tensorflow 2.0 version (use tf.compat.v1)

Requirement

require packages : environment.yml or conda-gpu-cuda10.yml (in preprocess folder) we tested on Cuda 10.0 + cudnn 7.5 + T4

$conda env create -f environment.yml or conda_gpu_cuda10.yml

tf 2.0.0 module don't support tf.contrib, so we change slim.conv2d to layers.conv2d (temporary)

inference of original pre-trained model

# change pre-trained model variable name using preprocess/change_conv2d.py
python main_woflow_up.py --model ckpt_woflow --use_gpu 1 --test_dir test_sample0

Training

prepare

make optical flow map (more information in pytorch_pwc folder)

use pytorch_pwc/run_data.py

training

python main_whole_up.py --h

inference

python test_main_whole.py --h