Tensorflow implementation of DrNet
ssh ... Build the Docker image with
$ docker build --no-cache -t img-dockername-video .
Build the docker container:
$ NV_GPU=0,2 nvidia-docker run -it -v ~/projects/drnet-tflow/:/drnet-tflow/ -v /mnt/:/mnt/ -p 8894:8888 --name container-name img-dockername-video bash
python run.py --num_gpus 1 --batch_size 50 --size_pose_embedding 5 --size_content_embedding 128 --max_steps 12 --num_epochs 2000 --run_name r1
python run_lstm.py --num_gpus 1 --batch_size 40 --size_post_embedding 5 --size_content_embedding 128 --evaluate False --training True --num_epocs 10000 -- run_name r1_lstm --restore_dir_D /some/dir/D --restore_dir_Ep /some/dir/Ep --restore_dir_Ec /some/dir/Ec
Other hyperparameters are described in the run.py and run_lstm.py files.
Pose encoder Dimensions = 5 Content encoder Dimensions = 128 DCGAN Unet + DCGAN Pose Encoder
Decoder Loss
After 60k iterations:
Original frame on the left, frame to be decoded the middle, decoder output on the right
Two 512 dimensional LSTMs, with tanh dense layer on top
L2 loss for predicted pose encodings:
Example of some generated video frames:
