Merge branch 'master' into features-to-readme

2 files changed, 99 insertions, 25 deletions

diff --git a/modules/hypernetworks/hypernetwork.py b/modules/hypernetworks/hypernetwork.py
index 4905710e..b1a5d0c7 100644
--- a/modules/hypernetworks/hypernetwork.py
+++ b/modules/hypernetworks/hypernetwork.py
@@ -22,25 +22,67 @@ from modules.textual_inversion.learn_schedule import LearnRateScheduler
 class HypernetworkModule(torch.nn.Module):

     multiplier = 1.0

 

-    def __init__(self, dim, state_dict=None):

+    def __init__(self, dim, state_dict=None, layer_structure=None, add_layer_norm=False, activation_func=None):

         super().__init__()

 

-        self.linear1 = torch.nn.Linear(dim, dim * 2)

-        self.linear2 = torch.nn.Linear(dim * 2, dim)

+        assert layer_structure is not None, "layer_structure must not be None"

+        assert layer_structure[0] == 1, "Multiplier Sequence should start with size 1!"

+        assert layer_structure[-1] == 1, "Multiplier Sequence should end with size 1!"

+

+        linears = []

+        for i in range(len(layer_structure) - 1):

+            linears.append(torch.nn.Linear(int(dim * layer_structure[i]), int(dim * layer_structure[i+1])))

+

+            if activation_func == "relu":

+                linears.append(torch.nn.ReLU())

+            elif activation_func == "leakyrelu":

+                linears.append(torch.nn.LeakyReLU())

+            elif activation_func == 'linear' or activation_func is None:

+                pass

+            else:

+                raise RuntimeError(f'hypernetwork uses an unsupported activation function: {activation_func}')

+

+            if add_layer_norm:

+                linears.append(torch.nn.LayerNorm(int(dim * layer_structure[i+1])))

+

+        self.linear = torch.nn.Sequential(*linears)

 

         if state_dict is not None:

-            self.load_state_dict(state_dict, strict=True)

+            self.fix_old_state_dict(state_dict)

+            self.load_state_dict(state_dict)

         else:

-

-            self.linear1.weight.data.normal_(mean=0.0, std=0.01)

-            self.linear1.bias.data.zero_()

-            self.linear2.weight.data.normal_(mean=0.0, std=0.01)

-            self.linear2.bias.data.zero_()

+            for layer in self.linear:

+                if type(layer) == torch.nn.Linear or type(layer) == torch.nn.LayerNorm:

+                    layer.weight.data.normal_(mean=0.0, std=0.01)

+                    layer.bias.data.zero_()

 

         self.to(devices.device)

 

+    def fix_old_state_dict(self, state_dict):

+        changes = {

+            'linear1.bias': 'linear.0.bias',

+            'linear1.weight': 'linear.0.weight',

+            'linear2.bias': 'linear.1.bias',

+            'linear2.weight': 'linear.1.weight',

+        }

+

+        for fr, to in changes.items():

+            x = state_dict.get(fr, None)

+            if x is None:

+                continue

+

+            del state_dict[fr]

+            state_dict[to] = x

+

     def forward(self, x):

-        return x + (self.linear2(self.linear1(x))) * self.multiplier

+        return x + self.linear(x) * self.multiplier

+

+    def trainables(self):

+        layer_structure = []

+        for layer in self.linear:

+            if type(layer) == torch.nn.Linear or type(layer) == torch.nn.LayerNorm:

+                layer_structure += [layer.weight, layer.bias]

+        return layer_structure

 

 

 def apply_strength(value=None):

@@ -51,16 +93,22 @@ class Hypernetwork:
     filename = None

     name = None

 

-    def __init__(self, name=None, enable_sizes=None):

+    def __init__(self, name=None, enable_sizes=None, layer_structure=None, add_layer_norm=False, activation_func=None):

         self.filename = None

         self.name = name

         self.layers = {}

         self.step = 0

         self.sd_checkpoint = None

         self.sd_checkpoint_name = None

+        self.layer_structure = layer_structure

+        self.add_layer_norm = add_layer_norm

+        self.activation_func = activation_func

 

         for size in enable_sizes or []:

-            self.layers[size] = (HypernetworkModule(size), HypernetworkModule(size))

+            self.layers[size] = (

+                HypernetworkModule(size, None, self.layer_structure, self.add_layer_norm, self.activation_func),

+                HypernetworkModule(size, None, self.layer_structure, self.add_layer_norm, self.activation_func),

+            )

 

     def weights(self):

         res = []

@@ -68,7 +116,7 @@ class Hypernetwork:
         for k, layers in self.layers.items():

             for layer in layers:

                 layer.train()

-                res += [layer.linear1.weight, layer.linear1.bias, layer.linear2.weight, layer.linear2.bias]

+                res += layer.trainables()

 

         return res

 

@@ -80,6 +128,9 @@ class Hypernetwork:
 

         state_dict['step'] = self.step

         state_dict['name'] = self.name

+        state_dict['layer_structure'] = self.layer_structure

+        state_dict['is_layer_norm'] = self.add_layer_norm

+        state_dict['activation_func'] = self.activation_func

         state_dict['sd_checkpoint'] = self.sd_checkpoint

         state_dict['sd_checkpoint_name'] = self.sd_checkpoint_name

 

@@ -92,9 +143,16 @@ class Hypernetwork:
 

         state_dict = torch.load(filename, map_location='cpu')

 

+        self.layer_structure = state_dict.get('layer_structure', [1, 2, 1])

+        self.add_layer_norm = state_dict.get('is_layer_norm', False)

+        self.activation_func = state_dict.get('activation_func', None)

+

         for size, sd in state_dict.items():

             if type(size) == int:

-                self.layers[size] = (HypernetworkModule(size, sd[0]), HypernetworkModule(size, sd[1]))

+                self.layers[size] = (

+                    HypernetworkModule(size, sd[0], self.layer_structure, self.add_layer_norm, self.activation_func),

+                    HypernetworkModule(size, sd[1], self.layer_structure, self.add_layer_norm, self.activation_func),

+                )

 

         self.name = state_dict.get('name', self.name)

         self.step = state_dict.get('step', 0)

@@ -196,7 +254,11 @@ def stack_conds(conds):
 

     return torch.stack(conds)

 

-def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, data_root, log_directory, steps, create_image_every, save_hypernetwork_every, template_file, preview_from_txt2img, preview_prompt, preview_negative_prompt, preview_steps, preview_sampler_index, preview_cfg_scale, preview_seed, preview_width, preview_height):

+

+def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, data_root, log_directory, training_width, training_height, steps, create_image_every, save_hypernetwork_every, template_file, preview_from_txt2img, preview_prompt, preview_negative_prompt, preview_steps, preview_sampler_index, preview_cfg_scale, preview_seed, preview_width, preview_height):

+    # images allows training previews to have infotext. Importing it at the top causes a circular import problem.

+    from modules import images

+

     assert hypernetwork_name, 'hypernetwork not selected'

 

     path = shared.hypernetworks.get(hypernetwork_name, None)

@@ -225,8 +287,7 @@ def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, data_root, log
 

     shared.state.textinfo = f"Preparing dataset from {html.escape(data_root)}..."

     with torch.autocast("cuda"):

-        ds = modules.textual_inversion.dataset.PersonalizedBase(data_root=data_root, width=512, height=512, repeats=shared.opts.training_image_repeats_per_epoch, placeholder_token=hypernetwork_name, model=shared.sd_model, device=devices.device, template_file=template_file, include_cond=True, batch_size=batch_size)

-

+        ds = modules.textual_inversion.dataset.PersonalizedBase(data_root=data_root, width=training_width, height=training_height, repeats=shared.opts.training_image_repeats_per_epoch, placeholder_token=hypernetwork_name, model=shared.sd_model, device=devices.device, template_file=template_file, include_cond=True, batch_size=batch_size)

     if unload:

         shared.sd_model.cond_stage_model.to(devices.cpu)

         shared.sd_model.first_stage_model.to(devices.cpu)

@@ -240,6 +301,7 @@ def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, data_root, log
 

     last_saved_file = "<none>"

     last_saved_image = "<none>"

+    forced_filename = "<none>"

 

     ititial_step = hypernetwork.step or 0

     if ititial_step > steps:

@@ -261,7 +323,7 @@ def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, data_root, log
 

         with torch.autocast("cuda"):

             c = stack_conds([entry.cond for entry in entries]).to(devices.device)

-#            c = torch.vstack([entry.cond for entry in entries]).to(devices.device)

+            # c = torch.vstack([entry.cond for entry in entries]).to(devices.device)

             x = torch.stack([entry.latent for entry in entries]).to(devices.device)

             loss = shared.sd_model(x, c)[0]

             del x

@@ -287,7 +349,8 @@ def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, data_root, log
         })

 

         if hypernetwork.step > 0 and images_dir is not None and hypernetwork.step % create_image_every == 0:

-            last_saved_image = os.path.join(images_dir, f'{hypernetwork_name}-{hypernetwork.step}.png')

+            forced_filename = f'{hypernetwork_name}-{hypernetwork.step}'

+            last_saved_image = os.path.join(images_dir, forced_filename)

 

             optimizer.zero_grad()

             shared.sd_model.cond_stage_model.to(devices.device)

@@ -323,7 +386,7 @@ def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, data_root, log
 

             if image is not None:

                 shared.state.current_image = image

-                image.save(last_saved_image)

+                last_saved_image, last_text_info = images.save_image(image, images_dir, "", p.seed, p.prompt, shared.opts.samples_format, processed.infotexts[0], p=p, forced_filename=forced_filename)

                 last_saved_image += f", prompt: {preview_text}"

 

         shared.state.job_no = hypernetwork.step

@@ -333,7 +396,7 @@ def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, data_root, log
 Loss: {mean_loss:.7f}<br/>

 Step: {hypernetwork.step}<br/>

 Last prompt: {html.escape(entries[0].cond_text)}<br/>

-Last saved embedding: {html.escape(last_saved_file)}<br/>

+Last saved hypernetwork: {html.escape(last_saved_file)}<br/>

 Last saved image: {html.escape(last_saved_image)}<br/>

 </p>

 """

diff --git a/modules/hypernetworks/ui.py b/modules/hypernetworks/ui.py
index dfa599af..266f04f6 100644
--- a/modules/hypernetworks/ui.py
+++ b/modules/hypernetworks/ui.py
@@ -1,5 +1,6 @@
 import html

 import os

+import re

 

 import gradio as gr

 

@@ -9,11 +10,21 @@ from modules import sd_hijack, shared, devices
 from modules.hypernetworks import hypernetwork

 

 

-def create_hypernetwork(name, enable_sizes):

+def create_hypernetwork(name, enable_sizes, overwrite_old, layer_structure=None, add_layer_norm=False, activation_func=None):

     fn = os.path.join(shared.cmd_opts.hypernetwork_dir, f"{name}.pt")

-    assert not os.path.exists(fn), f"file {fn} already exists"

-

-    hypernet = modules.hypernetworks.hypernetwork.Hypernetwork(name=name, enable_sizes=[int(x) for x in enable_sizes])

+    if not overwrite_old:

+        assert not os.path.exists(fn), f"file {fn} already exists"

+

+    if type(layer_structure) == str:

+        layer_structure = [float(x.strip()) for x in layer_structure.split(",")]

+

+    hypernet = modules.hypernetworks.hypernetwork.Hypernetwork(

+        name=name,

+        enable_sizes=[int(x) for x in enable_sizes],

+        layer_structure=layer_structure,

+        add_layer_norm=add_layer_norm,

+        activation_func=activation_func,

+    )

     hypernet.save(fn)

 

     shared.reload_hypernetworks()