Merge pull request #3842 from R-N/gradient-clipping

Gradient clipping in train tab

modules/hypernetworks/hypernetwork.py

@@ -402,10 +402,8 @@ def create_hypernetwork(name, enable_sizes, overwrite_old, layer_structure=None,
 
     shared.reload_hypernetworks()
 
-    return fn
 
-
-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):
+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):
     # images allows training previews to have infotext. Importing it at the top causes a circular import problem.
     from modules import images
 
@@ -448,6 +446,10 @@ def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, gradient_step,
         return hypernetwork, filename
 
     scheduler = LearnRateScheduler(learn_rate, steps, initial_step)
+
+    clip_grad = torch.nn.utils.clip_grad_value_ if clip_grad_mode == "value" else torch.nn.utils.clip_grad_norm_ if clip_grad_mode == "norm" else None
+    if clip_grad:
+        clip_grad_sched = LearnRateScheduler(clip_grad_value, steps, initial_step, verbose=False)
 
     # 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)}..."
@@ -466,7 +468,7 @@ def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, gradient_step,
         shared.parallel_processing_allowed = False
         shared.sd_model.cond_stage_model.to(devices.cpu)
         shared.sd_model.first_stage_model.to(devices.cpu)
-    
+
     weights = hypernetwork.weights()
     hypernetwork.train_mode()
 
@@ -525,6 +527,9 @@ def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, gradient_step,
                 if shared.state.interrupted:
                     break
 
+                if clip_grad:
+                    clip_grad_sched.step(hypernetwork.step)
+
                 with devices.autocast():
                     x = batch.latent_sample.to(devices.device, non_blocking=pin_memory)
                     if tag_drop_out != 0 or shuffle_tags:
@@ -539,14 +544,14 @@ def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, gradient_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}")
+
+                if clip_grad:
+                    clip_grad(weights, clip_grad_sched.learn_rate)
+
                 scaler.step(optimizer)
                 scaler.update()
                 hypernetwork.step += 1

modules/textual_inversion/learn_schedule.py

@@ -58,14 +58,19 @@ def __init__(self, learn_rate, max_steps, cur_step=0, verbose=True):
 
         self.finished = False
 
-    def apply(self, optimizer, step_number):
+    def step(self, step_number):
         if step_number < self.end_step:
-            return
+            return False
 
         try:
             (self.learn_rate, self.end_step) = next(self.schedules)
-        except Exception:
+        except StopIteration:
             self.finished = True
+            return False
+        return True
+
+    def apply(self, optimizer, step_number):
+        if not self.step(step_number):
             return
 
         if self.verbose:

modules/textual_inversion/textual_inversion.py

@@ -251,8 +251,7 @@ 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, 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, 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")
@@ -295,6 +294,11 @@ def train_embedding(embedding_name, learn_rate, batch_size, gradient_step, data_
         return embedding, filename
     scheduler = LearnRateScheduler(learn_rate, steps, initial_step)
 
+    clip_grad = torch.nn.utils.clip_grad_value_ if clip_grad_mode == "value" else \
+        torch.nn.utils.clip_grad_norm_ if clip_grad_mode == "norm" else \
+        None
+    if clip_grad:
+        clip_grad_sched = LearnRateScheduler(clip_grad_value, steps, ititial_step, verbose=False)
     # 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)}..."
     old_parallel_processing_allowed = shared.parallel_processing_allowed
@@ -361,6 +365,9 @@ def train_embedding(embedding_name, learn_rate, batch_size, gradient_step, data_
                 if shared.state.interrupted:
                     break
 
+                if clip_grad:
+                    clip_grad_sched.step(embedding.step)
+
                 with devices.autocast():
                     x = batch.latent_sample.to(devices.device, non_blocking=pin_memory)
                     c = shared.sd_model.cond_stage_model(batch.cond_text)
@@ -382,6 +389,10 @@ def train_embedding(embedding_name, learn_rate, batch_size, gradient_step, data_
                 # go back until we reach gradient accumulation steps
                 if (j + 1) % gradient_step != 0:
                     continue
+
+                if clip_grad:
+                    clip_grad(embedding.vec, clip_grad_sched.learn_rate)
+
                 scaler.step(optimizer)
                 scaler.update()
                 embedding.step += 1

modules/ui.py

@@ -1290,6 +1290,10 @@ def update_orig(image, state):
                     with gr.Row():
                         embedding_learn_rate = gr.Textbox(label='Embedding Learning rate', placeholder="Embedding Learning rate", value="0.005", elem_id="train_embedding_learn_rate")
                         hypernetwork_learn_rate = gr.Textbox(label='Hypernetwork Learning rate', placeholder="Hypernetwork Learning rate", value="0.00001", elem_id="train_hypernetwork_learn_rate")
+
+                    with gr.Row():
+                        clip_grad_mode = gr.Dropdown(value="disabled", label="Gradient Clipping", choices=["disabled", "value", "norm"])
+                        clip_grad_value = gr.Textbox(placeholder="Gradient clip value", value="0.1", show_label=False)
 
                     batch_size = gr.Number(label='Batch size', value=1, precision=0, elem_id="train_batch_size")
                     gradient_step = gr.Number(label='Gradient accumulation steps', value=1, precision=0, elem_id="train_gradient_step")
@@ -1402,6 +1406,8 @@ def update_orig(image, state):
                 training_width,
                 training_height,
                 steps,
+                clip_grad_mode,
+                clip_grad_value,
                 shuffle_tags,
                 tag_drop_out,
                 latent_sampling_method,
@@ -1431,6 +1437,8 @@ def update_orig(image, state):
                 training_width,
                 training_height,
                 steps,
+                clip_grad_mode,
+                clip_grad_value,
                 shuffle_tags,
                 tag_drop_out,
                 latent_sampling_method,