7 @@ def read_state_dict(checkpoint_file, print_global_state=False, map_location=None if extension.lower() == ".safetensors": device = map_location or shared.weight_load_location if device is None: - device = "cuda:0" if torch.cuda.is_available() else "cpu" + device = devices.get_cuda_device_string() if torch.cuda.is_available() else "cpu" pl_sd = safetensors.torch.load_file(checkpoint_file, device=device) else: pl_sd = torch.load(checkpoint_file, map_location=map_location or shared.weight_load_location) -- cgit v1.2.3 From 552d7b90bf483c160cd20740f7acd7fccbc02e6f Mon Sep 17 00:00:00 2001 From: Vladimir Mandic Date: Mon, 9 Jan 2023 18:34:26 -0500 Subject: allow model load if previous model failed --- modules/sd_models.py | 15 ++++++++++----- 1 file changed, 10 insertions(+), 5 deletions(-) (limited to 'modules/sd_models.py') diff --git a/modules/sd_models.py b/modules/sd_models.py index 76a89e88..0a6d55ca 100644 --- a/modules/sd_models.py +++ b/modules/sd_models.py @@ -49,6 +49,9 @@ def checkpoint_tiles(): def find_checkpoint_config(info): + if info is None: + return shared.cmd_opts.config + config = os.path.splitext(info.filename)[0] + ".yaml" if os.path.exists(config): return config @@ -345,14 +348,16 @@ def reload_model_weights(sd_model=None, info=None): if not sd_model: sd_model = shared.sd_model + if sd_model is None: # previous model load failed + current_checkpoint_info = None + else: + current_checkpoint_info = sd_model.sd_checkpoint_info + if sd_model.sd_model_checkpoint == checkpoint_info.filename: + return - current_checkpoint_info = sd_model.sd_checkpoint_info checkpoint_config = find_checkpoint_config(current_checkpoint_info) - if sd_model.sd_model_checkpoint == checkpoint_info.filename: - return - - if checkpoint_config != find_checkpoint_config(checkpoint_info) or should_hijack_inpainting(checkpoint_info) != should_hijack_inpainting(sd_model.sd_checkpoint_info): + if current_checkpoint_info is None or checkpoint_config != find_checkpoint_config(checkpoint_info) or should_hijack_inpainting(checkpoint_info) != should_hijack_inpainting(sd_model.sd_checkpoint_info): del sd_model checkpoints_loaded.clear() load_model(checkpoint_info) -- cgit v1.2.3 From 0c3feb202c5714abd50d879c1db2cd9a71ce93e3 Mon Sep 17 00:00:00 2001 From: AUTOMATIC <16777216c@gmail.com> Date: Tue, 10 Jan 2023 14:08:29 +0300 Subject: disable torch weight initialization and CLIP downloading/reading checkpoint to speedup creating sd model from config --- modules/sd_disable_initialization.py | 44 ++++++++++++++++++++++++++++++++++++ modules/sd_models.py | 5 ++-- 2 files changed, 47 insertions(+), 2 deletions(-) create mode 100644 modules/sd_disable_initialization.py (limited to 'modules/sd_models.py') diff --git a/modules/sd_disable_initialization.py b/modules/sd_disable_initialization.py new file mode 100644 index 00000000..c9a3b5e4 --- /dev/null +++ b/modules/sd_disable_initialization.py @@ -0,0 +1,44 @@ +import ldm.modules.encoders.modules +import open_clip +import torch + + +class DisableInitialization: + """ + When an object of this class enters a `with` block, it starts preventing torch's layer initialization + functions from working, and changes CLIP and OpenCLIP to not download model weights. When it leaves, + reverts everything to how it was. + + Use like this: + ``` + with DisableInitialization(): + do_things() + ``` + """ + + def __enter__(self): + def do_nothing(*args, **kwargs): + pass + + def create_model_and_transforms_without_pretrained(*args, pretrained=None, **kwargs): + return self.create_model_and_transforms(*args, pretrained=None, **kwargs) + + def CLIPTextModel_from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs): + return self.CLIPTextModel_from_pretrained(None, *model_args, config=pretrained_model_name_or_path, state_dict={}, **kwargs) + + self.init_kaiming_uniform = torch.nn.init.kaiming_uniform_ + self.init_no_grad_normal = torch.nn.init._no_grad_normal_ + self.create_model_and_transforms = open_clip.create_model_and_transforms + self.CLIPTextModel_from_pretrained = ldm.modules.encoders.modules.CLIPTextModel.from_pretrained + + torch.nn.init.kaiming_uniform_ = do_nothing + torch.nn.init._no_grad_normal_ = do_nothing + open_clip.create_model_and_transforms = create_model_and_transforms_without_pretrained + ldm.modules.encoders.modules.CLIPTextModel.from_pretrained = CLIPTextModel_from_pretrained + + def __exit__(self, exc_type, exc_val, exc_tb): + torch.nn.init.kaiming_uniform_ = self.init_kaiming_uniform + torch.nn.init._no_grad_normal_ = self.init_no_grad_normal + open_clip.create_model_and_transforms = self.create_model_and_transforms + ldm.modules.encoders.modules.CLIPTextModel.from_pretrained = self.CLIPTextModel_from_pretrained + diff --git a/modules/sd_models.py b/modules/sd_models.py index 0a6d55ca..ee241032 100644 --- a/modules/sd_models.py +++ b/modules/sd_models.py @@ -13,7 +13,7 @@ import ldm.modules.midas as midas from ldm.util import instantiate_from_config -from modules import shared, modelloader, devices, script_callbacks, sd_vae +from modules import shared, modelloader, devices, script_callbacks, sd_vae, sd_disable_initialization from modules.paths import models_path from modules.sd_hijack_inpainting import do_inpainting_hijack, should_hijack_inpainting @@ -319,7 +319,8 @@ def load_model(checkpoint_info=None): if shared.cmd_opts.no_half: sd_config.model.params.unet_config.params.use_fp16 = False - sd_model = instantiate_from_config(sd_config.model) + with sd_disable_initialization.DisableInitialization(): + sd_model = instantiate_from_config(sd_config.model) load_model_weights(sd_model, checkpoint_info) -- cgit v1.2.3 From ce3f639ec8758ce2bc90483336361d2dc25acd3a Mon Sep 17 00:00:00 2001 From: AUTOMATIC <16777216c@gmail.com> Date: Tue, 10 Jan 2023 16:51:04 +0300 Subject: add more stuff to ignore when creating model from config prevent .vae.safetensors files from being listed as stable diffusion models --- modules/modelloader.py | 4 +++- modules/sd_disable_initialization.py | 29 +++++++++++++++++++++++++---- modules/sd_models.py | 32 ++++++++++++++++++++++++++++---- 3 files changed, 56 insertions(+), 9 deletions(-) (limited to 'modules/sd_models.py') diff --git a/modules/modelloader.py b/modules/modelloader.py index 6a1a7ac8..e9aa514e 100644 --- a/modules/modelloader.py +++ b/modules/modelloader.py @@ -10,7 +10,7 @@ from modules.upscaler import Upscaler from modules.paths import script_path, models_path -def load_models(model_path: str, model_url: str = None, command_path: str = None, ext_filter=None, download_name=None) -> list: +def load_models(model_path: str, model_url: str = None, command_path: str = None, ext_filter=None, download_name=None, ext_blacklist=None) -> list: """ A one-and done loader to try finding the desired models in specified directories. @@ -45,6 +45,8 @@ def load_models(model_path: str, model_url: str = None, command_path: str = None full_path = file if os.path.isdir(full_path): continue + if ext_blacklist is not None and any([full_path.endswith(x) for x in ext_blacklist]): + continue if len(ext_filter) != 0: model_name, extension = os.path.splitext(file) if extension not in ext_filter: diff --git a/modules/sd_disable_initialization.py b/modules/sd_disable_initialization.py index c9a3b5e4..9942bd7e 100644 --- a/modules/sd_disable_initialization.py +++ b/modules/sd_disable_initialization.py @@ -1,15 +1,19 @@ import ldm.modules.encoders.modules import open_clip import torch +import transformers.utils.hub class DisableInitialization: """ - When an object of this class enters a `with` block, it starts preventing torch's layer initialization - functions from working, and changes CLIP and OpenCLIP to not download model weights. When it leaves, - reverts everything to how it was. + When an object of this class enters a `with` block, it starts: + - preventing torch's layer initialization functions from working + - changes CLIP and OpenCLIP to not download model weights + - changes CLIP to not make requests to check if there is a new version of a file you already have - Use like this: + When it leaves the block, it reverts everything to how it was before. + + Use it like this: ``` with DisableInitialization(): do_things() @@ -26,19 +30,36 @@ class DisableInitialization: def CLIPTextModel_from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs): return self.CLIPTextModel_from_pretrained(None, *model_args, config=pretrained_model_name_or_path, state_dict={}, **kwargs) + def transformers_utils_hub_get_from_cache(url, *args, local_files_only=False, **kwargs): + + # this file is always 404, prevent making request + if url == 'https://huggingface.co/openai/clip-vit-large-patch14/resolve/main/added_tokens.json': + raise transformers.utils.hub.EntryNotFoundError + + try: + return self.transformers_utils_hub_get_from_cache(url, *args, local_files_only=True, **kwargs) + except Exception as e: + return self.transformers_utils_hub_get_from_cache(url, *args, local_files_only=False, **kwargs) + self.init_kaiming_uniform = torch.nn.init.kaiming_uniform_ self.init_no_grad_normal = torch.nn.init._no_grad_normal_ + self.init_no_grad_uniform_ = torch.nn.init._no_grad_uniform_ self.create_model_and_transforms = open_clip.create_model_and_transforms self.CLIPTextModel_from_pretrained = ldm.modules.encoders.modules.CLIPTextModel.from_pretrained + self.transformers_utils_hub_get_from_cache = transformers.utils.hub.get_from_cache torch.nn.init.kaiming_uniform_ = do_nothing torch.nn.init._no_grad_normal_ = do_nothing + torch.nn.init._no_grad_uniform_ = do_nothing open_clip.create_model_and_transforms = create_model_and_transforms_without_pretrained ldm.modules.encoders.modules.CLIPTextModel.from_pretrained = CLIPTextModel_from_pretrained + transformers.utils.hub.get_from_cache = transformers_utils_hub_get_from_cache def __exit__(self, exc_type, exc_val, exc_tb): torch.nn.init.kaiming_uniform_ = self.init_kaiming_uniform torch.nn.init._no_grad_normal_ = self.init_no_grad_normal + torch.nn.init._no_grad_uniform_ = self.init_no_grad_uniform_ open_clip.create_model_and_transforms = self.create_model_and_transforms ldm.modules.encoders.modules.CLIPTextModel.from_pretrained = self.CLIPTextModel_from_pretrained + transformers.utils.hub.get_from_cache = self.transformers_utils_hub_get_from_cache diff --git a/modules/sd_models.py b/modules/sd_models.py index ee241032..1bb9088b 100644 --- a/modules/sd_models.py +++ b/modules/sd_models.py @@ -2,6 +2,7 @@ import collections import os.path import sys import gc +import time from collections import namedtuple import torch import re @@ -61,7 +62,7 @@ def find_checkpoint_config(info): def list_models(): checkpoints_list.clear() - model_list = modelloader.load_models(model_path=model_path, command_path=shared.cmd_opts.ckpt_dir, ext_filter=[".ckpt", ".safetensors"]) + model_list = modelloader.load_models(model_path=model_path, command_path=shared.cmd_opts.ckpt_dir, ext_filter=[".ckpt", ".safetensors"], ext_blacklist=[".vae.safetensors"]) def modeltitle(path, shorthash): abspath = os.path.abspath(path) @@ -288,6 +289,17 @@ def enable_midas_autodownload(): midas.api.load_model = load_model_wrapper +class Timer: + def __init__(self): + self.start = time.time() + + def elapsed(self): + end = time.time() + res = end - self.start + self.start = end + return res + + def load_model(checkpoint_info=None): from modules import lowvram, sd_hijack checkpoint_info = checkpoint_info or select_checkpoint() @@ -319,11 +331,17 @@ def load_model(checkpoint_info=None): if shared.cmd_opts.no_half: sd_config.model.params.unet_config.params.use_fp16 = False + timer = Timer() + with sd_disable_initialization.DisableInitialization(): sd_model = instantiate_from_config(sd_config.model) + elapsed_create = timer.elapsed() + load_model_weights(sd_model, checkpoint_info) + elapsed_load_weights = timer.elapsed() + if shared.cmd_opts.lowvram or shared.cmd_opts.medvram: lowvram.setup_for_low_vram(sd_model, shared.cmd_opts.medvram) else: @@ -338,7 +356,9 @@ def load_model(checkpoint_info=None): script_callbacks.model_loaded_callback(sd_model) - print("Model loaded.") + elapsed_the_rest = timer.elapsed() + + print(f"Model loaded in {elapsed_create + elapsed_load_weights + elapsed_the_rest:.1f}s ({elapsed_create:.1f}s create model, {elapsed_load_weights:.1f}s load weights).") return sd_model @@ -349,7 +369,7 @@ def reload_model_weights(sd_model=None, info=None): if not sd_model: sd_model = shared.sd_model - if sd_model is None: # previous model load failed + if sd_model is None: # previous model load failed current_checkpoint_info = None else: current_checkpoint_info = sd_model.sd_checkpoint_info @@ -371,6 +391,8 @@ def reload_model_weights(sd_model=None, info=None): sd_hijack.model_hijack.undo_hijack(sd_model) + timer = Timer() + try: load_model_weights(sd_model, checkpoint_info) except Exception as e: @@ -384,6 +406,8 @@ def reload_model_weights(sd_model=None, info=None): if not shared.cmd_opts.lowvram and not shared.cmd_opts.medvram: sd_model.to(devices.device) - print("Weights loaded.") + elapsed = timer.elapsed() + + print(f"Weights loaded in {elapsed:.1f}s.") return sd_model -- cgit v1.2.3 From 0f8603a55988d22616b17140e6c4a7e9d0736af5 Mon Sep 17 00:00:00 2001 From: AUTOMATIC <16777216c@gmail.com> Date: Tue, 10 Jan 2023 17:46:59 +0300 Subject: add support for transformers==4.25.1 add fallback for when quick model creation fails --- modules/sd_disable_initialization.py | 42 ++++++++++++++++++++++++++++++------ modules/sd_models.py | 8 +++++-- 2 files changed, 42 insertions(+), 8 deletions(-) (limited to 'modules/sd_models.py') diff --git a/modules/sd_disable_initialization.py b/modules/sd_disable_initialization.py index 9942bd7e..088ac24b 100644 --- a/modules/sd_disable_initialization.py +++ b/modules/sd_disable_initialization.py @@ -30,30 +30,53 @@ class DisableInitialization: def CLIPTextModel_from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs): return self.CLIPTextModel_from_pretrained(None, *model_args, config=pretrained_model_name_or_path, state_dict={}, **kwargs) - def transformers_utils_hub_get_from_cache(url, *args, local_files_only=False, **kwargs): + def transformers_modeling_utils_load_pretrained_model(*args, **kwargs): + args = args[0:3] + ('/', ) + args[4:] # resolved_archive_file; must set it to something to prevent what seems to be a bug + return self.transformers_modeling_utils_load_pretrained_model(*args, **kwargs) + + def transformers_utils_hub_get_file_from_cache(original, url, *args, **kwargs): # this file is always 404, prevent making request if url == 'https://huggingface.co/openai/clip-vit-large-patch14/resolve/main/added_tokens.json': raise transformers.utils.hub.EntryNotFoundError try: - return self.transformers_utils_hub_get_from_cache(url, *args, local_files_only=True, **kwargs) + return original(url, *args, local_files_only=True, **kwargs) except Exception as e: - return self.transformers_utils_hub_get_from_cache(url, *args, local_files_only=False, **kwargs) + return original(url, *args, local_files_only=False, **kwargs) + + def transformers_utils_hub_get_from_cache(url, *args, local_files_only=False, **kwargs): + return transformers_utils_hub_get_file_from_cache(self.transformers_utils_hub_get_from_cache, url, *args, **kwargs) + + def transformers_tokenization_utils_base_cached_file(url, *args, local_files_only=False, **kwargs): + return transformers_utils_hub_get_file_from_cache(self.transformers_tokenization_utils_base_cached_file, url, *args, **kwargs) + + def transformers_configuration_utils_cached_file(url, *args, local_files_only=False, **kwargs): + return transformers_utils_hub_get_file_from_cache(self.transformers_configuration_utils_cached_file, url, *args, **kwargs) self.init_kaiming_uniform = torch.nn.init.kaiming_uniform_ self.init_no_grad_normal = torch.nn.init._no_grad_normal_ self.init_no_grad_uniform_ = torch.nn.init._no_grad_uniform_ self.create_model_and_transforms = open_clip.create_model_and_transforms self.CLIPTextModel_from_pretrained = ldm.modules.encoders.modules.CLIPTextModel.from_pretrained - self.transformers_utils_hub_get_from_cache = transformers.utils.hub.get_from_cache + self.transformers_modeling_utils_load_pretrained_model = getattr(transformers.modeling_utils.PreTrainedModel, '_load_pretrained_model', None) + self.transformers_tokenization_utils_base_cached_file = getattr(transformers.tokenization_utils_base, 'cached_file', None) + self.transformers_configuration_utils_cached_file = getattr(transformers.configuration_utils, 'cached_file', None) + self.transformers_utils_hub_get_from_cache = getattr(transformers.utils.hub, 'get_from_cache', None) torch.nn.init.kaiming_uniform_ = do_nothing torch.nn.init._no_grad_normal_ = do_nothing torch.nn.init._no_grad_uniform_ = do_nothing open_clip.create_model_and_transforms = create_model_and_transforms_without_pretrained ldm.modules.encoders.modules.CLIPTextModel.from_pretrained = CLIPTextModel_from_pretrained - transformers.utils.hub.get_from_cache = transformers_utils_hub_get_from_cache + if self.transformers_modeling_utils_load_pretrained_model is not None: + transformers.modeling_utils.PreTrainedModel._load_pretrained_model = transformers_modeling_utils_load_pretrained_model + if self.transformers_tokenization_utils_base_cached_file is not None: + transformers.tokenization_utils_base.cached_file = transformers_tokenization_utils_base_cached_file + if self.transformers_configuration_utils_cached_file is not None: + transformers.configuration_utils.cached_file = transformers_configuration_utils_cached_file + if self.transformers_utils_hub_get_from_cache is not None: + transformers.utils.hub.get_from_cache = transformers_utils_hub_get_from_cache def __exit__(self, exc_type, exc_val, exc_tb): torch.nn.init.kaiming_uniform_ = self.init_kaiming_uniform @@ -61,5 +84,12 @@ class DisableInitialization: torch.nn.init._no_grad_uniform_ = self.init_no_grad_uniform_ open_clip.create_model_and_transforms = self.create_model_and_transforms ldm.modules.encoders.modules.CLIPTextModel.from_pretrained = self.CLIPTextModel_from_pretrained - transformers.utils.hub.get_from_cache = self.transformers_utils_hub_get_from_cache + if self.transformers_modeling_utils_load_pretrained_model is not None: + transformers.modeling_utils.PreTrainedModel._load_pretrained_model = self.transformers_modeling_utils_load_pretrained_model + if self.transformers_tokenization_utils_base_cached_file is not None: + transformers.utils.hub.cached_file = self.transformers_tokenization_utils_base_cached_file + if self.transformers_configuration_utils_cached_file is not None: + transformers.utils.hub.cached_file = self.transformers_configuration_utils_cached_file + if self.transformers_utils_hub_get_from_cache is not None: + transformers.utils.hub.get_from_cache = self.transformers_utils_hub_get_from_cache diff --git a/modules/sd_models.py b/modules/sd_models.py index 1bb9088b..b5bc12f0 100644 --- a/modules/sd_models.py +++ b/modules/sd_models.py @@ -14,7 +14,7 @@ import ldm.modules.midas as midas from ldm.util import instantiate_from_config -from modules import shared, modelloader, devices, script_callbacks, sd_vae, sd_disable_initialization +from modules import shared, modelloader, devices, script_callbacks, sd_vae, sd_disable_initialization, errors from modules.paths import models_path from modules.sd_hijack_inpainting import do_inpainting_hijack, should_hijack_inpainting @@ -333,7 +333,11 @@ def load_model(checkpoint_info=None): timer = Timer() - with sd_disable_initialization.DisableInitialization(): + try: + with sd_disable_initialization.DisableInitialization(): + sd_model = instantiate_from_config(sd_config.model) + except Exception as e: + print('Failed to create model quickly; will retry using slow method.', file=sys.stderr) sd_model = instantiate_from_config(sd_config.model) elapsed_create = timer.elapsed() -- cgit v1.2.3 From 4fdacd31e48c6a7a35c1c25c559932585e8addde Mon Sep 17 00:00:00 2001 From: AUTOMATIC <16777216c@gmail.com> Date: Wed, 11 Jan 2023 10:24:56 +0300 Subject: possible fix for fallback for fast model creation from config --- modules/sd_models.py | 3 +++ 1 file changed, 3 insertions(+) (limited to 'modules/sd_models.py') diff --git a/modules/sd_models.py b/modules/sd_models.py index b5bc12f0..a0a8a909 100644 --- a/modules/sd_models.py +++ b/modules/sd_models.py @@ -337,6 +337,9 @@ def load_model(checkpoint_info=None): with sd_disable_initialization.DisableInitialization(): sd_model = instantiate_from_config(sd_config.model) except Exception as e: + pass + + if sd_model is None: print('Failed to create model quickly; will retry using slow method.', file=sys.stderr) sd_model = instantiate_from_config(sd_config.model) -- cgit v1.2.3 From 1a23dc32ac5e16fac10115cafd0b841abd06e59f Mon Sep 17 00:00:00 2001 From: AUTOMATIC <16777216c@gmail.com> Date: Wed, 11 Jan 2023 10:34:36 +0300 Subject: possible fix for fallback for fast model creation from config, attempt 2 --- modules/sd_models.py | 1 + 1 file changed, 1 insertion(+) (limited to 'modules/sd_models.py') diff --git a/modules/sd_models.py b/modules/sd_models.py index a0a8a909..084ba7fa 100644 --- a/modules/sd_models.py +++ b/modules/sd_models.py @@ -333,6 +333,7 @@ def load_model(checkpoint_info=None): timer = Timer() + sd_model = None try: with sd_disable_initialization.DisableInitialization(): sd_model = instantiate_from_config(sd_config.model) -- cgit v1.2.3 From 4bd490727e156ff53107d53416d6b89be86f2a62 Mon Sep 17 00:00:00 2001 From: AUTOMATIC <16777216c@gmail.com> Date: Wed, 11 Jan 2023 18:54:04 +0300 Subject: fix for an error caused by skipping initialization, for realsies this time: TypeError: expected str, bytes or os.PathLike object, not NoneType --- modules/sd_disable_initialization.py | 71 ++++++++++++++++-------------------- modules/sd_models.py | 1 + 2 files changed, 33 insertions(+), 39 deletions(-) (limited to 'modules/sd_models.py') diff --git a/modules/sd_disable_initialization.py b/modules/sd_disable_initialization.py index 088ac24b..c72d8efc 100644 --- a/modules/sd_disable_initialization.py +++ b/modules/sd_disable_initialization.py @@ -20,6 +20,19 @@ class DisableInitialization: ``` """ + def __init__(self): + self.replaced = [] + + def replace(self, obj, field, func): + original = getattr(obj, field, None) + if original is None: + return None + + self.replaced.append((obj, field, original)) + setattr(obj, field, func) + + return original + def __enter__(self): def do_nothing(*args, **kwargs): pass @@ -37,11 +50,14 @@ class DisableInitialization: def transformers_utils_hub_get_file_from_cache(original, url, *args, **kwargs): # this file is always 404, prevent making request - if url == 'https://huggingface.co/openai/clip-vit-large-patch14/resolve/main/added_tokens.json': - raise transformers.utils.hub.EntryNotFoundError + if url == 'https://huggingface.co/openai/clip-vit-large-patch14/resolve/main/added_tokens.json' or url == 'openai/clip-vit-large-patch14' and args[0] == 'added_tokens.json': + return None try: - return original(url, *args, local_files_only=True, **kwargs) + res = original(url, *args, local_files_only=True, **kwargs) + if res is None: + res = original(url, *args, local_files_only=False, **kwargs) + return res except Exception as e: return original(url, *args, local_files_only=False, **kwargs) @@ -54,42 +70,19 @@ class DisableInitialization: def transformers_configuration_utils_cached_file(url, *args, local_files_only=False, **kwargs): return transformers_utils_hub_get_file_from_cache(self.transformers_configuration_utils_cached_file, url, *args, **kwargs) - self.init_kaiming_uniform = torch.nn.init.kaiming_uniform_ - self.init_no_grad_normal = torch.nn.init._no_grad_normal_ - self.init_no_grad_uniform_ = torch.nn.init._no_grad_uniform_ - self.create_model_and_transforms = open_clip.create_model_and_transforms - self.CLIPTextModel_from_pretrained = ldm.modules.encoders.modules.CLIPTextModel.from_pretrained - self.transformers_modeling_utils_load_pretrained_model = getattr(transformers.modeling_utils.PreTrainedModel, '_load_pretrained_model', None) - self.transformers_tokenization_utils_base_cached_file = getattr(transformers.tokenization_utils_base, 'cached_file', None) - self.transformers_configuration_utils_cached_file = getattr(transformers.configuration_utils, 'cached_file', None) - self.transformers_utils_hub_get_from_cache = getattr(transformers.utils.hub, 'get_from_cache', None) - - torch.nn.init.kaiming_uniform_ = do_nothing - torch.nn.init._no_grad_normal_ = do_nothing - torch.nn.init._no_grad_uniform_ = do_nothing - open_clip.create_model_and_transforms = create_model_and_transforms_without_pretrained - ldm.modules.encoders.modules.CLIPTextModel.from_pretrained = CLIPTextModel_from_pretrained - if self.transformers_modeling_utils_load_pretrained_model is not None: - transformers.modeling_utils.PreTrainedModel._load_pretrained_model = transformers_modeling_utils_load_pretrained_model - if self.transformers_tokenization_utils_base_cached_file is not None: - transformers.tokenization_utils_base.cached_file = transformers_tokenization_utils_base_cached_file - if self.transformers_configuration_utils_cached_file is not None: - transformers.configuration_utils.cached_file = transformers_configuration_utils_cached_file - if self.transformers_utils_hub_get_from_cache is not None: - transformers.utils.hub.get_from_cache = transformers_utils_hub_get_from_cache + self.replace(torch.nn.init, 'kaiming_uniform_', do_nothing) + self.replace(torch.nn.init, '_no_grad_normal_', do_nothing) + self.replace(torch.nn.init, '_no_grad_uniform_', do_nothing) + self.create_model_and_transforms = self.replace(open_clip, 'create_model_and_transforms', create_model_and_transforms_without_pretrained) + self.CLIPTextModel_from_pretrained = self.replace(ldm.modules.encoders.modules.CLIPTextModel, 'from_pretrained', CLIPTextModel_from_pretrained) + self.transformers_modeling_utils_load_pretrained_model = self.replace(transformers.modeling_utils.PreTrainedModel, '_load_pretrained_model', transformers_modeling_utils_load_pretrained_model) + self.transformers_tokenization_utils_base_cached_file = self.replace(transformers.tokenization_utils_base, 'cached_file', transformers_tokenization_utils_base_cached_file) + self.transformers_configuration_utils_cached_file = self.replace(transformers.configuration_utils, 'cached_file', transformers_configuration_utils_cached_file) + self.transformers_utils_hub_get_from_cache = self.replace(transformers.utils.hub, 'get_from_cache', transformers_utils_hub_get_from_cache) def __exit__(self, exc_type, exc_val, exc_tb): - torch.nn.init.kaiming_uniform_ = self.init_kaiming_uniform - torch.nn.init._no_grad_normal_ = self.init_no_grad_normal - torch.nn.init._no_grad_uniform_ = self.init_no_grad_uniform_ - open_clip.create_model_and_transforms = self.create_model_and_transforms - ldm.modules.encoders.modules.CLIPTextModel.from_pretrained = self.CLIPTextModel_from_pretrained - if self.transformers_modeling_utils_load_pretrained_model is not None: - transformers.modeling_utils.PreTrainedModel._load_pretrained_model = self.transformers_modeling_utils_load_pretrained_model - if self.transformers_tokenization_utils_base_cached_file is not None: - transformers.utils.hub.cached_file = self.transformers_tokenization_utils_base_cached_file - if self.transformers_configuration_utils_cached_file is not None: - transformers.utils.hub.cached_file = self.transformers_configuration_utils_cached_file - if self.transformers_utils_hub_get_from_cache is not None: - transformers.utils.hub.get_from_cache = self.transformers_utils_hub_get_from_cache + for obj, field, original in self.replaced: + setattr(obj, field, original) + + self.replaced.clear() diff --git a/modules/sd_models.py b/modules/sd_models.py index 084ba7fa..c466f273 100644 --- a/modules/sd_models.py +++ b/modules/sd_models.py @@ -334,6 +334,7 @@ def load_model(checkpoint_info=None): timer = Timer() sd_model = None + try: with sd_disable_initialization.DisableInitialization(): sd_model = instantiate_from_config(sd_config.model) -- cgit v1.2.3