def test_video_classifier_finetune_fiftyone(tmpdir):
with mock_encoded_video_dataset_folder(tmpdir) as (
dir_name,
total_duration,
):
half_duration = total_duration / 2 - 1e-9
train_dataset = fo.Dataset.from_dir(
dir_name,
dataset_type=fo.types.VideoClassificationDirectoryTree,
)
datamodule = VideoClassificationData.from_fiftyone(
train_dataset=train_dataset,
clip_sampler="uniform",
clip_duration=half_duration,
video_sampler=SequentialSampler,
decode_audio=False,
batch_size=1,
)
for sample in datamodule.train_dataset.data:
expected_t_shape = 5
assert sample["video"].shape[1] == expected_t_shape
model = VideoClassifier(num_classes=datamodule.num_classes,
pretrained=False,
backbone="slow_r50")
trainer = flash.Trainer(fast_dev_run=True,
gpus=torch.cuda.device_count())
trainer.finetune(model, datamodule=datamodule)
def test_video_classifier_finetune_fiftyone(tmpdir):
with mock_encoded_video_dataset_folder(tmpdir) as (
dir_name,
total_duration,
):
half_duration = total_duration / 2 - 1e-9
train_dataset = fo.Dataset.from_dir(
dir_name,
dataset_type=fo.types.VideoClassificationDirectoryTree,
)
datamodule = VideoClassificationData.from_fiftyone(
train_dataset=train_dataset,
clip_sampler="uniform",
clip_duration=half_duration,
video_sampler=SequentialSampler,
decode_audio=False,
)
for sample in datamodule.train_dataset.data:
expected_t_shape = 5
assert sample["video"].shape[1] == expected_t_shape
assert len(VideoClassifier.available_backbones()) > 5
train_transform = {
"post_tensor_transform": Compose([
ApplyTransformToKey(
key="video",
transform=Compose([
UniformTemporalSubsample(8),
RandomShortSideScale(min_size=256, max_size=320),
RandomCrop(244),
RandomHorizontalFlip(p=0.5),
]),
),
]),
"per_batch_transform_on_device": Compose([
ApplyTransformToKey(
key="video",
transform=K.VideoSequential(
K.Normalize(torch.tensor([0.45, 0.45, 0.45]), torch.tensor([0.225, 0.225, 0.225])),
K.augmentation.ColorJitter(0.1, 0.1, 0.1, 0.1, p=1.0),
data_format="BCTHW",
same_on_frame=False
)
),
]),
}
datamodule = VideoClassificationData.from_fiftyone(
train_dataset=train_dataset,
clip_sampler="uniform",
clip_duration=half_duration,
video_sampler=SequentialSampler,
decode_audio=False,
train_transform=train_transform
)
model = VideoClassifier(num_classes=datamodule.num_classes, pretrained=False)
trainer = flash.Trainer(fast_dev_run=True)
trainer.finetune(model, datamodule=datamodule)