House3D_RoomNav_baseline
This is the baseline model of the RoomNav task using House3D that I implemented.
It implements A3C with gated-LSTM policy for discrete actions.
In the paper, they used 120 or 200 processes but I only used 20 processes. https://arxiv.org/abs/1801.02209
requirements
- python 3.6+
- pytorch 0.4.1
pip install -r requirements.txt
you should input your path to config.json
and your gpu ids depending on your environment in Class Params() in main.py . (In my case, I used gpu 0 for tests and gpu 1, 2, 3 for training)
Training
python main.py
Success rate
Each line represents each env used in the learning.
Todo
- add batch structure 64(batch)*30(continuous actions)
- add decay the learning rate by a factor of 1.5 when the difference of KL-divergence becomes larger than 0.01