storytold/FineTrainers-Conditioning
2
0
Fork 0
mirror of https://github.com/storytold/FineTrainers-Conditioning.git synced 2026-10-25 08:29:35 +00:00
Code Issues Packages Projects Releases Wiki Activity

Adapter weights can be loaded.

Browse Source
This commit is contained in:
CrossProduct
2025-02-04 16:46:45 +00:00
parent ea3cd5c753
commit 28165e67d0
5 changed files with 41 additions and 8 deletions
Download Patch File Download Diff File Expand all files Collapse all files
finetrainers/trainer.py
+9 -3
View File
@@ -330,7 +330,7 @@ class Trainer:
def prepare_trainable_parameters(self) -> None:
logger.info("Initializing trainable parameters")
diffusion_components = self.model_config["load_diffusion_models"](**self._get_load_components_kwargs())
diffusion_components = self.model_config["load_diffusion_models"](output_dir=self.args.output_dir,**self._get_load_components_kwargs())
self._set_components(diffusion_components)
components = [self.text_encoder, self.text_encoder_2, self.text_encoder_3, self.vae]
@@ -1039,9 +1039,9 @@ class Trainer:
# loading videos for inference ..
if self.pose_condition:
if pose_video is not None:
pose_video = self.preprocess_condition_video(pose_video,max_num_frames=self.dataset.max_num_frames)
pose_video = self.dataset._preprocess_video(Path(pose_video)).unsqueeze(0) #self.preprocess_condition_video(pose_video,max_num_frames=self.dataset.max_num_frames)
if img_ref_video is not None:
img_ref_video = self.preprocess_condition_video_image_reference_video(path=img_ref_video,resolution_buckets=self.dataset.resolution_buckets,max_num_frames=self.dataset.max_num_frames)
img_ref_video = self.dataset._preprocess_video_image_reference_video(Path(img_ref_video)).unsqueeze(0) #self.preprocess_condition_video_image_reference_video(path=img_ref_video,resolution_buckets=self.dataset.resolution_buckets,max_num_frames=self.dataset.max_num_frames)
logger.debug(
f"Validating sample {i + 1}/{num_validation_samples} on process {accelerator.process_index}. Prompt: {prompt}",
@@ -1393,3 +1393,9 @@ class Trainer:
# adding singleton dimension for downstream inference.
frames = frames.unsqueeze(0)
return frames
@staticmethod
def _find_nearest_resolution(resolutions, height, width):
nearest_res = min(resolutions, key=lambda x: abs(x[1] - height) + abs(x[2] - width))
return nearest_res[1], nearest_res[2]