Added experimental wavelet loss
This commit is contained in:
1 parent
c101f07834
commit
ce4c5291a0
4 files changed
+45
-2
No files matched your search
@@ -34,6 +34,7 @@ from diffusers import EMAModel
|
||||
import math
|
||||
from toolkit.train_tools import precondition_model_outputs_flow_match
|
||||
from toolkit.models.diffusion_feature_extraction import DiffusionFeatureExtractor, load_dfe
|
||||
from toolkit.util.wavelet_loss import wavelet_loss
|
||||
|
||||
|
||||
def flush():
|
||||
@@ -481,6 +482,8 @@ class SDTrainer(BaseSDTrainProcess):
|
||||
|
||||
if self.train_config.loss_type == "mae":
|
||||
loss = torch.nn.functional.l1_loss(pred.float(), target.float(), reduction="none")
|
||||
elif self.train_config.loss_type == "wavelet":
|
||||
loss = wavelet_loss(pred, batch.latents, noise)
|
||||
else:
|
||||
loss = torch.nn.functional.mse_loss(pred.float(), target.float(), reduction="none")
|
||||
|
||||
|
||||
Reference in new issue
Block a user