diff options
Diffstat (limited to 'modules')
| -rw-r--r-- | modules/hypernetworks/hypernetwork.py | 17 | ||||
| -rw-r--r-- | modules/textual_inversion/learn_schedule.py | 11 | ||||
| -rw-r--r-- | modules/textual_inversion/textual_inversion.py | 17 | ||||
| -rw-r--r-- | modules/ui.py | 8 |
4 files changed, 47 insertions, 6 deletions
diff --git a/modules/hypernetworks/hypernetwork.py b/modules/hypernetworks/hypernetwork.py index 2e84583b..f45ce199 100644 --- a/modules/hypernetworks/hypernetwork.py +++ b/modules/hypernetworks/hypernetwork.py @@ -328,7 +328,7 @@ 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
@@ -384,8 +384,15 @@ def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, data_root, log ititial_step = hypernetwork.step or 0
if ititial_step > steps:
return hypernetwork, filename
+
+ clip_grad_mode_value = clip_grad_mode == "value"
+ clip_grad_mode_norm = clip_grad_mode == "norm"
+ clip_grad_enabled = clip_grad_mode_value or clip_grad_mode_norm
+ if clip_grad_enabled:
+ clip_grad_sched = LearnRateScheduler(clip_grad_value, steps, ititial_step, verbose=False)
scheduler = LearnRateScheduler(learn_rate, steps, ititial_step)
+
# if optimizer == "AdamW": or else Adam / AdamW / SGD, etc...
optimizer = torch.optim.AdamW(weights, lr=scheduler.learn_rate)
@@ -405,6 +412,9 @@ def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, data_root, log if shared.state.interrupted:
break
+ if clip_grad_enabled:
+ 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)
@@ -427,6 +437,11 @@ 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_mode_value:
+ torch.nn.utils.clip_grad_value_(weights, clip_value=clip_grad_sched.learn_rate)
+ elif clip_grad_mode_norm:
+ torch.nn.utils.clip_grad_norm_(weights, max_norm=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 3a736065..2627d585 100644 --- a/modules/textual_inversion/learn_schedule.py +++ b/modules/textual_inversion/learn_schedule.py @@ -51,14 +51,19 @@ 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 17dfb223..f272e536 100644 --- a/modules/textual_inversion/textual_inversion.py +++ b/modules/textual_inversion/textual_inversion.py @@ -205,7 +205,7 @@ def write_loss(log_directory, filename, step, epoch_len, values): })
-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):
assert embedding_name, 'embedding not selected'
shared.state.textinfo = "Initializing textual inversion training..."
@@ -254,6 +254,12 @@ def train_embedding(embedding_name, learn_rate, batch_size, data_root, log_direc ititial_step = embedding.step or 0
if ititial_step > steps:
return embedding, filename
+
+ clip_grad_mode_value = clip_grad_mode == "value"
+ clip_grad_mode_norm = clip_grad_mode == "norm"
+ clip_grad_enabled = clip_grad_mode_value or clip_grad_mode_norm
+ if clip_grad_enabled:
+ clip_grad_sched = LearnRateScheduler(clip_grad_value, steps, ititial_step, verbose=False)
scheduler = LearnRateScheduler(learn_rate, steps, ititial_step)
optimizer = torch.optim.AdamW([embedding.vec], lr=scheduler.learn_rate)
@@ -269,6 +275,9 @@ def train_embedding(embedding_name, learn_rate, batch_size, data_root, log_direc if shared.state.interrupted:
break
+ if clip_grad_enabled:
+ 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)
@@ -279,6 +288,12 @@ def train_embedding(embedding_name, learn_rate, batch_size, data_root, log_direc optimizer.zero_grad()
loss.backward()
+
+ if clip_grad_mode_value:
+ torch.nn.utils.clip_grad_value_(embedding.vec, clip_value=clip_grad_sched.learn_rate)
+ elif clip_grad_mode_norm:
+ torch.nn.utils.clip_grad_norm_(embedding.vec, max_norm=clip_grad_sched.learn_rate)
+
optimizer.step()
steps_done = embedding.step + 1
diff --git a/modules/ui.py b/modules/ui.py index 5055ca64..98f9565f 100644 --- a/modules/ui.py +++ b/modules/ui.py @@ -1254,7 +1254,9 @@ 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="1.0", 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")
@@ -1355,6 +1357,8 @@ 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,
@@ -1380,6 +1384,8 @@ 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,
|