2 This is an implementation of a deep residual network for predicting
3 the dynamics of 2D shapes as described in
5 F. Fleuret. Predicting the dynamics of 2d objects with a deep
6 residual network. CoRR, abs/1610.04032, 2016.
8 https://arxiv.org/abs/1610.04032
10 This package is composed of a simple 2d physics simulator called
11 'flatland' written in C++, to generate the data-set, and a deep
12 residual network 'dyncnn' written in the Lua/Torch7 framework.
14 You can run the reference experiment by executing the run.sh shell
19 (1) Generate the data-set of 40k triplets of images,
21 (2) Train the deep network, and output validation results every 100
22 epochs. This takes ~30h on a GTX 1080 with cuda 8.0, cudnn 5.1,
25 (3) Generate two pictures of the internal activations.
27 (4) Generate a graph with the loss curves if gnuplot is installed.