From 59afd55775db0d51ade8b4cb07b2434511e7aaf4 Mon Sep 17 00:00:00 2001 From: CrossProduct Date: Fri, 17 Jan 2025 22:43:20 +0000 Subject: [PATCH] Retested changes without conditioning it works. --- finetrainers/trainer.py | 12 ++++++------ 1 file changed, 6 insertions(+), 6 deletions(-) diff --git a/finetrainers/trainer.py b/finetrainers/trainer.py index cb2dc5e..a838821 100644 --- a/finetrainers/trainer.py +++ b/finetrainers/trainer.py @@ -100,10 +100,10 @@ class Trainer: self.state.model_name = self.args.model_name self.model_config = get_config_from_model_name(self.args.model_name, self.args.training_type) + self.pose_condition = False if self.args.pose_column != None: self.pose_condition = True - else: - self.pose_condition = False + def prepare_dataset(self) -> None: # TODO(aryan): Make a background process for fetching @@ -723,7 +723,7 @@ class Trainer: latent_conditions = make_contiguous(latent_conditions) - if self.pose_conditioning: + if self.pose_condition: # VAE output not patchified yet. pose_video_latents = condition_latents_prepare.prepare_latents_for_conditioning( vae=self.vae, @@ -739,7 +739,7 @@ class Trainer: img_refs_latents = condition_latents_prepare.prepare_latents_for_conditioning( vae=self.vae, - image_or_video=img_refs_latents, + image_or_video=img_refs, patch_size=self.transformer_config.patch_size, patch_size_t=self.transformer_config.patch_size_t, device=accelerator.device, @@ -773,7 +773,7 @@ class Trainer: generator=self.state.generator, ) timesteps = (sigmas * 1000.0).long() - if self.pose_conditioning: + if self.pose_condition: noise = torch.randn( latent_conditions["latents"].shape, generator=self.state.generator, @@ -826,7 +826,7 @@ class Trainer: ) weights = expand_tensor_dims(weights, noise.ndim) - if self.pose_conditioning: + if self.pose_condition: pred = self.model_config["forward_pass"]( transformer=self.transformer, scheduler=self.scheduler,