e-lab / torch-prednet

PredNet in Torch7

A model of PredNet in Torch7

get started:

Download dataset: moving MNIST examples. Small is a 100-sample test, otherwise use the larger ones with 8000 samples. This dataset originated from Viorica Patraucean and team.

to train PredNet:

Run: ``th main.lua --useGPU --batch 10 --nlayers 3 -s --savedir results --dataDir dataSets --dataName data-small``` to train on GPU on large dataset, visualize and save model and results.

to show PredNet outputs with saved model:

Run: qlua main.lua --model results --dataDir dataSets --dataName data-small --visOnly --useGPU

Notes:

• Data set format: name-train.t7 name-test.t7
• Memory usage: decrease nSeq parameter to use less memory. The model is cloned nSeq times and therefore memory usage is proportional to it.