Demo code along with pretrained models for "Video Prediction via Selective Sampling" (NeurIPS 2018)
Following the final review comments, we keep updating the code to make it more efficient and much lighter. To be specific, we want to unify "Sampler" and "Kernel Generator", and unify "Selector" and "Combiner", which avoids complex pre-training precedure for each sub-module. Therefore the detailed module implementation is different from that in orginal paper described. But the main idea is unchanged behind the code, which is sampling and selection.
For MovingMnist Datasets:
The used data is here:
https://www.dropbox.com/s/qqh4x3uq049z956/mnist.h5?dl=0
The pretrained model for one-digit moving (given 2 frames to predict 10 frames) is here:
https://1drv.ms/u/s!AnsWsC45wa-nbhT0AXdTCfsI_UQ
To run the demo code:
put the data into ./Data;
put the pretained model (should be unzipped first) into ./PretrainedModels;
run: python generate_vpss.py.
To train the demo code:
put the data into ./Data;
run: python train_sampler.py
run: python train_selector.py
For RobotPush Datasets:
The used data [1] is here:
https://sites.google.com/site/brainrobotdata/home/push-dataset
The pretrained model (given 2 frames to predict 10 frames) is here:
https://1drv.ms/u/s!AnsWsC45wa-nb_cGCaWMRSbyhgk
To run the demo code:
put the data into ./Data;
put the pretained model (should be unzipped first) into ./PretrainedModels;
run: python generate_vpss_RobotPush.py.
To train the demo code:
put the data into ./Data;
run: python train_feat_vpss_RobotPush.py
run: python train_sampler_vpss_RobotPush.py
run: python train_selector_vpss_RobotPush.py
TO DO: Demo code for Human3.6M Datasets.
[1] Unsupervised Learning for Physical Interaction through Video Prediction, Chelsea Finn, Ian Goodfellow, Sergey Levine, NIPS, 2016.