fork from https://github.com/junyanz/pytorch-CycleGAN-and-pix2pix
a.sh is train cmd
python train.py --dataroot ./dataset/ucf-npy/ --dataset_mode v --model pix2pix --which_model_netG unet_256 --which_direction AtoB --norm batch --niter 10 --niter_decay 10 --batchSize 1 --name virtualkittidepth --depth 7 --max_dataset_size 150 --output_nc 3 --input_nc 3 --gpu_ids 0 --data_dir ./dataset/UCF/v_BabyCrawling**.avi --load_video 1 ::run ok
Video2XYZ 已经实现: [vid2dep demo show] [vid2vid demo show]
[video generate demo: children]
正在推进: vid2sped vid2angle vid2action vid2reward vid2nlp ref: #3
更完善的期望是sensorfuse2representation, representation2XYZ, SF2Rpstt2XYZ,
对于progressGAN--pix2pixHD想法是 can train pred angle /2(左右) /4(大左小左 大右小右) /8 图像的越来越精细到功能训练的准确度越来越高。
!Curiosity-driven Exploration by Self-supervised Prediction