diff options
Diffstat (limited to 'modules')
| -rw-r--r-- | modules/hypernetworks/hypernetwork.py | 16 | ||||
| -rw-r--r-- | modules/textual_inversion/learn_schedule.py | 11 | ||||
| -rw-r--r-- | modules/textual_inversion/textual_inversion.py | 14 | ||||
| -rw-r--r-- | modules/ui.py | 8 |
4 files changed, 42 insertions, 7 deletions
diff --git a/modules/hypernetworks/hypernetwork.py b/modules/hypernetworks/hypernetwork.py index 7f182712..3371b18e 100644 --- a/modules/hypernetworks/hypernetwork.py +++ b/modules/hypernetworks/hypernetwork.py @@ -367,7 +367,7 @@ 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, data_root, log_directory, training_width, training_height, steps, clip_grad_mode, clip_grad_value, 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
@@ -410,6 +410,12 @@ def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, data_root, log scheduler = LearnRateScheduler(learn_rate, steps, ititial_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)}..."
with torch.autocast("cuda"):
@@ -425,7 +431,7 @@ def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, data_root, log 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
@@ -466,6 +472,9 @@ def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, data_root, log if shared.state.interrupted:
break
+ if clip_grad:
+ clip_grad_sched.step(hypernetwork.step)
+
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)
@@ -488,6 +497,9 @@ def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, data_root, log 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'
+ if clip_grad:
+ clip_grad(weights, clip_grad_sched.learn_rate)
+
optimizer.step()
steps_done = hypernetwork.step + 1
diff --git a/modules/textual_inversion/learn_schedule.py b/modules/textual_inversion/learn_schedule.py index dd0c0ad1..f63fc72f 100644 --- a/modules/textual_inversion/learn_schedule.py +++ b/modules/textual_inversion/learn_schedule.py @@ -58,14 +58,19 @@ class LearnRateScheduler: 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:
diff --git a/modules/textual_inversion/textual_inversion.py b/modules/textual_inversion/textual_inversion.py index 0aeb0459..687d97bb 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -224,7 +224,7 @@ 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, data_root, log_directory, training_width, training_height, steps, clip_grad_mode, clip_grad_value, 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")
@@ -269,6 +269,11 @@ def train_embedding(embedding_name, learn_rate, batch_size, data_root, log_direc scheduler = LearnRateScheduler(learn_rate, steps, ititial_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)}..."
with torch.autocast("cuda"):
@@ -297,6 +302,9 @@ def train_embedding(embedding_name, learn_rate, batch_size, data_root, log_direc if shared.state.interrupted:
break
+ if clip_grad:
+ clip_grad_sched.step(embedding.step)
+
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)
@@ -307,6 +315,10 @@ def train_embedding(embedding_name, learn_rate, batch_size, data_root, log_direc optimizer.zero_grad()
loss.backward()
+
+ if clip_grad:
+ clip_grad(embedding.vec, clip_grad_sched.learn_rate)
+
optimizer.step()
steps_done = embedding.step + 1
diff --git a/modules/ui.py b/modules/ui.py index 7ea1177f..67d787a7 100644 --- a/modules/ui.py +++ b/modules/ui.py @@ -1252,7 +1252,9 @@ def create_ui(wrap_gradio_gpu_call): with gr.Row():
embedding_learn_rate = gr.Textbox(label='Embedding Learning rate', placeholder="Embedding Learning rate", value="0.005")
hypernetwork_learn_rate = gr.Textbox(label='Hypernetwork Learning rate', placeholder="Hypernetwork Learning rate", value="0.00001")
-
+ 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)
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")
@@ -1357,6 +1359,8 @@ def create_ui(wrap_gradio_gpu_call): training_width,
training_height,
steps,
+ clip_grad_mode,
+ clip_grad_value,
create_image_every,
save_embedding_every,
template_file,
@@ -1382,6 +1386,8 @@ def create_ui(wrap_gradio_gpu_call): training_width,
training_height,
steps,
+ clip_grad_mode,
+ clip_grad_value,
create_image_every,
save_embedding_every,
template_file,
|