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

new ref https://github.com/createamind/keras-cpcgan