Qwen-Image-Layered-MRP-MLX/src/mflux/models/flux/variants/dreambooth/dreambooth.py

77 lines
3.1 KiB
Python

from mlx import nn
from tqdm import tqdm
from mflux.models.common.config.config import Config
from mflux.models.common.lora.layer.linear_lora_layer import LoRALinear
from mflux.models.flux.variants.dreambooth.optimization.dreambooth_loss import DreamBoothLoss
from mflux.models.flux.variants.dreambooth.state.training_spec import TrainingSpec
from mflux.models.flux.variants.dreambooth.state.training_state import TrainingState
from mflux.models.flux.variants.dreambooth.statistics.plotter import Plotter
from mflux.models.flux.variants.txt2img.flux import Flux1
class DreamBooth:
@staticmethod
def train(
flux: Flux1,
config: Config,
training_spec: TrainingSpec,
training_state: TrainingState,
):
# Freeze the model (LoRA layers are already applied to transformer in from_spec)
flux.freeze()
# Unfreeze LoRA layers so they can be trained
DreamBooth._unfreeze_lora_layers(flux.transformer)
# Define loss computation as a function of a batch 'b'
train_step_function = nn.value_and_grad(
model=flux,
fn=lambda b: DreamBoothLoss.compute_loss(flux, config, b),
)
# Setup progress bar
batches = tqdm(
training_state.iterator,
total=training_state.iterator.total_number_of_steps(),
initial=training_state.iterator.num_iterations,
)
# Training loop
for batch in batches:
# Perform one gradient update on the LoRA the weights
loss, grads = train_step_function(batch)
training_state.optimizer.optimizer.update(model=flux, gradients=grads)
del loss, grads
# Plot loss progress periodically
if training_state.should_plot_loss(training_spec):
validation_batch = training_state.iterator.get_validation_batch()
validation_loss = DreamBoothLoss.compute_loss(flux, config, validation_batch)
training_state.statistics.append_values(step=training_state.iterator.num_iterations, loss=validation_loss) # fmt: off
Plotter.update_loss_plot(training_spec=training_spec, training_state=training_state)
del validation_loss
# Generate a test image from the model periodically
if training_state.should_generate_image(training_spec):
image = flux.generate_image(
seed=training_spec.seed,
prompt=training_spec.instrumentation.validation_prompt,
)
image.save(path=training_state.get_current_validation_image_path(training_spec))
del image
flux.prompt_cache = {}
# Save checkpoint periodically
if training_state.should_save(training_spec):
training_state.save(flux, training_spec)
# Save the final state
training_state.save(flux, training_spec)
@staticmethod
def _unfreeze_lora_layers(module: nn.Module) -> None:
for name, child in module.named_modules():
if isinstance(child, LoRALinear):
child.unfreeze(keys=["lora_A", "lora_B"], strict=False)