def test_uniform_crop(self):
# For videos with height < width.
video = thwc_to_cthw(create_dummy_video_frames(
20, 30, 40)).to(dtype=torch.float32)
# Left crop.
actual = uniform_crop(video, size=20, spatial_idx=0)
self.assertTrue(actual.equal(video[:, :, 5:25, :20]))
# Center crop.
actual = uniform_crop(video, size=20, spatial_idx=1)
self.assertTrue(actual.equal(video[:, :, 5:25, 10:30]))
# Right crop.
actual = uniform_crop(video, size=20, spatial_idx=2)
self.assertTrue(actual.equal(video[:, :, 5:25, 20:]))
# For videos with height > width.
video = thwc_to_cthw(create_dummy_video_frames(
20, 40, 30)).to(dtype=torch.float32)
# Top crop.
actual = uniform_crop(video, size=20, spatial_idx=0)
self.assertTrue(actual.equal(video[:, :, :20, 5:25]))
# Center crop.
actual = uniform_crop(video, size=20, spatial_idx=1)
self.assertTrue(actual.equal(video[:, :, 10:30, 5:25]))
# Bottom crop.
actual = uniform_crop(video, size=20, spatial_idx=2)
self.assertTrue(actual.equal(video[:, :, 20:, 5:25]))
def test_reading_from_directory_structure(self, decoder):
# For an unknown reason this import has to be here for `buck test` to work.
import torchvision.io as io
with tempfile.TemporaryDirectory() as root_dir:
# Create test directory structure with two classes and a video in each.
root_dir_name = pathlib.Path(root_dir)
test_class_1 = root_dir_name / "running"
test_class_1.mkdir()
data_1 = create_dummy_video_frames(15, 10, 10)
test_class_2 = root_dir_name / "cleaning windows"
test_class_2.mkdir()
data_2 = create_dummy_video_frames(20, 15, 15)
with tempfile.NamedTemporaryFile(
suffix=".mp4",
dir=test_class_1) as f_1, tempfile.NamedTemporaryFile(
suffix=".mp4", dir=test_class_2) as f_2:
f_1.close()
f_2.close()
# Write lossless video for each class.
io.write_video(
f_1.name,
data_1,
fps=30,
video_codec="libx264rgb",
options={"crf": "0"},
)
io.write_video(
f_2.name,
data_2,
fps=30,
video_codec="libx264rgb",
options={"crf": "0"},
)
clip_sampler = make_clip_sampler("uniform", 3)
labeled_video_paths = LabeledVideoPaths.from_path(root_dir)
dataset = LabeledVideoDataset(
labeled_video_paths,
clip_sampler=clip_sampler,
video_sampler=SequentialSampler,
decode_audio=False,
decoder=decoder,
)
# Videos are sorted alphabetically so "cleaning windows" (i.e. data_2)
# will be first.
sample_1 = next(dataset)
self.assertEqual(sample_1["label"], 0)
self.assertTrue(sample_1["video"].equal(
thwc_to_cthw(data_2).to(torch.float32)))
sample_2 = next(dataset)
self.assertEqual(sample_2["label"], 1)
self.assertTrue(sample_2["video"].equal(
thwc_to_cthw(data_1).to(torch.float32)))
def test_uniform_temporal_subsample(self):
video = thwc_to_cthw(create_dummy_video_frames(
20, 30, 40)).to(dtype=torch.float32)
actual = uniform_temporal_subsample(video, video.shape[1])
self.assertTrue(actual.equal(video))
video = thwc_to_cthw(create_dummy_video_frames(
20, 30, 40)).to(dtype=torch.float32)
actual = uniform_temporal_subsample(video, video.shape[1] // 2)
self.assertTrue(
actual.equal(video[:, [0, 2, 4, 6, 8, 10, 12, 14, 16, 19]]))
video = thwc_to_cthw(create_dummy_video_frames(
20, 30, 40)).to(dtype=torch.float32)
actual = uniform_temporal_subsample(video, 1)
self.assertTrue(actual.equal(video[:, 0:1]))
def test_compose_with_video_transforms(self):
video = thwc_to_cthw(create_dummy_video_frames(
20, 30, 40)).to(dtype=torch.float32)
test_clip = {"video": video, "label": 0}
# Compose using torchvision and pytorchvideo transformst to ensure they interact
# correctly.
num_subsample = 10
transform = Compose([
ApplyTransformToKey(
key="video",
transform=Compose([
UniformTemporalSubsample(num_subsample),
NormalizeVideo([video.mean()] * 3, [video.std()] * 3),
RandomShortSideScale(min_size=15, max_size=25),
RandomCropVideo(10),
RandomHorizontalFlipVideo(p=0.5),
]),
)
])
actual = transform(test_clip)
c, t, h, w = actual["video"].shape
self.assertEqual(c, 3)
self.assertEqual(t, num_subsample)
self.assertEqual(h, 10)
self.assertEqual(w, 10)
def test_repeat_temporal_frames_subsample(self):
video = thwc_to_cthw(create_dummy_video_frames(
32, 10, 10)).to(dtype=torch.float32)
actual = repeat_temporal_frames_subsample(video, (1, 4))
expected_shape = ((3, 32, 10, 10), (3, 8, 10, 10))
for idx in range(len(actual)):
self.assertEqual(actual[idx].shape, expected_shape[idx])
def test_normalize(self):
video = thwc_to_cthw(create_dummy_video_frames(
10, 30, 40)).to(dtype=torch.float32)
transform = Normalize(video.mean(), video.std())
actual = transform(video)
self.assertAlmostEqual(actual.mean().item(), 0)
self.assertAlmostEqual(actual.std().item(), 1)
def _init_benchmark_short_side_scale(**kwargs) -> Callable:
x = thwc_to_cthw(
create_dummy_video_frames(
kwargs["temporal_size"],
kwargs["ori_spatial_size"][0],
kwargs["ori_spatial_size"][1],
)).to(dtype=torch.float32)
def func_to_benchmark() -> None:
_ = short_side_scale(x, kwargs["dst_short_size"])
return
return func_to_benchmark
def test_uniform_crop_transform(self):
video = thwc_to_cthw(create_dummy_video_frames(
10, 30, 40)).to(dtype=torch.float32)
test_clip = {"video": video, "aug_index": 1, "label": 0}
transform = UniformCropVideo(20)
actual = transform(test_clip)
c, t, h, w = actual["video"].shape
self.assertEqual(c, 3)
self.assertEqual(t, 10)
self.assertEqual(h, 20)
self.assertEqual(w, 20)
self.assertTrue(actual["video"].equal(video[:, :, 5:25, 10:30]))
def test_torchscriptable_input_output(self):
video = thwc_to_cthw(create_dummy_video_frames(20, 30, 40)).to(
dtype=torch.float32
)
# Test all the torchscriptable tensors.
for transform in [UniformTemporalSubsample(10), RandomShortSideScale(10, 20)]:
transform_script = torch.jit.script(transform)
self.assertTrue(isinstance(transform_script, torch.jit.ScriptModule))
# Seed before each transform to force determinism.
torch.manual_seed(0)
output = transform(video)
torch.manual_seed(0)
script_output = transform_script(video)
self.assertTrue(output.equal(script_output))
def test_short_side_scale_height_shorter_opencv(self):
video = thwc_to_cthw(create_dummy_video_frames(
20, 10, 20)).to(dtype=torch.float32)
actual = short_side_scale(video, 5, backend="opencv")
self.assertEqual(actual.shape, (3, 20, 5, 10))
def test_short_side_scale_equal_size_pytorch(self):
video = thwc_to_cthw(create_dummy_video_frames(
20, 10, 10)).to(dtype=torch.float32)
actual = short_side_scale(video, 10, backend="pytorch")
self.assertEqual(actual.shape, (3, 20, 10, 10))
def test_reading_from_directory_structure_hmdb51(self, decoder):
# For an unknown reason this import has to be here for `buck test` to work.
import torchvision.io as io
with tempfile.TemporaryDirectory() as root_dir:
# Create test directory structure with two classes and a video in each.
root_dir_name = pathlib.Path(root_dir)
action_1 = "running"
action_2 = "cleaning_windows"
videos_root_dir = root_dir_name / "videos"
videos_root_dir.mkdir()
test_class_1 = videos_root_dir / action_1
test_class_1.mkdir()
data_1 = create_dummy_video_frames(15, 10, 10)
test_class_2 = videos_root_dir / action_2
test_class_2.mkdir()
data_2 = create_dummy_video_frames(20, 15, 15)
test_splits = root_dir_name / "folds"
test_splits.mkdir()
with tempfile.NamedTemporaryFile(
suffix="_u_nm_np1_ba_goo_19.avi",
dir=test_class_1) as f_1, tempfile.NamedTemporaryFile(
suffix="_u_nm_np1_fr_med_1.avi",
dir=test_class_2) as f_2:
f_1.close()
f_2.close()
# Write lossless video for each class.
io.write_video(
f_1.name,
data_1,
fps=30,
video_codec="libx264rgb",
options={"crf": "0"},
)
io.write_video(
f_2.name,
data_2,
fps=30,
video_codec="libx264rgb",
options={"crf": "0"},
)
_, video_name_1 = os.path.split(f_1.name)
_, video_name_2 = os.path.split(f_2.name)
with open(
os.path.join(test_splits,
action_1 + "_test_split1.txt"), "w") as f:
f.write(f"{video_name_1} 1\n")
with open(
os.path.join(test_splits,
action_2 + "_test_split1.txt"), "w") as f:
f.write(f"{video_name_2} 1\n")
clip_sampler = make_clip_sampler("uniform", 3)
dataset = Hmdb51(
data_path=test_splits,
video_path_prefix=root_dir_name / "videos",
clip_sampler=clip_sampler,
video_sampler=SequentialSampler,
split_id=1,
split_type="train",
decode_audio=False,
decoder=decoder,
)
# Videos are sorted alphabetically so "cleaning windows" (i.e. data_2)
# will be first.
sample_1 = next(dataset)
sample_2 = next(dataset)
self.assertTrue(sample_1["label"] in [action_1, action_2])
if sample_1["label"] == action_2:
sample_1, sample_2 = sample_2, sample_1
self.assertEqual(sample_1["label"], action_1)
self.assertEqual(5, len(sample_1["meta_tags"]))
self.assertTrue(sample_1["video"].equal(
thwc_to_cthw(data_1).to(torch.float32)))
self.assertEqual(sample_2["label"], action_2)
self.assertEqual(5, len(sample_2["meta_tags"]))
self.assertTrue(sample_2["video"].equal(
thwc_to_cthw(data_2).to(torch.float32)))