From 4fafc34e498130dcbb2d1a44fbc55fdba31e32d4 Mon Sep 17 00:00:00 2001 From: catboxanon <122327233+catboxanon@users.noreply.github.com> Date: Thu, 10 Aug 2023 23:42:58 -0400 Subject: Fix to make LoRA old method setting work --- extensions-builtin/Lora/networks.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) (limited to 'extensions-builtin/Lora') diff --git a/extensions-builtin/Lora/networks.py b/extensions-builtin/Lora/networks.py index bc722e90..7e3415ac 100644 --- a/extensions-builtin/Lora/networks.py +++ b/extensions-builtin/Lora/networks.py @@ -357,7 +357,7 @@ def network_forward(module, input, original_forward): if module is None: continue - y = module.forward(y, input) + y = module.forward(input, y) return y -- cgit v1.2.3 From bd4da4474bef5c9c1f690c62b971704ee73d2860 Mon Sep 17 00:00:00 2001 From: Kohaku-Blueleaf <59680068+KohakuBlueleaf@users.noreply.github.com> Date: Sun, 13 Aug 2023 02:27:39 +0800 Subject: Add extra norm module into built-in lora ext refer to LyCORIS 1.9.0.dev6 add new option and module for training norm layer (Which is reported to be good for style) --- extensions-builtin/Lora/network.py | 7 ++- extensions-builtin/Lora/network_norm.py | 29 ++++++++++++ extensions-builtin/Lora/networks.py | 64 ++++++++++++++++++++++---- extensions-builtin/Lora/scripts/lora_script.py | 16 +++++++ 4 files changed, 105 insertions(+), 11 deletions(-) create mode 100644 extensions-builtin/Lora/network_norm.py (limited to 'extensions-builtin/Lora') diff --git a/extensions-builtin/Lora/network.py b/extensions-builtin/Lora/network.py index 0a18d69e..b7b89061 100644 --- a/extensions-builtin/Lora/network.py +++ b/extensions-builtin/Lora/network.py @@ -133,7 +133,7 @@ class NetworkModule: return 1.0 - def finalize_updown(self, updown, orig_weight, output_shape): + def finalize_updown(self, updown, orig_weight, output_shape, ex_bias=None): if self.bias is not None: updown = updown.reshape(self.bias.shape) updown += self.bias.to(orig_weight.device, dtype=orig_weight.dtype) @@ -145,7 +145,10 @@ class NetworkModule: if orig_weight.size().numel() == updown.size().numel(): updown = updown.reshape(orig_weight.shape) - return updown * self.calc_scale() * self.multiplier() + if ex_bias is None: + ex_bias = 0 + + return updown * self.calc_scale() * self.multiplier(), ex_bias * self.multiplier() def calc_updown(self, target): raise NotImplementedError() diff --git a/extensions-builtin/Lora/network_norm.py b/extensions-builtin/Lora/network_norm.py new file mode 100644 index 00000000..dab8b684 --- /dev/null +++ b/extensions-builtin/Lora/network_norm.py @@ -0,0 +1,29 @@ +import network + + +class ModuleTypeNorm(network.ModuleType): + def create_module(self, net: network.Network, weights: network.NetworkWeights): + if all(x in weights.w for x in ["w_norm", "b_norm"]): + return NetworkModuleNorm(net, weights) + + return None + + +class NetworkModuleNorm(network.NetworkModule): + def __init__(self, net: network.Network, weights: network.NetworkWeights): + super().__init__(net, weights) + print("NetworkModuleNorm") + + self.w_norm = weights.w.get("w_norm") + self.b_norm = weights.w.get("b_norm") + + def calc_updown(self, orig_weight): + output_shape = self.w_norm.shape + updown = self.w_norm.to(orig_weight.device, dtype=orig_weight.dtype) + + if self.b_norm is not None: + ex_bias = self.b_norm.to(orig_weight.device, dtype=orig_weight.dtype) + else: + ex_bias = None + + return self.finalize_updown(updown, orig_weight, output_shape, ex_bias) diff --git a/extensions-builtin/Lora/networks.py b/extensions-builtin/Lora/networks.py index 7e3415ac..74cefe43 100644 --- a/extensions-builtin/Lora/networks.py +++ b/extensions-builtin/Lora/networks.py @@ -7,6 +7,7 @@ import network_hada import network_ia3 import network_lokr import network_full +import network_norm import torch from typing import Union @@ -19,6 +20,7 @@ module_types = [ network_ia3.ModuleTypeIa3(), network_lokr.ModuleTypeLokr(), network_full.ModuleTypeFull(), + network_norm.ModuleTypeNorm(), ] @@ -31,6 +33,8 @@ suffix_conversion = { "resnets": { "conv1": "in_layers_2", "conv2": "out_layers_3", + "norm1": "in_layers_0", + "norm2": "out_layers_0", "time_emb_proj": "emb_layers_1", "conv_shortcut": "skip_connection", } @@ -258,20 +262,25 @@ def load_networks(names, te_multipliers=None, unet_multipliers=None, dyn_dims=No purge_networks_from_memory() -def network_restore_weights_from_backup(self: Union[torch.nn.Conv2d, torch.nn.Linear, torch.nn.MultiheadAttention]): +def network_restore_weights_from_backup(self: Union[torch.nn.Conv2d, torch.nn.Linear, torch.nn.GroupNorm, torch.nn.LayerNorm, torch.nn.MultiheadAttention]): weights_backup = getattr(self, "network_weights_backup", None) + bias_backup = getattr(self, "network_bias_backup", None) - if weights_backup is None: + if weights_backup is None and bias_backup is None: return - if isinstance(self, torch.nn.MultiheadAttention): - self.in_proj_weight.copy_(weights_backup[0]) - self.out_proj.weight.copy_(weights_backup[1]) - else: - self.weight.copy_(weights_backup) + if weights_backup is not None: + if isinstance(self, torch.nn.MultiheadAttention): + self.in_proj_weight.copy_(weights_backup[0]) + self.out_proj.weight.copy_(weights_backup[1]) + else: + self.weight.copy_(weights_backup) + if bias_backup is not None: + self.bias.copy_(bias_backup) -def network_apply_weights(self: Union[torch.nn.Conv2d, torch.nn.Linear, torch.nn.MultiheadAttention]): + +def network_apply_weights(self: Union[torch.nn.Conv2d, torch.nn.Linear, torch.nn.GroupNorm, torch.nn.LayerNorm, torch.nn.MultiheadAttention]): """ Applies the currently selected set of networks to the weights of torch layer self. If weights already have this particular set of networks applied, does nothing. @@ -294,6 +303,11 @@ def network_apply_weights(self: Union[torch.nn.Conv2d, torch.nn.Linear, torch.nn self.network_weights_backup = weights_backup + bias_backup = getattr(self, "network_bias_backup", None) + if bias_backup is None and getattr(self, 'bias', None) is not None: + bias_backup = self.bias.to(devices.cpu, copy=True) + self.network_bias_backup = bias_backup + if current_names != wanted_names: network_restore_weights_from_backup(self) @@ -301,13 +315,15 @@ def network_apply_weights(self: Union[torch.nn.Conv2d, torch.nn.Linear, torch.nn module = net.modules.get(network_layer_name, None) if module is not None and hasattr(self, 'weight'): with torch.no_grad(): - updown = module.calc_updown(self.weight) + updown, ex_bias = module.calc_updown(self.weight) if len(self.weight.shape) == 4 and self.weight.shape[1] == 9: # inpainting model. zero pad updown to make channel[1] 4 to 9 updown = torch.nn.functional.pad(updown, (0, 0, 0, 0, 0, 5)) self.weight += updown + if getattr(self, 'bias', None) is not None: + self.bias += ex_bias continue module_q = net.modules.get(network_layer_name + "_q_proj", None) @@ -397,6 +413,36 @@ def network_Conv2d_load_state_dict(self, *args, **kwargs): return torch.nn.Conv2d_load_state_dict_before_network(self, *args, **kwargs) +def network_GroupNorm_forward(self, input): + if shared.opts.lora_functional: + return network_forward(self, input, torch.nn.GroupNorm_forward_before_network) + + network_apply_weights(self) + + return torch.nn.GroupNorm_forward_before_network(self, input) + + +def network_GroupNorm_load_state_dict(self, *args, **kwargs): + network_reset_cached_weight(self) + + return torch.nn.GroupNorm_load_state_dict_before_network(self, *args, **kwargs) + + +def network_LayerNorm_forward(self, input): + if shared.opts.lora_functional: + return network_forward(self, input, torch.nn.LayerNorm_forward_before_network) + + network_apply_weights(self) + + return torch.nn.LayerNorm_forward_before_network(self, input) + + +def network_LayerNorm_load_state_dict(self, *args, **kwargs): + network_reset_cached_weight(self) + + return torch.nn.LayerNorm_load_state_dict_before_network(self, *args, **kwargs) + + def network_MultiheadAttention_forward(self, *args, **kwargs): network_apply_weights(self) diff --git a/extensions-builtin/Lora/scripts/lora_script.py b/extensions-builtin/Lora/scripts/lora_script.py index 6ab8b6e7..dc307f8c 100644 --- a/extensions-builtin/Lora/scripts/lora_script.py +++ b/extensions-builtin/Lora/scripts/lora_script.py @@ -40,6 +40,18 @@ if not hasattr(torch.nn, 'Conv2d_forward_before_network'): if not hasattr(torch.nn, 'Conv2d_load_state_dict_before_network'): torch.nn.Conv2d_load_state_dict_before_network = torch.nn.Conv2d._load_from_state_dict +if not hasattr(torch.nn, 'GroupNorm_forward_before_network'): + torch.nn.GroupNorm_forward_before_network = torch.nn.GroupNorm.forward + +if not hasattr(torch.nn, 'GroupNorm_load_state_dict_before_network'): + torch.nn.GroupNorm_load_state_dict_before_network = torch.nn.GroupNorm._load_from_state_dict + +if not hasattr(torch.nn, 'LayerNorm_forward_before_network'): + torch.nn.LayerNorm_forward_before_network = torch.nn.LayerNorm.forward + +if not hasattr(torch.nn, 'LayerNorm_load_state_dict_before_network'): + torch.nn.LayerNorm_load_state_dict_before_network = torch.nn.LayerNorm._load_from_state_dict + if not hasattr(torch.nn, 'MultiheadAttention_forward_before_network'): torch.nn.MultiheadAttention_forward_before_network = torch.nn.MultiheadAttention.forward @@ -50,6 +62,10 @@ torch.nn.Linear.forward = networks.network_Linear_forward torch.nn.Linear._load_from_state_dict = networks.network_Linear_load_state_dict torch.nn.Conv2d.forward = networks.network_Conv2d_forward torch.nn.Conv2d._load_from_state_dict = networks.network_Conv2d_load_state_dict +torch.nn.GroupNorm.forward = networks.network_GroupNorm_forward +torch.nn.GroupNorm._load_from_state_dict = networks.network_GroupNorm_load_state_dict +torch.nn.LayerNorm.forward = networks.network_LayerNorm_forward +torch.nn.LayerNorm._load_from_state_dict = networks.network_LayerNorm_load_state_dict torch.nn.MultiheadAttention.forward = networks.network_MultiheadAttention_forward torch.nn.MultiheadAttention._load_from_state_dict = networks.network_MultiheadAttention_load_state_dict -- cgit v1.2.3 From a2b83050965a1a117f2762d3b5fa8b4841777e8f Mon Sep 17 00:00:00 2001 From: Kohaku-Blueleaf <59680068+KohakuBlueleaf@users.noreply.github.com> Date: Sun, 13 Aug 2023 02:35:04 +0800 Subject: return None if no ex_bias --- extensions-builtin/Lora/network.py | 6 +++--- extensions-builtin/Lora/networks.py | 2 +- 2 files changed, 4 insertions(+), 4 deletions(-) (limited to 'extensions-builtin/Lora') diff --git a/extensions-builtin/Lora/network.py b/extensions-builtin/Lora/network.py index b7b89061..d8e8dfb7 100644 --- a/extensions-builtin/Lora/network.py +++ b/extensions-builtin/Lora/network.py @@ -145,10 +145,10 @@ class NetworkModule: if orig_weight.size().numel() == updown.size().numel(): updown = updown.reshape(orig_weight.shape) - if ex_bias is None: - ex_bias = 0 + if ex_bias is not None: + ex_bias = ex_bias * self.multiplier() - return updown * self.calc_scale() * self.multiplier(), ex_bias * self.multiplier() + return updown * self.calc_scale() * self.multiplier(), ex_bias def calc_updown(self, target): raise NotImplementedError() diff --git a/extensions-builtin/Lora/networks.py b/extensions-builtin/Lora/networks.py index 74cefe43..ba621139 100644 --- a/extensions-builtin/Lora/networks.py +++ b/extensions-builtin/Lora/networks.py @@ -322,7 +322,7 @@ def network_apply_weights(self: Union[torch.nn.Conv2d, torch.nn.Linear, torch.nn updown = torch.nn.functional.pad(updown, (0, 0, 0, 0, 0, 5)) self.weight += updown - if getattr(self, 'bias', None) is not None: + if ex_bias is not None and getattr(self, 'bias', None) is not None: self.bias += ex_bias continue -- cgit v1.2.3 From 5881dcb8873b3f87b9c6545e9cb8d1d77023f4fe Mon Sep 17 00:00:00 2001 From: Kohaku-Blueleaf <59680068+KohakuBlueleaf@users.noreply.github.com> Date: Sun, 13 Aug 2023 02:36:02 +0800 Subject: remove debug print --- extensions-builtin/Lora/network_norm.py | 1 - 1 file changed, 1 deletion(-) (limited to 'extensions-builtin/Lora') diff --git a/extensions-builtin/Lora/network_norm.py b/extensions-builtin/Lora/network_norm.py index dab8b684..ce450158 100644 --- a/extensions-builtin/Lora/network_norm.py +++ b/extensions-builtin/Lora/network_norm.py @@ -12,7 +12,6 @@ class ModuleTypeNorm(network.ModuleType): class NetworkModuleNorm(network.NetworkModule): def __init__(self, net: network.Network, weights: network.NetworkWeights): super().__init__(net, weights) - print("NetworkModuleNorm") self.w_norm = weights.w.get("w_norm") self.b_norm = weights.w.get("b_norm") -- cgit v1.2.3 From 7fa5ee54b15904bef6598800df76ba1291d44ec6 Mon Sep 17 00:00:00 2001 From: catboxanon <122327233+catboxanon@users.noreply.github.com> Date: Sun, 13 Aug 2023 02:32:54 -0400 Subject: Support search and display of hashes for all extra network items --- extensions-builtin/Lora/ui_extra_networks_lora.py | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) (limited to 'extensions-builtin/Lora') diff --git a/extensions-builtin/Lora/ui_extra_networks_lora.py b/extensions-builtin/Lora/ui_extra_networks_lora.py index 3629e5c0..55409a78 100644 --- a/extensions-builtin/Lora/ui_extra_networks_lora.py +++ b/extensions-builtin/Lora/ui_extra_networks_lora.py @@ -25,9 +25,10 @@ class ExtraNetworksPageLora(ui_extra_networks.ExtraNetworksPage): item = { "name": name, "filename": lora_on_disk.filename, + "shorthash": lora_on_disk.shorthash, "preview": self.find_preview(path), "description": self.find_description(path), - "search_term": self.search_terms_from_path(lora_on_disk.filename), + "search_term": self.search_terms_from_path(lora_on_disk.filename) + " " + (lora_on_disk.hash or ""), "local_preview": f"{path}.{shared.opts.samples_format}", "metadata": lora_on_disk.metadata, "sort_keys": {'default': index, **self.get_sort_keys(lora_on_disk.filename)}, -- cgit v1.2.3 From d8419762c1454ba51baa710d9ce8e762efc056ef Mon Sep 17 00:00:00 2001 From: AUTOMATIC1111 <16777216c@gmail.com> Date: Sun, 13 Aug 2023 15:07:37 +0300 Subject: Lora: output warnings in UI rather than fail for unfitting loras; switch to logging for error output in console --- extensions-builtin/Lora/extra_networks_lora.py | 12 ++++- extensions-builtin/Lora/networks.py | 61 ++++++++++++++++---------- extensions-builtin/Lora/scripts/lora_script.py | 6 +-- 3 files changed, 50 insertions(+), 29 deletions(-) (limited to 'extensions-builtin/Lora') diff --git a/extensions-builtin/Lora/extra_networks_lora.py b/extensions-builtin/Lora/extra_networks_lora.py index ba2945c6..32e32cab 100644 --- a/extensions-builtin/Lora/extra_networks_lora.py +++ b/extensions-builtin/Lora/extra_networks_lora.py @@ -1,4 +1,4 @@ -from modules import extra_networks, shared +from modules import extra_networks, shared, sd_hijack import networks @@ -6,9 +6,14 @@ class ExtraNetworkLora(extra_networks.ExtraNetwork): def __init__(self): super().__init__('lora') + self.errors = {} + """mapping of network names to the number of errors the network had during operation""" + def activate(self, p, params_list): additional = shared.opts.sd_lora + self.errors.clear() + if additional != "None" and additional in networks.available_networks and not any(x for x in params_list if x.items[0] == additional): p.all_prompts = [x + f"" for x in p.all_prompts] params_list.append(extra_networks.ExtraNetworkParams(items=[additional, shared.opts.extra_networks_default_multiplier])) @@ -56,4 +61,7 @@ class ExtraNetworkLora(extra_networks.ExtraNetwork): p.extra_generation_params["Lora hashes"] = ", ".join(network_hashes) def deactivate(self, p): - pass + if self.errors: + p.comment("Networks with errors: " + ", ".join(f"{k} ({v})" for k, v in self.errors.items())) + + self.errors.clear() diff --git a/extensions-builtin/Lora/networks.py b/extensions-builtin/Lora/networks.py index ba621139..c252ed9e 100644 --- a/extensions-builtin/Lora/networks.py +++ b/extensions-builtin/Lora/networks.py @@ -1,3 +1,4 @@ +import logging import os import re @@ -194,7 +195,7 @@ def load_network(name, network_on_disk): net.modules[key] = net_module if keys_failed_to_match: - print(f"Failed to match keys when loading network {network_on_disk.filename}: {keys_failed_to_match}") + logging.debug(f"Network {network_on_disk.filename} didn't match keys: {keys_failed_to_match}") return net @@ -207,7 +208,6 @@ def purge_networks_from_memory(): devices.torch_gc() - def load_networks(names, te_multipliers=None, unet_multipliers=None, dyn_dims=None): already_loaded = {} @@ -248,7 +248,7 @@ def load_networks(names, te_multipliers=None, unet_multipliers=None, dyn_dims=No if net is None: failed_to_load_networks.append(name) - print(f"Couldn't find network with name {name}") + logging.info(f"Couldn't find network with name {name}") continue net.te_multiplier = te_multipliers[i] if te_multipliers else 1.0 @@ -257,7 +257,7 @@ def load_networks(names, te_multipliers=None, unet_multipliers=None, dyn_dims=No loaded_networks.append(net) if failed_to_load_networks: - sd_hijack.model_hijack.comments.append("Failed to find networks: " + ", ".join(failed_to_load_networks)) + sd_hijack.model_hijack.comments.append("Networks not found: " + ", ".join(failed_to_load_networks)) purge_networks_from_memory() @@ -314,17 +314,22 @@ def network_apply_weights(self: Union[torch.nn.Conv2d, torch.nn.Linear, torch.nn for net in loaded_networks: module = net.modules.get(network_layer_name, None) if module is not None and hasattr(self, 'weight'): - with torch.no_grad(): - updown, ex_bias = module.calc_updown(self.weight) + try: + with torch.no_grad(): + updown, ex_bias = module.calc_updown(self.weight) - if len(self.weight.shape) == 4 and self.weight.shape[1] == 9: - # inpainting model. zero pad updown to make channel[1] 4 to 9 - updown = torch.nn.functional.pad(updown, (0, 0, 0, 0, 0, 5)) + if len(self.weight.shape) == 4 and self.weight.shape[1] == 9: + # inpainting model. zero pad updown to make channel[1] 4 to 9 + updown = torch.nn.functional.pad(updown, (0, 0, 0, 0, 0, 5)) - self.weight += updown - if ex_bias is not None and getattr(self, 'bias', None) is not None: - self.bias += ex_bias - continue + self.weight += updown + if ex_bias is not None and getattr(self, 'bias', None) is not None: + self.bias += ex_bias + except RuntimeError as e: + logging.debug(f"Network {net.name} layer {network_layer_name}: {e}") + extra_network_lora.errors[net.name] = extra_network_lora.errors.get(net.name, 0) + 1 + + continue module_q = net.modules.get(network_layer_name + "_q_proj", None) module_k = net.modules.get(network_layer_name + "_k_proj", None) @@ -332,21 +337,28 @@ def network_apply_weights(self: Union[torch.nn.Conv2d, torch.nn.Linear, torch.nn module_out = net.modules.get(network_layer_name + "_out_proj", None) if isinstance(self, torch.nn.MultiheadAttention) and module_q and module_k and module_v and module_out: - with torch.no_grad(): - updown_q = module_q.calc_updown(self.in_proj_weight) - updown_k = module_k.calc_updown(self.in_proj_weight) - updown_v = module_v.calc_updown(self.in_proj_weight) - updown_qkv = torch.vstack([updown_q, updown_k, updown_v]) - updown_out = module_out.calc_updown(self.out_proj.weight) - - self.in_proj_weight += updown_qkv - self.out_proj.weight += updown_out - continue + try: + with torch.no_grad(): + updown_q = module_q.calc_updown(self.in_proj_weight) + updown_k = module_k.calc_updown(self.in_proj_weight) + updown_v = module_v.calc_updown(self.in_proj_weight) + updown_qkv = torch.vstack([updown_q, updown_k, updown_v]) + updown_out = module_out.calc_updown(self.out_proj.weight) + + self.in_proj_weight += updown_qkv + self.out_proj.weight += updown_out + + except RuntimeError as e: + logging.debug(f"Network {net.name} layer {network_layer_name}: {e}") + extra_network_lora.errors[net.name] = extra_network_lora.errors.get(net.name, 0) + 1 + + continue if module is None: continue - print(f'failed to calculate network weights for layer {network_layer_name}') + logging.debug(f"Network {net.name} layer {network_layer_name}: couldn't find supported operation") + extra_network_lora.errors[net.name] = extra_network_lora.errors.get(net.name, 0) + 1 self.network_current_names = wanted_names @@ -519,6 +531,7 @@ def infotext_pasted(infotext, params): if added: params["Prompt"] += "\n" + "".join(added) +extra_network_lora = None available_networks = {} available_network_aliases = {} diff --git a/extensions-builtin/Lora/scripts/lora_script.py b/extensions-builtin/Lora/scripts/lora_script.py index dc307f8c..4c6e774a 100644 --- a/extensions-builtin/Lora/scripts/lora_script.py +++ b/extensions-builtin/Lora/scripts/lora_script.py @@ -23,9 +23,9 @@ def unload(): def before_ui(): ui_extra_networks.register_page(ui_extra_networks_lora.ExtraNetworksPageLora()) - extra_network = extra_networks_lora.ExtraNetworkLora() - extra_networks.register_extra_network(extra_network) - extra_networks.register_extra_network_alias(extra_network, "lyco") + networks.extra_network_lora = extra_networks_lora.ExtraNetworkLora() + extra_networks.register_extra_network(networks.extra_network_lora) + extra_networks.register_extra_network_alias(networks.extra_network_lora, "lyco") if not hasattr(torch.nn, 'Linear_forward_before_network'): -- cgit v1.2.3 From db40d26d086b6e980bd28ce7eb3008ec09b1f0e8 Mon Sep 17 00:00:00 2001 From: AUTOMATIC1111 <16777216c@gmail.com> Date: Sun, 13 Aug 2023 16:38:10 +0300 Subject: linter --- extensions-builtin/Lora/extra_networks_lora.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) (limited to 'extensions-builtin/Lora') diff --git a/extensions-builtin/Lora/extra_networks_lora.py b/extensions-builtin/Lora/extra_networks_lora.py index 32e32cab..005ff32c 100644 --- a/extensions-builtin/Lora/extra_networks_lora.py +++ b/extensions-builtin/Lora/extra_networks_lora.py @@ -1,4 +1,4 @@ -from modules import extra_networks, shared, sd_hijack +from modules import extra_networks, shared import networks -- cgit v1.2.3 From d9cc27cb29926c9cc5dce331da8fbaf996cf4973 Mon Sep 17 00:00:00 2001 From: Kohaku-Blueleaf <59680068+KohakuBlueleaf@users.noreply.github.com> Date: Mon, 14 Aug 2023 13:32:51 +0800 Subject: Fix MHA updown err and support ex-bias for no-bias layer --- extensions-builtin/Lora/networks.py | 37 +++++++++++++++++++++++++++++-------- 1 file changed, 29 insertions(+), 8 deletions(-) (limited to 'extensions-builtin/Lora') diff --git a/extensions-builtin/Lora/networks.py b/extensions-builtin/Lora/networks.py index ba621139..1645b822 100644 --- a/extensions-builtin/Lora/networks.py +++ b/extensions-builtin/Lora/networks.py @@ -277,7 +277,15 @@ def network_restore_weights_from_backup(self: Union[torch.nn.Conv2d, torch.nn.Li self.weight.copy_(weights_backup) if bias_backup is not None: - self.bias.copy_(bias_backup) + if isinstance(self, torch.nn.MultiheadAttention): + self.out_proj.bias.copy_(bias_backup) + else: + self.bias.copy_(bias_backup) + else: + if isinstance(self, torch.nn.MultiheadAttention): + self.out_proj.bias = None + else: + self.bias = None def network_apply_weights(self: Union[torch.nn.Conv2d, torch.nn.Linear, torch.nn.GroupNorm, torch.nn.LayerNorm, torch.nn.MultiheadAttention]): @@ -305,7 +313,12 @@ def network_apply_weights(self: Union[torch.nn.Conv2d, torch.nn.Linear, torch.nn bias_backup = getattr(self, "network_bias_backup", None) if bias_backup is None and getattr(self, 'bias', None) is not None: - bias_backup = self.bias.to(devices.cpu, copy=True) + if isinstance(self, torch.nn.MultiheadAttention) and self.out_proj.bias is not None: + bias_backup = self.out_proj.bias.to(devices.cpu, copy=True) + elif getattr(self, 'bias', None) is not None: + bias_backup = self.bias.to(devices.cpu, copy=True) + else: + bias_backup = None self.network_bias_backup = bias_backup if current_names != wanted_names: @@ -322,8 +335,11 @@ def network_apply_weights(self: Union[torch.nn.Conv2d, torch.nn.Linear, torch.nn updown = torch.nn.functional.pad(updown, (0, 0, 0, 0, 0, 5)) self.weight += updown - if ex_bias is not None and getattr(self, 'bias', None) is not None: - self.bias += ex_bias + if ex_bias is not None and hasattr(self, 'bias'): + if self.bias is None: + self.bias = torch.nn.Parameter(ex_bias) + else: + self.bias += ex_bias continue module_q = net.modules.get(network_layer_name + "_q_proj", None) @@ -333,14 +349,19 @@ def network_apply_weights(self: Union[torch.nn.Conv2d, torch.nn.Linear, torch.nn if isinstance(self, torch.nn.MultiheadAttention) and module_q and module_k and module_v and module_out: with torch.no_grad(): - updown_q = module_q.calc_updown(self.in_proj_weight) - updown_k = module_k.calc_updown(self.in_proj_weight) - updown_v = module_v.calc_updown(self.in_proj_weight) + updown_q, _ = module_q.calc_updown(self.in_proj_weight) + updown_k, _ = module_k.calc_updown(self.in_proj_weight) + updown_v, _ = module_v.calc_updown(self.in_proj_weight) updown_qkv = torch.vstack([updown_q, updown_k, updown_v]) - updown_out = module_out.calc_updown(self.out_proj.weight) + updown_out, ex_bias = module_out.calc_updown(self.out_proj.weight) self.in_proj_weight += updown_qkv self.out_proj.weight += updown_out + if ex_bias is not None: + if self.out_proj.bias is None: + self.out_proj.bias = torch.nn.Parameter(ex_bias) + else: + self.out_proj.bias += ex_bias continue if module is None: -- cgit v1.2.3 From f70ded89365f71d42b6a60a561e8fccfdd25c159 Mon Sep 17 00:00:00 2001 From: Kohaku-Blueleaf <59680068+KohakuBlueleaf@users.noreply.github.com> Date: Mon, 14 Aug 2023 13:53:40 +0800 Subject: remove "if bias exist" check --- extensions-builtin/Lora/networks.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) (limited to 'extensions-builtin/Lora') diff --git a/extensions-builtin/Lora/networks.py b/extensions-builtin/Lora/networks.py index 96d14344..22fdff4a 100644 --- a/extensions-builtin/Lora/networks.py +++ b/extensions-builtin/Lora/networks.py @@ -312,7 +312,7 @@ def network_apply_weights(self: Union[torch.nn.Conv2d, torch.nn.Linear, torch.nn self.network_weights_backup = weights_backup bias_backup = getattr(self, "network_bias_backup", None) - if bias_backup is None and getattr(self, 'bias', None) is not None: + if bias_backup is None: if isinstance(self, torch.nn.MultiheadAttention) and self.out_proj.bias is not None: bias_backup = self.out_proj.bias.to(devices.cpu, copy=True) elif getattr(self, 'bias', None) is not None: -- cgit v1.2.3