RAFT (Recurrent All Pairs Field Transforms for Optical Flow, Teed et. al., ECCV2020) implementation via tf.keras
$ pip install tf-raft
or you can simply clone this repository.
- TensorFlow
- TensorFlow-addons
- albumentations
see details in pyoroject.toml
MPI-Sintel or FlyingChairs datasets are relatively light. See more datasets in the oirignal repository
from tf_raft.model import RAFT, SmallRAFT
from tf_raft.losses import sequence_loss, end_point_error
# iters/iters_pred are the number of recurrent update of flow in training/prediction
raft = RAFT(iters=iters, iters_pred=iters_pred)
raft.compile(
optimizer=optimizer,
clip_norm=clip_norm,
loss=sequence_loss,
epe=end_point_error
)
raft.fit(
ds_train,
epochs=epochs,
callbacks=callbacks,
steps_per_epoch=train_size//batch_size,
validation_data=ds_val,
validation_steps=val_size
)In practice, you are required to prepare dataset, optimizer, callbacks etc, check details in train_sintel.py or train_chairs.py.
train_chairs.py and train_sintel.py train RAFT model via YAML configuration. Sample configs are in configs directory. Run;
$ python train_chairs.py /path/to/config.ymlI made the pre-trained weights (on both FlyingChairs and MPI-Sintel) public.
You can download them via gsutil or curl.
$ gsutil cp -r gs://tf-raft-pretrained/2020-09-26T18-38/checkpoints .or
$ mkdir checkpoints
$ curl -OL https://storage.googleapis.com/tf-raft-pretrained/2020-09-26T18-38/checkpoints/model.data-00000-of-00001
$ curl -OL https://storage.googleapis.com/tf-raft-pretrained/2020-09-26T18-38/checkpoints/model.index
$ mv model* checkpoints/$ gsutil cp -r gs://tf-raft-pretrained/2020-09-26T08-51/checkpoints .or
$ mkdir checkpoints
$ curl -OL https://storage.googleapis.com/tf-raft-pretrained/2020-09-26T08-51/checkpoints/model.data-00000-of-00001
$ curl -OL https://storage.googleapis.com/tf-raft-pretrained/2020-09-26T08-51/checkpoints/model.index
$ mv model* checkpoints/raft = RAFT(iters=iters, iters_pred=iters_pred)
raft.load_weights('checkpoints/model')
# forward (with dummy inputs)
x1 = np.random.uniform(0, 255, (1, 448, 512, 3)).astype(np.float32)
x2 = np.random.uniform(0, 255, (1, 448, 512, 3)).astype(np.float32)
flow_predictions = model([x1, x2], training=False)
print(flow_predictions[-1].shape) # >> (1, 448, 512, 2)Though I have tried to reproduce the original implementation faithfully, there is some difference between the original one and mine (mainly because of used framework: PyTorch/TensorFlow);
- The original implementations provides cuda-based correlation function but I don't. My TF-based implementation works well, but cuda-based one may run faster.
- I have trained my model on FlyingChairs and MPI-Sintel separately in my private environment (GCP with P100 accelerator). The model has been trained well, but not reached the best score reported in the paper (trained on multiple datasets).
- The original one uses mixed-precision. This may get training much faster, but I don't. TensorFlow also enables mixed-precision with few additional lines, see https://www.tensorflow.org/guide/mixed_precision if interested.
Additional, global gradient clipping seems to be essential for stable training though it is not emphasized in the original paper. This operation can be done via torch.nn.utils.clip_grad_norm_(model.parameters(), clip) in PyTorch, tf.clip_by_global_norm(grads, clip_norm) in TF (coded at self.train_step in tf_raft/model.py).