add option to use batch size for training

modules/hypernetworks/hypernetwork.py

@@ -182,7 +182,21 @@ def attention_CrossAttention_forward(self, x, context=None, mask=None):
     return self.to_out(out)
 
 
-def train_hypernetwork(hypernetwork_name, learn_rate, 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 stack_conds(conds):
+    if len(conds) == 1:
+        return torch.stack(conds)
+
+    # same as in reconstruct_multicond_batch
+    token_count = max([x.shape[0] for x in conds])
+    for i in range(len(conds)):
+        if conds[i].shape[0] != token_count:
+            last_vector = conds[i][-1:]
+            last_vector_repeated = last_vector.repeat([token_count - conds[i].shape[0], 1])
+            conds[i] = torch.vstack([conds[i], last_vector_repeated])
+
+    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):
     assert hypernetwork_name, 'hypernetwork not selected'
 
     path = shared.hypernetworks.get(hypernetwork_name, None)
@@ -211,7 +225,7 @@ def train_hypernetwork(hypernetwork_name, learn_rate, data_root, log_directory,
 
     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)
+        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)
 
     if unload:
         shared.sd_model.cond_stage_model.to(devices.cpu)
@@ -235,7 +249,7 @@ def train_hypernetwork(hypernetwork_name, learn_rate, data_root, log_directory,
     optimizer = torch.optim.AdamW(weights, lr=scheduler.learn_rate)
 
     pbar = tqdm.tqdm(enumerate(ds), total=steps - ititial_step)
-    for i, entry in pbar:
+    for i, entries in pbar:
         hypernetwork.step = i + ititial_step
 
         scheduler.apply(optimizer, hypernetwork.step)
@@ -246,11 +260,12 @@ def train_hypernetwork(hypernetwork_name, learn_rate, data_root, log_directory,
             break
 
         with torch.autocast("cuda"):
-            cond = entry.cond.to(devices.device)
-            x = entry.latent.to(devices.device)
-            loss = shared.sd_model(x.unsqueeze(0), cond)[0]
+            c = stack_conds([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
-            del cond
+            del c
 
             losses[hypernetwork.step % losses.shape[0]] = loss.item()
 
@@ -292,7 +307,7 @@ def train_hypernetwork(hypernetwork_name, learn_rate, data_root, log_directory,
                 p.width = preview_width
                 p.height = preview_height
             else:
-                p.prompt = entry.cond_text
+                p.prompt = entries[0].cond_text
                 p.steps = 20
 
             preview_text = p.prompt
@@ -315,7 +330,7 @@ def train_hypernetwork(hypernetwork_name, learn_rate, data_root, log_directory,
 <p>
 Loss: {losses.mean():.7f}<br/>
 Step: {hypernetwork.step}<br/>
-Last prompt: {html.escape(entry.cond_text)}<br/>
+Last prompt: {html.escape(entries[0].cond_text)}<br/>
 Last saved embedding: {html.escape(last_saved_file)}<br/>
 Last saved image: {html.escape(last_saved_image)}<br/>
 </p>

modules/textual_inversion/dataset.py

@@ -24,11 +24,12 @@ class DatasetEntry:
 
 
 class PersonalizedBase(Dataset):
-    def __init__(self, data_root, width, height, repeats, flip_p=0.5, placeholder_token="*", model=None, device=None, template_file=None, include_cond=False):
-        re_word = re.compile(shared.opts.dataset_filename_word_regex) if len(shared.opts.dataset_filename_word_regex)>0 else None
+    def __init__(self, data_root, width, height, repeats, flip_p=0.5, placeholder_token="*", model=None, device=None, template_file=None, include_cond=False, batch_size=1):
+        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.batch_size = batch_size
         self.width = width
         self.height = height
         self.flip = transforms.RandomHorizontalFlip(p=flip_p)
@@ -78,13 +79,13 @@ class PersonalizedBase(Dataset):
 
             if include_cond:
                 entry.cond_text = self.create_text(filename_text)
-                entry.cond = cond_model([entry.cond_text]).to(devices.cpu)
+                entry.cond = cond_model([entry.cond_text]).to(devices.cpu).squeeze(0)
 
             self.dataset.append(entry)
 
-        self.length = len(self.dataset) * repeats
+        self.length = len(self.dataset) * repeats // batch_size
 
-        self.initial_indexes = np.arange(self.length) % len(self.dataset)
+        self.initial_indexes = np.arange(len(self.dataset))
         self.indexes = None
         self.shuffle()
 
@@ -101,13 +102,19 @@ class PersonalizedBase(Dataset):
         return self.length
 
     def __getitem__(self, i):
-        if i % len(self.dataset) == 0:
-            self.shuffle()
+        res = []
 
-        index = self.indexes[i % len(self.indexes)]
-        entry = self.dataset[index]
+        for j in range(self.batch_size):
+            position = i * self.batch_size + j
+            if position % len(self.indexes) == 0:
+                self.shuffle()
 
-        if entry.cond is None:
-            entry.cond_text = self.create_text(entry.filename_text)
+            index = self.indexes[position % len(self.indexes)]
+            entry = self.dataset[index]
 
-        return entry
+            if entry.cond is None:
+                entry.cond_text = self.create_text(entry.filename_text)
+
+            res.append(entry)
+
+        return res

modules/textual_inversion/textual_inversion.py

@@ -199,7 +199,7 @@ def write_loss(log_directory, filename, step, epoch_len, values):
         })
 
 
-def train_embedding(embedding_name, learn_rate, data_root, log_directory, training_width, training_height, steps, 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, data_root, log_directory, training_width, training_height, steps, 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):
     assert embedding_name, 'embedding not selected'
 
     shared.state.textinfo = "Initializing textual inversion training..."
@@ -231,7 +231,7 @@ def train_embedding(embedding_name, learn_rate, data_root, log_directory, traini
 
     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=training_width, height=training_height, repeats=shared.opts.training_image_repeats_per_epoch, placeholder_token=embedding_name, model=shared.sd_model, device=devices.device, template_file=template_file)
+        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, device=devices.device, template_file=template_file, batch_size=batch_size)
 
     hijack = sd_hijack.model_hijack
 
@@ -251,7 +251,7 @@ def train_embedding(embedding_name, learn_rate, data_root, log_directory, traini
     optimizer = torch.optim.AdamW([embedding.vec], lr=scheduler.learn_rate)
 
     pbar = tqdm.tqdm(enumerate(ds), total=steps-ititial_step)
-    for i, entry in pbar:
+    for i, entries in pbar:
         embedding.step = i + ititial_step
 
         scheduler.apply(optimizer, embedding.step)
@@ -262,10 +262,9 @@ def train_embedding(embedding_name, learn_rate, data_root, log_directory, traini
             break
 
         with torch.autocast("cuda"):
-            c = cond_model([entry.cond_text])
-
-            x = entry.latent.to(devices.device)
-            loss = shared.sd_model(x.unsqueeze(0), c)[0]
+            c = cond_model([entry.cond_text for entry in entries])
+            x = torch.stack([entry.latent for entry in entries]).to(devices.device)
+            loss = shared.sd_model(x, c)[0]
             del x
 
             losses[embedding.step % losses.shape[0]] = loss.item()
@@ -307,7 +306,7 @@ def train_embedding(embedding_name, learn_rate, data_root, log_directory, traini
                 p.width = preview_width
                 p.height = preview_height
             else:
-                p.prompt = entry.cond_text
+                p.prompt = entries[0].cond_text
                 p.steps = 20
                 p.width = training_width
                 p.height = training_height
@@ -348,7 +347,7 @@ def train_embedding(embedding_name, learn_rate, data_root, log_directory, traini
 <p>
 Loss: {losses.mean():.7f}<br/>
 Step: {embedding.step}<br/>
-Last prompt: {html.escape(entry.cond_text)}<br/>
+Last prompt: {html.escape(entries[0].cond_text)}<br/>
 Last saved embedding: {html.escape(last_saved_file)}<br/>
 Last saved image: {html.escape(last_saved_image)}<br/>
 </p>

modules/ui.py

@@ -1166,6 +1166,7 @@ def create_ui(wrap_gradio_gpu_call):
                     train_embedding_name = gr.Dropdown(label='Embedding', choices=sorted(sd_hijack.model_hijack.embedding_db.word_embeddings.keys()))
                     train_hypernetwork_name = gr.Dropdown(label='Hypernetwork', choices=[x for x in shared.hypernetworks.keys()])
                     learn_rate = gr.Textbox(label='Learning rate', placeholder="Learning rate", value="0.005")
+                    batch_size = gr.Number(label='Batch size', value=1, precision=0)
                     dataset_directory = gr.Textbox(label='Dataset directory', placeholder="Path to directory with input images")
                     log_directory = gr.Textbox(label='Log directory', placeholder="Path to directory where to write outputs", value="textual_inversion")
                     template_file = gr.Textbox(label='Prompt template file', value=os.path.join(script_path, "textual_inversion_templates", "style_filewords.txt"))
@@ -1244,6 +1245,7 @@ def create_ui(wrap_gradio_gpu_call):
             inputs=[
                 train_embedding_name,
                 learn_rate,
+                batch_size,
                 dataset_directory,
                 log_directory,
                 training_width,
@@ -1268,6 +1270,7 @@ def create_ui(wrap_gradio_gpu_call):
             inputs=[
                 train_hypernetwork_name,
                 learn_rate,
+                batch_size,
                 dataset_directory,
                 log_directory,
                 steps,