def test_video_classifier_finetune_from_csv(tmpdir):
with mock_video_csv_file(tmpdir) as (mock_csv, total_duration):
half_duration = total_duration / 2 - 1e-9
datamodule = VideoClassificationData.from_csv(
"file",
"target",
train_file=mock_csv,
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(default_root_dir=tmpdir,
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,
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 from_kinetics(
clip_sampler: str = "uniform",
clip_duration: int = 1,
decode_audio: bool = False,
batch_size=1,
**data_module_kwargs,
) -> VideoClassificationData:
"""Downloads and loads the Kinetics data set."""
download_data("https://pl-flash-data.s3.amazonaws.com/kinetics.zip", "./data")
return VideoClassificationData.from_folders(
train_folder=os.path.join(os.getcwd(), "data/kinetics/train"),
val_folder=os.path.join(os.getcwd(), "data/kinetics/val"),
clip_sampler=clip_sampler,
clip_duration=clip_duration,
decode_audio=decode_audio,
batch_size=batch_size,
**data_module_kwargs,
)
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)
ApplyTransformToKey(key="video",
transform=K.VideoSequential(
*per_batch_transform_on_device,
data_format="BCTHW",
same_on_frame=False)),
]),
}
# 3. Load the data from directories.
datamodule = VideoClassificationData.from_folders(
train_folder=os.path.join(flash.PROJECT_ROOT, "data/kinetics/train"),
val_folder=os.path.join(flash.PROJECT_ROOT, "data/kinetics/val"),
predict_folder=os.path.join(flash.PROJECT_ROOT,
"data/kinetics/predict"),
train_transform=make_transform(train_post_tensor_transform),
val_transform=make_transform(val_post_tensor_transform),
predict_transform=make_transform(val_post_tensor_transform),
batch_size=8,
clip_sampler="uniform",
clip_duration=1,
video_sampler=RandomSampler,
decode_audio=False,
num_workers=8)
# 4. List the available models
print(VideoClassifier.available_backbones())
# out: ['efficient_x3d_s', 'efficient_x3d_xs', ... ,slowfast_r50', 'x3d_m', 'x3d_s', 'x3d_xs']
print(VideoClassifier.get_backbone_details("x3d_xs"))
# 5. Build the VideoClassifier with a PyTorchVideo backbone.
model = VideoClassifier(backbone="x3d_xs",
num_classes=datamodule.num_classes,
import os
import torch
import flash
from flash.core.data.utils import download_data
from flash.video import VideoClassificationData, VideoClassifier
# 1. Create the DataModule
# Find more datasets at https://pytorchvideo.readthedocs.io/en/latest/data.html
download_data("https://pl-flash-data.s3.amazonaws.com/kinetics.zip", "./data")
datamodule = VideoClassificationData.from_folders(
train_folder=os.path.join(os.getcwd(), "data/kinetics/train"),
val_folder=os.path.join(os.getcwd(), "data/kinetics/val"),
clip_sampler="uniform",
clip_duration=1,
decode_audio=False,
)
# 2. Build the task
model = VideoClassifier(backbone="x3d_xs",
num_classes=datamodule.num_classes,
pretrained=False)
# 3. Create the trainer and finetune the model
trainer = flash.Trainer(max_epochs=3, gpus=torch.cuda.device_count())
trainer.finetune(model, datamodule=datamodule, strategy="freeze")
# 4. Make a prediction
predictions = model.predict(os.path.join(os.getcwd(), "data/kinetics/predict"))
import flash
from flash.core.data.utils import download_data
from flash.core.integrations.labelstudio.visualizer import launch_app
from flash.video import VideoClassificationData, VideoClassifier
# 1 Download data
download_data(
"https://label-studio-testdata.s3.us-east-2.amazonaws.com/lightning-flash/video_data.zip"
)
# 2. Load export data
datamodule = VideoClassificationData.from_labelstudio(
export_json="data/project.json",
data_folder="data/upload/",
val_split=0.2,
clip_sampler="uniform",
clip_duration=1,
decode_audio=False,
)
# 3. Build the task
model = VideoClassifier(
backbone="slow_r50",
num_classes=datamodule.num_classes,
)
# 4. Create the trainer and finetune the model
trainer = flash.Trainer(max_epochs=3)
trainer.finetune(model, datamodule=datamodule, strategy="freeze")
# 5. Make a prediction
# See the License for the specific language governing permissions and
# limitations under the License.
import torch
import flash
from flash.core.data.utils import download_data
from flash.video import VideoClassificationData, VideoClassifier
# 1. Create the DataModule
# Find more datasets at https://pytorchvideo.readthedocs.io/en/latest/data.html
download_data("https://pl-flash-data.s3.amazonaws.com/kinetics.zip", "./data")
datamodule = VideoClassificationData.from_folders(
train_folder="data/kinetics/train",
val_folder="data/kinetics/val",
clip_sampler="uniform",
clip_duration=1,
decode_audio=False,
batch_size=1,
)
# 2. Build the task
model = VideoClassifier(backbone="x3d_xs", labels=datamodule.labels, pretrained=False)
# 3. Create the trainer and finetune the model
trainer = flash.Trainer(
max_epochs=1, gpus=torch.cuda.device_count(), strategy="ddp" if torch.cuda.device_count() > 1 else None
)
trainer.finetune(model, datamodule=datamodule, strategy="freeze")
# 4. Make a prediction
datamodule = VideoClassificationData.from_folders(predict_folder="data/kinetics/predict", batch_size=1)
ApplyTransformToKey(key="video",
transform=K.VideoSequential(
*per_batch_transform_on_device,
data_format="BCTHW",
same_on_frame=False)),
]),
}
# 3. Load the data from directories.
datamodule = VideoClassificationData.from_paths(
train_data_path=os.path.join(_PATH_ROOT, "data/kinetics/train"),
val_data_path=os.path.join(_PATH_ROOT, "data/kinetics/val"),
predict_data_path=os.path.join(_PATH_ROOT, "data/kinetics/predict"),
clip_sampler="uniform",
clip_duration=2,
video_sampler=RandomSampler,
decode_audio=False,
train_transform=make_transform(train_post_tensor_transform),
val_transform=make_transform(val_post_tensor_transform),
predict_transform=make_transform(val_post_tensor_transform),
num_workers=8,
batch_size=8,
)
# 4. List the available models
print(VideoClassifier.available_models())
# out: ['efficient_x3d_s', 'efficient_x3d_xs', ... ,slowfast_r50', 'x3d_m', 'x3d_s', 'x3d_xs']
print(VideoClassifier.get_model_details("x3d_xs"))
# 5. Build the model - `x3d_xs` comes with `nn.Softmax` by default for their `head_activation`.
model = VideoClassifier(model="x3d_xs", num_classes=datamodule.num_classes)
model.serializer = Labels()
