Skip to content

Delta transfer learning between models (e.g. from Base to Turbo models) #1353

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Draft
dxqb wants to merge 1 commit into
base: master
Choose a base branch
from
+33 −1
Draft

Delta transfer learning between models (e.g. from Base to Turbo models) #1353

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
34 changes: 33 additions & 1 deletion modules/trainer/GenericTrainer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -607,6 +607,12 @@ def __before_eval(self):
self.model.optimizer.eval()

def train(self):


transfer_step1 = False
Copy link
Copy Markdown
Collaborator Author

@dxqb dxqb Mar 1, 2026 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

configuration is here
don't forget to restart OneTrainer when you have changed something

transfer_step2 = False
transfer_guidance = 3.0

train_device = torch.device(self.config.train_device)

train_progress = self.model.train_progress
Expand Down Expand Up @@ -739,7 +745,33 @@ def sample_commands_fun():
model_output_data['target'][prior_pred_indices] = prior_model_prediction[prior_pred_indices]
model_output_data['prior_target'] = prior_model_prediction
else:
model_output_data = self.model_setup.predict(self.model, batch, self.config, train_progress)
transfer_filename = f"transfer-{train_progress.global_step}.pt"
if transfer_step1:
with torch.no_grad():
with self.model_setup.prior_model(self.model, self.config):
model_output_data_without_lora = self.model_setup.predict(self.model, batch, self.config, train_progress)
model_output_data_with_lora = self.model_setup.predict(self.model, batch, self.config, train_progress)

torch.save((
model_output_data_without_lora['predicted'],
model_output_data_with_lora['predicted'],
batch['image_path'],
model_output_data_with_lora['timestep'],
), transfer_filename)

train_progress.next_step(self.config.batch_size)
self.callbacks.on_update_train_progress(train_progress, current_epoch_length, self.config.epochs)
continue
if transfer_step2:
base_without_lora, base_with_lora, image_path, timestep = torch.load(transfer_filename, map_location=train_device)
with self.model_setup.prior_model(self.model, self.config), torch.no_grad():
model_output_data_prior = self.model_setup.predict(self.model, batch, self.config, train_progress)

model_output_data = self.model_setup.predict(self.model, batch, self.config, train_progress)
assert image_path == batch['image_path'] and model_output_data['timestep'] == timestep, "transfer step 1 data does not match"
model_output_data['target'] = model_output_data_prior['predicted'] + transfer_guidance * (base_with_lora - base_without_lora)
else:
model_output_data = self.model_setup.predict(self.model, batch, self.config, train_progress)

loss = self.model_setup.calculate_loss(self.model, batch, model_output_data, self.config)

Expand Down