Browse Source
modify video shape
Signed-off-by: xujinling <jinling.xu@zilliz.com>
main
xujinling
3 years ago
1 changed files with
3 additions and
1 deletions
-
frozen_in_time.py
|
|
|
@ -70,7 +70,7 @@ class FrozenInTime(NNOperator): |
|
|
|
video_model_type='SpaceTimeTransformer', |
|
|
|
text_is_load_pretrained=False, |
|
|
|
device=self.device) |
|
|
|
|
|
|
|
self.model.to(self.device) |
|
|
|
self.tokenizer = AutoTokenizer.from_pretrained('distilbert-base-uncased', TOKENIZERS_PARALLELISM=False) |
|
|
|
self.transform_cfgs = get_configs( |
|
|
|
side_size=256, |
|
|
|
@ -101,6 +101,8 @@ class FrozenInTime(NNOperator): |
|
|
|
# Convert list of towhee.types.Image to numpy.ndarray in float32 |
|
|
|
video = numpy.stack([img.astype(numpy.float32) / 255. for img in data], axis=0) |
|
|
|
assert len(video.shape) == 4 |
|
|
|
if video.shape[0] != 4: |
|
|
|
self.transform_cfgs.update(num_frames=4) |
|
|
|
video = video.transpose(3, 0, 1, 2) # twhc -> ctwh |
|
|
|
video = transform_video( |
|
|
|
video=video, |
|
|
|
|
