Context aware Interpolation Pipeline

Test-running the Context-Aware Interpolation training

Creating a TFRecord dataset

1. Download DAVIS.

2. If you wish to add more images, be sure that they are organized like DAVIS. For example, in DAVIS/JPEGImages/480p, the structure is:

   <directory to images>
       <shot_0>
           <image_0>.(png | jpg)
           ...
           <image_n>.(png | jpg)
       <shot_1>
       ...
       <shot_n>
   

3. Run the following command from the project root directory:

   python -m mains.create_davis_dataset -d path/to/DAVIS -o path/to/tfrecords_dir
   

4. Expected output in the specified output directory should be:

   <tfrecords_dir>
       ...
       0_interp_dataset_train.tfrecords
       ...
       n_interp_dataset_train.tfrecords
       
       0_interp_dataset_valid.tfrecords
       ...
       n_interp_dataset_valid.tfrecords
   

Training

1. Have your tf records and pre-trained PWC-Net weights prepared.

2. Run the following command:

   python -m mains.train_context_interp -d path/to/tfrecords_dir -c path/to/checkpoints_dir -w path/to/pwcnet_weights.npz
   

3. Launch tensorboard.

   tensorboard --logdir=path/to/checkpoints_dir