Newer
Older
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
# Weight Agnostic Neural Networks
Weight Agnostic Networks: network topologies evolved to work with a variety of shared weights. Adapted from the [prettyNEAT](../prettyNEAT) package.
This repository represents a snapshot of the code used to produce the results in the original [WANN paper](https://weightagnostic.github.io/). The goal was to create minimal rather than general code for clearer understanding. If you are interested in a more extendable and general implementation for your own experiments, we recommend using the [WANN fork](../prettyNEAT_WANN) of the [prettyNEAT](../prettyNEAT) package.
## Dependencies
Core algorithm tested with:
- Python 3.5.3
- NumPy 1.15.2 (`pip install numpy`)
- mpi4py 3.0.1 (`pip install mpi4py`)
- OpenAI Gym 0.9.6 (`pip install gym` -- installation details [here](https://github.com/openai/gym))
Domains tested with:
- Cart-pole Swing-up (included, but requires OpenAI gym)
- Bipedal Walker: Box2d (see OpenAI gym installation)
- Quadruped (Ant) Walker: PyBullet 1.6.3 (`pip install pybullet`)
- MNIST: Mnist utilities 0.2.2 (`pip install mnist`)
- VAE Racer:
- Tensorflow 1.8 (`pip install tensorflow==1.8.0`)
- Pretrained VAE (in [wannRelease](../) -- copy to root to use, e.g: `cp -r ../vae .`)
## Training Weight Agnostic Neural Networks
To get started and see that everything is set up you can test the swing up domain:
```
python wann_train.py
```
which is the same as the default hyperparameter:
```
python wann_train.py -p p/laptop_swing.json -n 8
```
Where `8` is the number of workers you have on your computer, and `p/laptop_swing.json` contains hyperparameters.
Evaluation of the population is embarrassingly parallel so every extra core you have will really speed things up. Here is an example training curve on cart-pole with a population of 64 on an 8 core laptop:
Where `Fitness` is the mean reward earned over all trials and weight values. `Median Fitness` is the fitness of the median performing member of the population, `Max Fitness` is the fitness of the best performing member of the population, `Top Fitness` is the best performing member ever found. `Peak Fitness` is the mean reward earned by the best performing member with its best performing weight value. To reproduce this graph see this [jupyter notebook](log/viewRunStats.ipynb).
The full list of hyperparameters and their meaning is explained in [hypkey.txt](p/hypkey.txt)
## Testing and Tuning Weight Agnostic Neural Networks
To view or test a WANN:
```
python wann_test.py -p p/swingup.json -i champions/swing.out --nReps 3 --view True
```
WANNs are saved as 2D numpy arrays and can be retested, train, and viewed using the [WANNTool](../WANNTool) provided.
---
## Amira
### General
To run on windows change line 236 in `wann_train.py` to `subprocess.check_call(["mpiexec","-np", str(n), sys.executable] +['-u']+ sys.argv, env=env)` instead of `subprocess.check_call(["mpirun","-np", str(n), sys.executable] +['-u']+ sys.argv, env=env)`
### Frep
```
python wann_train.py -p p/frep.json -n 2
```
### Rover training
### How to add new training "myName"
1. in `./p` add `myName.json` file and change task name
2. in `./domain`
1. add `myName.py`
1. change class to `myNameEnv`
2. in `make_env.py`
1. add else if statement
3. in `config.py`
1. add games['myName'] entry and change layer sizes and activation function types
### how to show the results/plot graph