Implementation of Deep Tracking: Seeing Beyond Seeing Using Recurrent Neural Networks, Ondruska & Posner 2016, using Python.
Doesn't completely work yet, it reproduces the early parts of training pretty well but runs into numerical instability in later stages.
####Requires:
- numpy
- theano
- lasagne
- h5py
If using original dataset:
chmod +x get_data.sh && ./get_data.sh
If generating data:
python generate_moving_objects.py --dest DATA_FILE --filestype hdf5 --frames 3000
To train:
python train.py
Options:
--num_epochs, default 10
--seq_len, default 100
--batch_size, default 2