Tensorflow Implementation of Visual Interaction Networks from Google Deepmind.

Implementation is on Tensorflow r1.2.

https://arxiv.org/abs/1706.01433

"Another key part of relational reasoning involves predicting the future in a physical scene. From just a glance, humans can infer not only what objects are where, but also what will happen to them over the upcoming seconds, minutes and even longer in some cases. For example, if you kick a football against a wall, your brain predicts what will happen when the ball hits the wall and how their movements will be affected afterwards (the ball will ricochet at a speed proportional to the kick and - in most cases - the wall will remain where it is)." From an article of Deepmind

https://deepmind.com/blog/neural-approach-relational-reasoning/

For changing configure values, please check constants script.

cat constracts.py

For generating images and data,

python physical_engines.py

For modeling Visual Interaction Networks

python gravity_vin.py

The Data are gathered from my own implemented physics engine, which is same of interaction network git repo. https://github.com/jaesik817/Interaction-networks_tensorflow

One different thing from IN physics engines is time difference between frames, which was 0.0001 in IN repo and is 0.001 in this. Because 0.0001 secs frame cannot be recognized in 32 x 32 images.

Settings are written in constants.py and gravity_vin.py. The number of objects, frames on each simulations and rollout frames are 3, 50 and 20. The training max epoches are 80000. In physical_engines code, every frames are saved as image and coded data, and those things are used in gravity_vin script. Each image has background ones from CIFAR 10 training data set as the paper.

The loss decreased as followed, which is summarized value of losses on near future 8 frames and encoding-decoding losses on input images.

The quilititive results are as followed.

True :

Modeling :