Gradient accumulation, autocast fix, new latent sampling method, etc

modules/hypernetworks/hypernetwork.py

@@ -367,13 +367,13 @@ def report_statistics(loss_info:dict):
 
 
 
-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):
+def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, gradient_step, data_root, log_directory, training_width, training_height, steps, 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
 
     save_hypernetwork_every = save_hypernetwork_every or 0
     create_image_every = create_image_every or 0
-    textual_inversion.validate_train_inputs(hypernetwork_name, learn_rate, batch_size, data_root, template_file, steps, save_hypernetwork_every, create_image_every, log_directory, name="hypernetwork")
+    textual_inversion.validate_train_inputs(hypernetwork_name, learn_rate, batch_size, gradient_step, data_root, template_file, steps, save_hypernetwork_every, create_image_every, log_directory, name="hypernetwork")
 
     path = shared.hypernetworks.get(hypernetwork_name, None)
     shared.loaded_hypernetwork = Hypernetwork()
@@ -403,29 +403,25 @@ def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, data_root, log
     hypernetwork = shared.loaded_hypernetwork
     checkpoint = sd_models.select_checkpoint()
 
-    ititial_step = hypernetwork.step or 0
-    if ititial_step >= steps:
+    initial_step = hypernetwork.step or 0
+    if initial_step >= steps:
         shared.state.textinfo = f"Model has already been trained beyond specified max steps"
         return hypernetwork, filename
 
-    scheduler = LearnRateScheduler(learn_rate, steps, ititial_step)
+    scheduler = LearnRateScheduler(learn_rate, steps, initial_step)
 
     # 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)}..."
-    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=hypernetwork_name, model=shared.sd_model, device=devices.device, template_file=template_file, include_cond=True, batch_size=batch_size)
+
+    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)
+    dl = modules.textual_inversion.dataset.PersonalizedDataLoader(ds, batch_size=ds.batch_size, pin_memory=pin_memory)
 
     if unload:
         shared.sd_model.cond_stage_model.to(devices.cpu)
         shared.sd_model.first_stage_model.to(devices.cpu)
     
-    size = len(ds.indexes)
-    loss_dict = defaultdict(lambda : deque(maxlen = 1024))
-    losses = torch.zeros((size,))
-    previous_mean_losses = [0]
-    previous_mean_loss = 0
-    print("Mean loss of {} elements".format(size))
-    
     weights = hypernetwork.weights()
     for weight in weights:
         weight.requires_grad = True
@@ -446,62 +442,81 @@ def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, data_root, log
             print("Cannot resume from saved optimizer!")
             print(e)
 
+    scaler = torch.cuda.amp.GradScaler()
+    
+    batch_size = ds.batch_size
+    gradient_step = ds.gradient_step
+    # n steps = batch_size * gradient_step * n image processed
+    steps_per_epoch = len(ds) // batch_size // gradient_step
+    max_steps_per_epoch = len(ds) // batch_size - (len(ds) // batch_size) % gradient_step
+    loss_step = 0
+    _loss_step = 0 #internal
+    # size = len(ds.indexes)
+    # loss_dict = defaultdict(lambda : deque(maxlen = 1024))
+    # losses = torch.zeros((size,))
+    # previous_mean_losses = [0]
+    # previous_mean_loss = 0
+    # print("Mean loss of {} elements".format(size))
+
     steps_without_grad = 0
 
     last_saved_file = "<none>"
     last_saved_image = "<none>"
     forced_filename = "<none>"
 
-    pbar = tqdm.tqdm(enumerate(ds), total=steps - ititial_step)
-    for i, entries in pbar:
-        hypernetwork.step = i + ititial_step
-        if len(loss_dict) > 0:
-            previous_mean_losses = [i[-1] for i in loss_dict.values()]
-            previous_mean_loss = mean(previous_mean_losses)
-            
+    pbar = tqdm.tqdm(total=steps - initial_step)
+    try:
+        for i in range((steps-initial_step) * gradient_step):
+            if scheduler.finished:
+                break
+            if shared.state.interrupted:
+                break
+            for j, batch in enumerate(dl):
+                # works as a drop_last=True for gradient accumulation
+                if j == max_steps_per_epoch:
+                    break
                 scheduler.apply(optimizer, hypernetwork.step)
                 if scheduler.finished:
                     break
-
                 if shared.state.interrupted:
                     break
 
                 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)
-            x = torch.stack([entry.latent for entry in entries]).to(devices.device)
-            loss = shared.sd_model(x, c)[0]
+                    x = batch.latent_sample.to(devices.device, non_blocking=pin_memory)
+                    if tag_drop_out != 0 or shuffle_tags:
+                        shared.sd_model.cond_stage_model.to(devices.device)
+                        c = shared.sd_model.cond_stage_model(batch.cond_text).to(devices.device, non_blocking=pin_memory)
+                        shared.sd_model.cond_stage_model.to(devices.cpu)
+                    else:
+                        c = stack_conds(batch.cond).to(devices.device, non_blocking=pin_memory)
+                    loss = shared.sd_model(x, c)[0] / gradient_step
                     del x
                     del c
 
-            losses[hypernetwork.step % losses.shape[0]] = loss.item()
-            for entry in entries:
-                loss_dict[entry.filename].append(loss.item())
-                
-            optimizer.zero_grad()
-            weights[0].grad = None
-            loss.backward()
-
-            if weights[0].grad is None:
-                steps_without_grad += 1
-            else:
-                steps_without_grad = 0
-            assert steps_without_grad < 10, 'no gradient found for the trained weight after backward() for 10 steps in a row; this is a bug; training cannot continue'
-
-            optimizer.step()
+                    _loss_step += loss.item()
+                scaler.scale(loss).backward()
+                # go back until we reach gradient accumulation steps
+                if (j + 1) % gradient_step != 0:
+                    continue
+                # print(f"grad:{weights[0].grad.detach().cpu().abs().mean().item():.7f}")
+                # scaler.unscale_(optimizer)
+                # print(f"grad:{weights[0].grad.detach().cpu().abs().mean().item():.15f}")
+                # torch.nn.utils.clip_grad_norm_(weights, max_norm=1.0)
+                # print(f"grad:{weights[0].grad.detach().cpu().abs().mean().item():.15f}")
+                scaler.step(optimizer)
+                scaler.update()
+                hypernetwork.step += 1
+                pbar.update()
+                optimizer.zero_grad(set_to_none=True)
+                loss_step = _loss_step
+                _loss_step = 0
 
                 steps_done = hypernetwork.step + 1
                 
-        if torch.isnan(losses[hypernetwork.step % losses.shape[0]]): 
-            raise RuntimeError("Loss diverged.")
-        
-        if len(previous_mean_losses) > 1:
-            std = stdev(previous_mean_losses)
-        else:
-            std = 0
-        dataset_loss_info = f"dataset loss:{mean(previous_mean_losses):.3f}" + u"\u00B1" + f"({std / (len(previous_mean_losses) ** 0.5):.3f})"
-        pbar.set_description(dataset_loss_info)
+                epoch_num = hypernetwork.step // steps_per_epoch
+                epoch_step = hypernetwork.step % steps_per_epoch
 
+                pbar.set_description(f"[Epoch {epoch_num}: {epoch_step+1}/{steps_per_epoch}]loss: {loss_step:.7f}")
                 if hypernetwork_dir is not None and steps_done % save_hypernetwork_every == 0:
                     # Before saving, change name to match current checkpoint.
                     hypernetwork_name_every = f'{hypernetwork_name}-{steps_done}'
@@ -512,8 +527,8 @@ def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, data_root, log
                     save_hypernetwork(hypernetwork, checkpoint, hypernetwork_name, last_saved_file)
                     hypernetwork.optimizer_state_dict = None  # dereference it after saving, to save memory.
 
-        textual_inversion.write_loss(log_directory, "hypernetwork_loss.csv", hypernetwork.step, len(ds), {
-            "loss": f"{previous_mean_loss:.7f}",
+                textual_inversion.write_loss(log_directory, "hypernetwork_loss.csv", hypernetwork.step, steps_per_epoch, {
+                    "loss": f"{loss_step:.7f}",
                     "learn_rate": scheduler.learn_rate
                 })
 
@@ -521,7 +536,6 @@ def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, data_root, log
                     forced_filename = f'{hypernetwork_name}-{steps_done}'
                     last_saved_image = os.path.join(images_dir, forced_filename)
 
-            optimizer.zero_grad()
                     shared.sd_model.cond_stage_model.to(devices.device)
                     shared.sd_model.first_stage_model.to(devices.device)
 
@@ -541,13 +555,15 @@ def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, data_root, log
                         p.width = preview_width
                         p.height = preview_height
                     else:
-                p.prompt = entries[0].cond_text
+                        p.prompt = batch.cond_text[0]
                         p.steps = 20
+                        p.width = training_width
+                        p.height = training_height
 
                     preview_text = p.prompt
 
                     processed = processing.process_images(p)
-            image = processed.images[0] if len(processed.images)>0 else None
+                    image = processed.images[0] if len(processed.images) > 0 else None
 
                     if unload:
                         shared.sd_model.cond_stage_model.to(devices.cpu)
@@ -562,15 +578,19 @@ def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, data_root, log
 
                 shared.state.textinfo = f"""
 <p>
-Loss: {previous_mean_loss:.7f}<br/>
+Loss: {loss_step:.7f}<br/>
 Step: {hypernetwork.step}<br/>
-Last prompt: {html.escape(entries[0].cond_text)}<br/>
+Last prompt: {html.escape(batch.cond_text[0])}<br/>
 Last saved hypernetwork: {html.escape(last_saved_file)}<br/>
 Last saved image: {html.escape(last_saved_image)}<br/>
 </p>
 """
-        
-    report_statistics(loss_dict)
+    except Exception:
+        print(traceback.format_exc(), file=sys.stderr)
+    finally:
+        pbar.leave = False
+        pbar.close()
+        #report_statistics(loss_dict)
 
     filename = os.path.join(shared.cmd_opts.hypernetwork_dir, f'{hypernetwork_name}.pt')
     hypernetwork.optimizer_name = optimizer_name
@@ -579,6 +599,9 @@ Last saved image: {html.escape(last_saved_image)}<br/>
     save_hypernetwork(hypernetwork, checkpoint, hypernetwork_name, filename)
     del optimizer
     hypernetwork.optimizer_state_dict = None  # dereference it after saving, to save memory.
+    shared.sd_model.cond_stage_model.to(devices.device)
+    shared.sd_model.first_stage_model.to(devices.device)
+
     return hypernetwork, filename
 
 def save_hypernetwork(hypernetwork, checkpoint, hypernetwork_name, filename):

modules/sd_hijack.py

@@ -8,7 +8,7 @@ from torch import einsum
 from torch.nn.functional import silu
 
 import modules.textual_inversion.textual_inversion
-from modules import prompt_parser, devices, sd_hijack_optimizations, shared
+from modules import prompt_parser, devices, sd_hijack_optimizations, shared, sd_hijack_checkpoint
 from modules.shared import opts, device, cmd_opts
 from modules.sd_hijack_optimizations import invokeAI_mps_available
 
@@ -59,6 +59,10 @@ def undo_optimizations():
 def get_target_prompt_token_count(token_count):
     return math.ceil(max(token_count, 1) / 75) * 75
 
+def fix_checkpoint():
+    ldm.modules.attention.BasicTransformerBlock.forward = sd_hijack_checkpoint.BasicTransformerBlock_forward
+    ldm.modules.diffusionmodules.openaimodel.ResBlock.forward = sd_hijack_checkpoint.ResBlock_forward
+    ldm.modules.diffusionmodules.openaimodel.AttentionBlock.forward = sd_hijack_checkpoint.AttentionBlock_forward
 
 class StableDiffusionModelHijack:
     fixes = None
@@ -78,6 +82,7 @@ class StableDiffusionModelHijack:
         self.clip = m.cond_stage_model
 
         apply_optimizations()
+        fix_checkpoint()
 
         def flatten(el):
             flattened = [flatten(children) for children in el.children()]

modules/sd_hijack_checkpoint.py

@@ -0,0 +1,10 @@
+from torch.utils.checkpoint import checkpoint
+
+def BasicTransformerBlock_forward(self, x, context=None):
+    return checkpoint(self._forward, x, context)
+
+def AttentionBlock_forward(self, x):
+    return checkpoint(self._forward, x)
+
+def ResBlock_forward(self, x, emb):
+    return checkpoint(self._forward, x, emb)

modules/shared.py

@@ -322,8 +322,7 @@ options_templates.update(options_section(('system', "System"), {
 
 options_templates.update(options_section(('training', "Training"), {
     "unload_models_when_training": OptionInfo(False, "Move VAE and CLIP to RAM when training if possible. Saves VRAM."),
-    "shuffle_tags": OptionInfo(False, "Shuffleing tags by ',' when create texts."),
-    "tag_drop_out": OptionInfo(0, "Dropout tags when create texts", gr.Slider, {"minimum": 0, "maximum": 1, "step": 0.1}),
+    "pin_memory": OptionInfo(False, "Turn on pin_memory for DataLoader. Makes training slightly faster but can increase memory usage."),
     "save_optimizer_state": OptionInfo(False, "Saves Optimizer state as separate *.optim file. Training can be resumed with HN itself and matching optim file."),
     "dataset_filename_word_regex": OptionInfo("", "Filename word regex"),
     "dataset_filename_join_string": OptionInfo(" ", "Filename join string"),

modules/textual_inversion/dataset.py

@@ -3,7 +3,7 @@ import numpy as np
 import PIL
 import torch
 from PIL import Image
-from torch.utils.data import Dataset
+from torch.utils.data import Dataset, DataLoader
 from torchvision import transforms
 
 import random
@@ -11,25 +11,28 @@ import tqdm
 from modules import devices, shared
 import re
 
+from ldm.modules.distributions.distributions import DiagonalGaussianDistribution
+
 re_numbers_at_start = re.compile(r"^[-\d]+\s*")
 
 
 class DatasetEntry:
-    def __init__(self, filename=None, latent=None, filename_text=None):
+    def __init__(self, filename=None, filename_text=None, latent_dist=None, latent_sample=None, cond=None, cond_text=None, pixel_values=None):
         self.filename = filename
-        self.latent = latent
         self.filename_text = filename_text
-        self.cond = None
-        self.cond_text = None
+        self.latent_dist = latent_dist
+        self.latent_sample = latent_sample
+        self.cond = cond
+        self.cond_text = cond_text
+        self.pixel_values = pixel_values
 
 
 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, batch_size=1):
+    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'):        
         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)
@@ -45,11 +48,16 @@ class PersonalizedBase(Dataset):
         assert os.path.isdir(data_root), "Dataset directory doesn't exist"
         assert os.listdir(data_root), "Dataset directory is empty"
 
-        cond_model = shared.sd_model.cond_stage_model
-
         self.image_paths = [os.path.join(data_root, file_path) for file_path in os.listdir(data_root)]
+
+        
+        self.shuffle_tags = shuffle_tags
+        self.tag_drop_out = tag_drop_out
+
         print("Preparing dataset...")
         for path in tqdm.tqdm(self.image_paths):
+            if shared.state.interrupted:
+                raise Exception("inturrupted")
             try:
                 image = Image.open(path).convert('RGB').resize((self.width, self.height), PIL.Image.BICUBIC)
             except Exception:
@@ -71,37 +79,58 @@ class PersonalizedBase(Dataset):
             npimage = np.array(image).astype(np.uint8)
             npimage = (npimage / 127.5 - 1.0).astype(np.float32)
 
-            torchdata = torch.from_numpy(npimage).to(device=device, dtype=torch.float32)
-            torchdata = torch.moveaxis(torchdata, 2, 0)
+            torchdata = torch.from_numpy(npimage).permute(2, 0, 1).to(device=device, dtype=torch.float32)
+            latent_sample = None
 
-            init_latent = model.get_first_stage_encoding(model.encode_first_stage(torchdata.unsqueeze(dim=0))).squeeze()
-            init_latent = init_latent.to(devices.cpu)
+            with torch.autocast("cuda"):
+                latent_dist = model.encode_first_stage(torchdata.unsqueeze(dim=0))
 
-            entry = DatasetEntry(filename=path, filename_text=filename_text, latent=init_latent)
+            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)
+            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)
+            elif latent_sampling_method == "random":
+                entry = DatasetEntry(filename=path, filename_text=filename_text, latent_dist=latent_dist)
 
-            if include_cond:
+            if not (self.tag_drop_out != 0 or self.shuffle_tags):
                 entry.cond_text = self.create_text(filename_text)
+
+            if include_cond and not (self.tag_drop_out != 0 or self.shuffle_tags):
+                with torch.autocast("cuda"):
                     entry.cond = cond_model([entry.cond_text]).to(devices.cpu).squeeze(0)
+            # elif not include_cond:
+            #     _, _, _, _, hijack_fixes, token_count = cond_model.process_text([entry.cond_text])
+            #     max_n = token_count // 75
+            #     index_list = [ [] for _ in range(max_n + 1) ]
+            #     for n, (z, _) in hijack_fixes[0]:
+            #         index_list[n].append(z)
+            #     with torch.autocast("cuda"):
+            #         entry.cond = cond_model([entry.cond_text]).to(devices.cpu).squeeze(0)
+            #     entry.emb_index = index_list
 
             self.dataset.append(entry)
+            del torchdata
+            del latent_dist
+            del latent_sample
 
-        assert len(self.dataset) > 0, "No images have been found in the dataset."
-        self.length = len(self.dataset) * repeats // batch_size
-
-        self.dataset_length = len(self.dataset)
-        self.indexes = None
-        self.shuffle()
-
-    def shuffle(self):
-        self.indexes = np.random.permutation(self.dataset_length)
+        self.length = len(self.dataset)
+        assert self.length > 0, "No images have been found in the dataset."
+        self.batch_size = min(batch_size, self.length)
+        self.gradient_step = min(gradient_step, self.length // self.batch_size)
+        self.latent_sampling_method = latent_sampling_method
 
     def create_text(self, filename_text):
         text = random.choice(self.lines)
         text = text.replace("[name]", self.placeholder_token)
         tags = filename_text.split(',')
-        if shared.opts.tag_drop_out != 0:
-            tags = [t for t in tags if random.random() > shared.opts.tag_drop_out]
-        if shared.opts.shuffle_tags:
+        if self.tag_drop_out != 0:
+            tags = [t for t in tags if random.random() > self.tag_drop_out]
+        if self.shuffle_tags:
             random.shuffle(tags)
         text = text.replace("[filewords]", ','.join(tags))
         return text
@@ -110,19 +139,28 @@ class PersonalizedBase(Dataset):
         return self.length
 
     def __getitem__(self, i):
-        res = []
-
-        for j in range(self.batch_size):
-            position = i * self.batch_size + j
-            if position % len(self.indexes) == 0:
-                self.shuffle()
-
-            index = self.indexes[position % len(self.indexes)]
-            entry = self.dataset[index]
-
-            if entry.cond is None:
+        entry = self.dataset[i]
+        if self.tag_drop_out != 0 or self.shuffle_tags:
             entry.cond_text = self.create_text(entry.filename_text)
+        if self.latent_sampling_method == "random":
+            entry.latent_sample = shared.sd_model.get_first_stage_encoding(entry.latent_dist)
+        return entry
 
-            res.append(entry)
+class PersonalizedDataLoader(DataLoader):
+    def __init__(self, *args, **kwargs):
+        super(PersonalizedDataLoader, self).__init__(shuffle=True, drop_last=True, *args, **kwargs)
+        self.collate_fn = collate_wrapper
         
-        return res
+
+class BatchLoader:
+    def __init__(self, data):
+        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)
+
+    def pin_memory(self):
+        self.latent_sample = self.latent_sample.pin_memory()
+        return self
+
+def collate_wrapper(batch):
+    return BatchLoader(batch)

modules/textual_inversion/textual_inversion.py

@@ -184,7 +184,7 @@ def write_loss(log_directory, filename, step, epoch_len, values):
     if shared.opts.training_write_csv_every == 0:
         return
 
-    if (step + 1) % shared.opts.training_write_csv_every != 0:
+    if step % shared.opts.training_write_csv_every != 0:
         return
     write_csv_header = False if os.path.exists(os.path.join(log_directory, filename)) else True
 
@@ -194,21 +194,23 @@ def write_loss(log_directory, filename, step, epoch_len, values):
         if write_csv_header:
             csv_writer.writeheader()
 
-        epoch = step // epoch_len
-        epoch_step = step % epoch_len 
+        epoch = (step - 1) // epoch_len
+        epoch_step = (step - 1) % epoch_len 
 
         csv_writer.writerow({
-            "step": step + 1,
+            "step": step,
             "epoch": epoch,
-            "epoch_step": epoch_step + 1,
+            "epoch_step": epoch_step,
             **values,
         })
 
-def validate_train_inputs(model_name, learn_rate, batch_size, data_root, template_file, steps, save_model_every, create_image_every, log_directory, name="embedding"):
+def validate_train_inputs(model_name, learn_rate, batch_size, gradient_step, data_root, template_file, steps, save_model_every, create_image_every, log_directory, name="embedding"):
     assert model_name, f"{name} not selected"
     assert learn_rate, "Learning rate is empty or 0"
     assert isinstance(batch_size, int), "Batch size must be integer"
     assert batch_size > 0, "Batch size must be positive"
+    assert isinstance(gradient_step, int), "Gradient accumulation step must be integer"
+    assert gradient_step > 0, "Gradient accumulation step must be positive"
     assert data_root, "Dataset directory is empty"
     assert os.path.isdir(data_root), "Dataset directory doesn't exist"
     assert os.listdir(data_root), "Dataset directory is empty"
@@ -224,10 +226,10 @@ def validate_train_inputs(model_name, learn_rate, batch_size, data_root, templat
     if save_model_every or create_image_every:
         assert log_directory, "Log directory is empty"
 
-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):
+def train_embedding(embedding_name, learn_rate, batch_size, gradient_step, data_root, log_directory, training_width, training_height, steps, 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, data_root, template_file, steps, save_embedding_every, create_image_every, log_directory, name="embedding")
+    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")
 
     shared.state.textinfo = "Initializing textual inversion training..."
     shared.state.job_count = steps
@@ -255,76 +257,112 @@ def train_embedding(embedding_name, learn_rate, batch_size, data_root, log_direc
     else:
         images_embeds_dir = None
 
-    cond_model = shared.sd_model.cond_stage_model
-
     hijack = sd_hijack.model_hijack
 
     embedding = hijack.embedding_db.word_embeddings[embedding_name]
     checkpoint = sd_models.select_checkpoint()
 
-    ititial_step = embedding.step or 0
-    if ititial_step >= steps:
+    initial_step = embedding.step or 0
+    if initial_step >= steps:
         shared.state.textinfo = f"Model has already been trained beyond specified max steps"
         return embedding, filename
-
-    scheduler = LearnRateScheduler(learn_rate, steps, ititial_step)
+    scheduler = LearnRateScheduler(learn_rate, steps, initial_step)
 
    # 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)}..."
-    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, batch_size=batch_size)
+    
+    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)
+
+    latent_sampling_method = ds.latent_sampling_method
+
+    dl = modules.textual_inversion.dataset.PersonalizedDataLoader(ds, batch_size=ds.batch_size, pin_memory=False)
+
     if unload:
         shared.sd_model.first_stage_model.to(devices.cpu)
 
     embedding.vec.requires_grad = True
     optimizer = torch.optim.AdamW([embedding.vec], lr=scheduler.learn_rate)
+    scaler = torch.cuda.amp.GradScaler()
+
+    batch_size = ds.batch_size
+    gradient_step = ds.gradient_step
+    # n steps = batch_size * gradient_step * n image processed
+    steps_per_epoch = len(ds) // batch_size // gradient_step
+    max_steps_per_epoch = len(ds) // batch_size - (len(ds) // batch_size) % gradient_step
+    loss_step = 0
+    _loss_step = 0 #internal
 
-    losses = torch.zeros((32,))
     
     last_saved_file = "<none>"
     last_saved_image = "<none>"
     forced_filename = "<none>"
     embedding_yet_to_be_embedded = False
     
-    pbar = tqdm.tqdm(enumerate(ds), total=steps-ititial_step)
-    for i, entries in pbar:
-        embedding.step = i + ititial_step
-
+    pbar = tqdm.tqdm(total=steps - initial_step)
+    try:
+        for i in range((steps-initial_step) * gradient_step):
+            if scheduler.finished:
+                break
+            if shared.state.interrupted:
+                break
+            for j, batch in enumerate(dl):
+                # works as a drop_last=True for gradient accumulation
+                if j == max_steps_per_epoch:
+                    break
                 scheduler.apply(optimizer, embedding.step)
                 if scheduler.finished:
                     break
-
                 if shared.state.interrupted:
                     break
 
                 with torch.autocast("cuda"):
-            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]
+                    # c = stack_conds(batch.cond).to(devices.device)
+                    # mask = torch.tensor(batch.emb_index).to(devices.device, non_blocking=pin_memory)
+                    # print(mask)
+                    # c[:, 1:1+embedding.vec.shape[0]] = embedding.vec.to(devices.device, non_blocking=pin_memory)
+                    x = batch.latent_sample.to(devices.device, non_blocking=pin_memory)
+                    c = shared.sd_model.cond_stage_model(batch.cond_text)
+                    loss = shared.sd_model(x, c)[0] / gradient_step
                     del x
                     
-            losses[embedding.step % losses.shape[0]] = loss.item()
+                    _loss_step += loss.item()
+                scaler.scale(loss).backward()
                 
-            optimizer.zero_grad()
-            loss.backward()
-            optimizer.step()
+                # go back until we reach gradient accumulation steps
+                if (j + 1) % gradient_step != 0:
+                    continue
+                #print(f"grad:{embedding.vec.grad.detach().cpu().abs().mean().item():.7f}")
+                #scaler.unscale_(optimizer)
+                #print(f"grad:{embedding.vec.grad.detach().cpu().abs().mean().item():.7f}")
+                #torch.nn.utils.clip_grad_norm_(embedding.vec, max_norm=1.0)
+                #print(f"grad:{embedding.vec.grad.detach().cpu().abs().mean().item():.7f}")
+                scaler.step(optimizer)
+                scaler.update()
+                embedding.step += 1
+                pbar.update()
+                optimizer.zero_grad(set_to_none=True)
+                loss_step = _loss_step
+                _loss_step = 0
 
                 steps_done = embedding.step + 1
 
-        epoch_num = embedding.step // len(ds)
-        epoch_step = embedding.step % len(ds)
-
-        pbar.set_description(f"[Epoch {epoch_num}: {epoch_step+1}/{len(ds)}]loss: {losses.mean():.7f}")
+                epoch_num = embedding.step // steps_per_epoch
+                epoch_step = embedding.step % steps_per_epoch
 
+                pbar.set_description(f"[Epoch {epoch_num}: {epoch_step+1}/{steps_per_epoch}]loss: {loss_step:.7f}")
                 if embedding_dir is not None and steps_done % save_embedding_every == 0:
                     # Before saving, change name to match current checkpoint.
                     embedding_name_every = f'{embedding_name}-{steps_done}'
                     last_saved_file = os.path.join(embedding_dir, f'{embedding_name_every}.pt')
+                    #if shared.opts.save_optimizer_state:
+                        #embedding.optimizer_state_dict = optimizer.state_dict()
                     save_embedding(embedding, checkpoint, embedding_name_every, last_saved_file, remove_cached_checksum=True)
                     embedding_yet_to_be_embedded = True
 
-        write_loss(log_directory, "textual_inversion_loss.csv", embedding.step, len(ds), {
-            "loss": f"{losses.mean():.7f}",
+                write_loss(log_directory, "textual_inversion_loss.csv", embedding.step, steps_per_epoch, {
+                    "loss": f"{loss_step:.7f}",
                     "learn_rate": scheduler.learn_rate
                 })
 
@@ -351,7 +389,7 @@ def train_embedding(embedding_name, learn_rate, batch_size, data_root, log_direc
                         p.width = preview_width
                         p.height = preview_height
                     else:
-                p.prompt = entries[0].cond_text
+                        p.prompt = batch.cond_text[0]
                         p.steps = 20
                         p.width = training_width
                         p.height = training_height
@@ -359,12 +397,15 @@ def train_embedding(embedding_name, learn_rate, batch_size, data_root, log_direc
                     preview_text = p.prompt
 
                     processed = processing.process_images(p)
-            image = processed.images[0]
+                    image = processed.images[0] if len(processed.images) > 0 else None
 
                     if unload:
                         shared.sd_model.first_stage_model.to(devices.cpu)
 
+                    if image is not None:
                         shared.state.current_image = 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, save_to_dirs=False)
+                        last_saved_image += f", prompt: {preview_text}"
 
                     if save_image_with_stored_embedding and os.path.exists(last_saved_file) and embedding_yet_to_be_embedded:
 
@@ -399,16 +440,21 @@ def train_embedding(embedding_name, learn_rate, batch_size, data_root, log_direc
 
                 shared.state.textinfo = f"""
 <p>
-Loss: {losses.mean():.7f}<br/>
+Loss: {loss_step:.7f}<br/>
 Step: {embedding.step}<br/>
-Last prompt: {html.escape(entries[0].cond_text)}<br/>
+Last prompt: {html.escape(batch.cond_text[0])}<br/>
 Last saved embedding: {html.escape(last_saved_file)}<br/>
 Last saved image: {html.escape(last_saved_image)}<br/>
 </p>
 """
-
         filename = os.path.join(shared.cmd_opts.embeddings_dir, f'{embedding_name}.pt')
         save_embedding(embedding, checkpoint, embedding_name, filename, remove_cached_checksum=True)
+    except Exception:
+        print(traceback.format_exc(), file=sys.stderr)
+        pass
+    finally:
+        pbar.leave = False
+        pbar.close()
         shared.sd_model.first_stage_model.to(devices.device)
 
     return embedding, filename

modules/ui.py

@@ -1289,6 +1289,7 @@ def create_ui(wrap_gradio_gpu_call):
                         hypernetwork_learn_rate = gr.Textbox(label='Hypernetwork Learning rate', placeholder="Hypernetwork Learning rate", value="0.00001")
 
                     batch_size = gr.Number(label='Batch size', value=1, precision=0)
+                    gradient_step = gr.Number(label='Gradient accumulation steps', 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"))
@@ -1299,6 +1300,11 @@ def create_ui(wrap_gradio_gpu_call):
                     save_embedding_every = gr.Number(label='Save a copy of embedding to log directory every N steps, 0 to disable', value=500, precision=0)
                     save_image_with_stored_embedding = gr.Checkbox(label='Save images with embedding in PNG chunks', value=True)
                     preview_from_txt2img = gr.Checkbox(label='Read parameters (prompt, etc...) from txt2img tab when making previews', value=False)
+                    with gr.Row():
+                        shuffle_tags = gr.Checkbox(label="Shuffle tags by ',' when creating prompts.", value=False)
+                        tag_drop_out = gr.Slider(minimum=0, maximum=1, step=0.1, label="Drop out tags when creating prompts.", value=0)
+                    with gr.Row():
+                        latent_sampling_method = gr.Radio(label='Choose latent sampling method', value="once", choices=['once', 'deterministic', 'random'])
 
                     with gr.Row():
                         interrupt_training = gr.Button(value="Interrupt")
@@ -1387,11 +1393,15 @@ def create_ui(wrap_gradio_gpu_call):
                 train_embedding_name,
                 embedding_learn_rate,
                 batch_size,
+                gradient_step,
                 dataset_directory,
                 log_directory,
                 training_width,
                 training_height,
                 steps,
+                shuffle_tags,
+                tag_drop_out,
+                latent_sampling_method,
                 create_image_every,
                 save_embedding_every,
                 template_file,
@@ -1412,11 +1422,15 @@ def create_ui(wrap_gradio_gpu_call):
                 train_hypernetwork_name,
                 hypernetwork_learn_rate,
                 batch_size,
+                gradient_step,
                 dataset_directory,
                 log_directory,
                 training_width,
                 training_height,
                 steps,
+                shuffle_tags,
+                tag_drop_out,
+                latent_sampling_method,
                 create_image_every,
                 save_embedding_every,
                 template_file,