This code reimplements the Prednet model (mostly copy & paste): [Deep Predictive Coding Networks for Video Prediction and Unsupervised Learning] [github], in gridworld sequence data.
I adopt the source implementations mostly, except:
- Using
tf.contrib.keras(it's all the same if you use 'TF' backends in Keras) - Bug fixed. (extrap_start_time, see [this])
- New data_generator in Gridworld Simulator. The grid-world simulator simulates a free energy loss physics in a 2D scene.
- Add roi-loss.
All the hyper-parameters are in train.py.
Run:
python train.py
Evaluate:
python evaluate.py
- Python 2
- Tensorflow 1.1 (keras in tf.contrib)
- Pygame
- Numpy
I completed this code when I was an intern at Horizon Robotics. I will greatly thank the paper of William Lotter in ICLR 2017, and his implementations:PredNet. Also greately thanks my mentor Penghong Lin, and other colleagues (Lisen Mu, Jingchu Liu and Henglai Wei) for helpful discussion.