storytold/FineTrainers-Conditioning
2
0
Fork 0
mirror of https://github.com/storytold/FineTrainers-Conditioning.git synced 2026-10-28 08:57:10 +00:00
Code Issues Packages Projects Releases Wiki Activity

Ok adapter trains need to get the pass for validation using a skeleton.

Browse Source
This commit is contained in:
CrossProduct
2025-01-22 22:07:52 +00:00
parent d631737329
commit d6b3615421
1 changed files with 11 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files
finetrainers/trainer.py
+11 -9
View File
@@ -825,15 +825,8 @@ class Trainer:
width=latent_conditions["width"], width=latent_conditions["width"],
patch_size = 1, patch_size = 1,
patch_size_t = 1) patch_size_t = 1)
# target # target latent patchified
latent_conditions["latents"] = post_conditioned_latent_patchify(
latents=latent_conditions["latents"],
num_frames=latent_conditions["num_frames"],
height=latent_conditions["height"],
width=latent_conditions["width"],
patch_size = 1,
patch_size_t = 1
)
# pose information latent # pose information latent
latent_conditions.update({"noisy_latents": latent_final["latents"]}) latent_conditions.update({"noisy_latents": latent_final["latents"]})
latent_conditions.update({"noisy_latents_residual":noisy_latents_residual["latents"]}) latent_conditions.update({"noisy_latents_residual":noisy_latents_residual["latents"]})
@@ -875,6 +868,15 @@ class Trainer:
scheduler=self.scheduler, noise=noise, latents=latent_conditions["latents"] scheduler=self.scheduler, noise=noise, latents=latent_conditions["latents"]
) )
if self.pose_condition:
# patchify target
target = post_conditioned_latent_patchify(latents=target,
num_frames=latent_conditions["num_frames"],
height=latent_conditions["height"],
width=latent_conditions["width"],
patch_size = 1,
patch_size_t = 1)["latents"]
loss = weights.float() * (pred["latents"].float() - target.float()).pow(2) loss = weights.float() * (pred["latents"].float() - target.float()).pow(2)
# Average loss across all but batch dimension # Average loss across all but batch dimension
loss = loss.mean(list(range(1, loss.ndim))) loss = loss.mean(list(range(1, loss.ndim)))