riaz.somc/transformers
1
0
Fork 0
mirror of https://github.com/huggingface/transformers.git synced 2025-08-03 03:31:05 +06:00
Code Issues Actions 2 Packages Projects Releases Wiki Activity

Fixed num_channels!=3 normalization training (#20630)

Browse Source 
* Fixed num_channels!=3 normalization training

* empty commit to trigger CI

* Empty-Commit for CircleCI

* Empty-Commit

* Empty Commit try-3: https://discuss.circleci.com/t/github-code-checkout-suddenly-failing/31558

* Empty commit to trigger CI

Co-authored-by: Lay Jain <layjain@basil.csail.mit.edu>
Co-authored-by: ydshieh <ydshieh@users.noreply.github.com>
This commit is contained in:
layjain 2023-01-17 23:36:20 +05:30 committed by GitHub
parent 865da84abb
commit 44caf4f6f4
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
1 changed files with 13 additions and 5 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

18
src/transformers/models/videomae/modeling_videomae.py
View File

@ -819,11 +819,15 @@ class VideoMAEForPreTraining(VideoMAEPreTrainedModel):
loss = None
with torch.no_grad():
# calculate the labels to be predicted
# first, unnormalize the frames
device = pixel_values.device
mean = torch.as_tensor(IMAGENET_DEFAULT_MEAN).to(device)[None, None, :, None, None]
std = torch.as_tensor(IMAGENET_DEFAULT_STD).to(device)[None, None, :, None, None]
frames = pixel_values * std + mean # in [0, 1]
if self.config.num_channels != 3:
# Can't unnormalize with default means/stds
frames = pixel_values
else:
# first, unnormalize the frames
device = pixel_values.device
mean = torch.as_tensor(IMAGENET_DEFAULT_MEAN).to(device)[None, None, :, None, None]
std = torch.as_tensor(IMAGENET_DEFAULT_STD).to(device)[None, None, :, None, None]
frames = pixel_values * std + mean # in [0, 1]
batch_size, time, num_channels, height, width = frames.shape
tubelet_size, patch_size = self.config.tubelet_size, self.config.patch_size
@ -859,6 +863,10 @@ class VideoMAEForPreTraining(VideoMAEPreTrainedModel):
tubelet_size * patch_size * patch_size * num_channels,
)
else:
if self.config.num_channels != 3:
raise ValueError(
"Can't unnormalize non-RGB images. Consider setting config.norm_pix_loss to False."
)
# step 1: split up dimensions (time by tubelet_size, height by patch_size, width by patch_size)
frames = frames.view(
batch_size,