Merge pull request #6481 from guaneec/varsize

Allow mixed image sizes in TI/HN training

modules/hypernetworks/hypernetwork.py

@@ -403,7 +403,7 @@ def create_hypernetwork(name, enable_sizes, overwrite_old, layer_structure=None,
     shared.reload_hypernetworks()
 
 
-def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, gradient_step, data_root, log_directory, training_width, training_height, steps, clip_grad_mode, clip_grad_value, shuffle_tags, tag_drop_out, latent_sampling_method, 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, gradient_step, data_root, log_directory, training_width, training_height, varsize, steps, clip_grad_mode, clip_grad_value, shuffle_tags, tag_drop_out, latent_sampling_method, 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
 
@@ -456,7 +456,7 @@ def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, gradient_step,
 
     pin_memory = shared.opts.pin_memory
 
-    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, cond_model=shared.sd_model.cond_stage_model, device=devices.device, template_file=template_file, include_cond=True, batch_size=batch_size, gradient_step=gradient_step, shuffle_tags=shuffle_tags, tag_drop_out=tag_drop_out, latent_sampling_method=latent_sampling_method)
+    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, cond_model=shared.sd_model.cond_stage_model, device=devices.device, template_file=template_file, include_cond=True, batch_size=batch_size, gradient_step=gradient_step, shuffle_tags=shuffle_tags, tag_drop_out=tag_drop_out, latent_sampling_method=latent_sampling_method, varsize=varsize)
 
     if shared.opts.save_training_settings_to_txt:
         saved_params = dict(

modules/textual_inversion/dataset.py

@@ -17,7 +17,7 @@ re_numbers_at_start = re.compile(r"^[-\d]+\s*")
 
 
 class DatasetEntry:
-    def __init__(self, filename=None, filename_text=None, latent_dist=None, latent_sample=None, cond=None, cond_text=None, pixel_values=None):
+    def __init__(self, filename=None, filename_text=None, latent_dist=None, latent_sample=None, cond=None, cond_text=None, pixel_values=None, img_shape=None):
         self.filename = filename
         self.filename_text = filename_text
         self.latent_dist = latent_dist
@@ -25,16 +25,15 @@ class DatasetEntry:
         self.cond = cond
         self.cond_text = cond_text
         self.pixel_values = pixel_values
+        self.img_shape = img_shape
 
 
 class PersonalizedBase(Dataset):
-    def __init__(self, data_root, width, height, repeats, flip_p=0.5, placeholder_token="*", model=None, cond_model=None, device=None, template_file=None, include_cond=False, batch_size=1, gradient_step=1, shuffle_tags=False, tag_drop_out=0, latent_sampling_method='once'):
+    def __init__(self, data_root, width, height, repeats, flip_p=0.5, placeholder_token="*", model=None, cond_model=None, device=None, template_file=None, include_cond=False, batch_size=1, gradient_step=1, shuffle_tags=False, tag_drop_out=0, latent_sampling_method='once', varsize=False):
         re_word = re.compile(shared.opts.dataset_filename_word_regex) if len(shared.opts.dataset_filename_word_regex) > 0 else None
 
         self.placeholder_token = placeholder_token
 
-        self.width = width
-        self.height = height
         self.flip = transforms.RandomHorizontalFlip(p=flip_p)
 
         self.dataset = []
@@ -47,6 +46,8 @@ class PersonalizedBase(Dataset):
         assert data_root, 'dataset directory not specified'
         assert os.path.isdir(data_root), "Dataset directory doesn't exist"
         assert os.listdir(data_root), "Dataset directory is empty"
+        if varsize:
+            assert batch_size == 1, 'variable img size must have batch size 1'
 
         self.image_paths = [os.path.join(data_root, file_path) for file_path in os.listdir(data_root)]
 
@@ -59,7 +60,9 @@ class PersonalizedBase(Dataset):
             if shared.state.interrupted:
                 raise Exception("interrupted")
             try:
-                image = Image.open(path).convert('RGB').resize((self.width, self.height), PIL.Image.BICUBIC)
+                image = Image.open(path).convert('RGB')
+                if not varsize:
+                    image = image.resize((width, height), PIL.Image.BICUBIC)
             except Exception:
                 continue
 
@@ -88,14 +91,14 @@ class PersonalizedBase(Dataset):
             if latent_sampling_method == "once" or (latent_sampling_method == "deterministic" and not isinstance(latent_dist, DiagonalGaussianDistribution)):
                 latent_sample = model.get_first_stage_encoding(latent_dist).squeeze().to(devices.cpu)
                 latent_sampling_method = "once"
-                entry = DatasetEntry(filename=path, filename_text=filename_text, latent_sample=latent_sample)
+                entry = DatasetEntry(filename=path, filename_text=filename_text, latent_sample=latent_sample, img_shape=image.size)
             elif latent_sampling_method == "deterministic":
                 # Works only for DiagonalGaussianDistribution
                 latent_dist.std = 0
                 latent_sample = model.get_first_stage_encoding(latent_dist).squeeze().to(devices.cpu)
-                entry = DatasetEntry(filename=path, filename_text=filename_text, latent_sample=latent_sample)
+                entry = DatasetEntry(filename=path, filename_text=filename_text, latent_sample=latent_sample, img_shape=image.size)
             elif latent_sampling_method == "random":
-                entry = DatasetEntry(filename=path, filename_text=filename_text, latent_dist=latent_dist)
+                entry = DatasetEntry(filename=path, filename_text=filename_text, latent_dist=latent_dist, img_shape=image.size)
 
             if not (self.tag_drop_out != 0 or self.shuffle_tags):
                 entry.cond_text = self.create_text(filename_text)
@@ -151,6 +154,7 @@ class BatchLoader:
         self.cond_text = [entry.cond_text for entry in data]
         self.cond = [entry.cond for entry in data]
         self.latent_sample = torch.stack([entry.latent_sample for entry in data]).squeeze(1)
+        self.img_shape = [entry.img_shape for entry in data]
         #self.emb_index = [entry.emb_index for entry in data]
         #print(self.latent_sample.device)
 

modules/textual_inversion/textual_inversion.py

@@ -296,8 +296,7 @@ def validate_train_inputs(model_name, learn_rate, batch_size, gradient_step, dat
     if save_model_every or create_image_every:
         assert log_directory, "Log directory is empty"
 
-
-def train_embedding(embedding_name, learn_rate, batch_size, gradient_step, data_root, log_directory, training_width, training_height, steps, clip_grad_mode, clip_grad_value, shuffle_tags, tag_drop_out, latent_sampling_method, create_image_every, save_embedding_every, template_file, save_image_with_stored_embedding, preview_from_txt2img, preview_prompt, preview_negative_prompt, preview_steps, preview_sampler_index, preview_cfg_scale, preview_seed, preview_width, preview_height):
+def train_embedding(embedding_name, learn_rate, batch_size, gradient_step, data_root, log_directory, training_width, training_height, varsize, steps, clip_grad_mode, clip_grad_value, shuffle_tags, tag_drop_out, latent_sampling_method, create_image_every, save_embedding_every, template_file, save_image_with_stored_embedding, preview_from_txt2img, preview_prompt, preview_negative_prompt, preview_steps, preview_sampler_index, preview_cfg_scale, preview_seed, preview_width, preview_height):
     save_embedding_every = save_embedding_every or 0
     create_image_every = create_image_every or 0
     validate_train_inputs(embedding_name, learn_rate, batch_size, gradient_step, data_root, template_file, steps, save_embedding_every, create_image_every, log_directory, name="embedding")
@@ -351,7 +350,7 @@ def train_embedding(embedding_name, learn_rate, batch_size, gradient_step, data_
 
     pin_memory = shared.opts.pin_memory
 
-    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=embedding_name, model=shared.sd_model, cond_model=shared.sd_model.cond_stage_model, device=devices.device, template_file=template_file, batch_size=batch_size, gradient_step=gradient_step, shuffle_tags=shuffle_tags, tag_drop_out=tag_drop_out, latent_sampling_method=latent_sampling_method)
+    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=embedding_name, model=shared.sd_model, cond_model=shared.sd_model.cond_stage_model, device=devices.device, template_file=template_file, batch_size=batch_size, gradient_step=gradient_step, shuffle_tags=shuffle_tags, tag_drop_out=tag_drop_out, latent_sampling_method=latent_sampling_method, varsize=varsize)
 
     if shared.opts.save_training_settings_to_txt:
         save_settings_to_file(log_directory, {**dict(model_name=checkpoint.model_name, model_hash=checkpoint.hash, num_of_dataset_images=len(ds), num_vectors_per_token=len(embedding.vec)), **locals()})
@@ -493,8 +492,8 @@ def train_embedding(embedding_name, learn_rate, batch_size, gradient_step, data_
                     else:
                         p.prompt = batch.cond_text[0]
                         p.steps = 20
-                        p.width = training_width
-                        p.height = training_height
+                        p.width = batch.img_shape[0][0]
+                        p.height = batch.img_shape[0][1]
 
                     preview_text = p.prompt
 

modules/ui.py

@@ -1348,6 +1348,7 @@ def create_ui():
                     template_file = gr.Textbox(label='Prompt template file', value=os.path.join(script_path, "textual_inversion_templates", "style_filewords.txt"), elem_id="train_template_file")
                     training_width = gr.Slider(minimum=64, maximum=2048, step=8, label="Width", value=512, elem_id="train_training_width")
                     training_height = gr.Slider(minimum=64, maximum=2048, step=8, label="Height", value=512, elem_id="train_training_height")
+                    varsize = gr.Checkbox(label="Ignore dimension settings and do not resize images", value=False, elem_id="train_varsize")
                     steps = gr.Number(label='Max steps', value=100000, precision=0, elem_id="train_steps")
 
                     with FormRow():
@@ -1454,6 +1455,7 @@ def create_ui():
                 log_directory,
                 training_width,
                 training_height,
+                varsize,
                 steps,
                 clip_grad_mode,
                 clip_grad_value,
@@ -1485,6 +1487,7 @@ def create_ui():
                 log_directory,
                 training_width,
                 training_height,
+                varsize,
                 steps,
                 clip_grad_mode,
                 clip_grad_value,