Help tool: debug and plot network activity

[snowgoon88]

Hello,

I've made myself a "tool" to help me understand/debug/build player behavior. The idea is to have, along side, the agent moving in the top-down view and some plots which monitor some selected neurons activity.

Running:

$> python debug_plot.py info_switcher.json

using this JSON file will automatically use the model from player_switcher.player_switcher() for debug+plot

{
    "player": {
        "module": "player_switcher",
        "func":   "switcher_player"
    },
    "neurons": {
        "nni": {
            "indices": [0, 63, 33, 65],
            "labels":  ["L", "R", "C", "nrj"],
            "lim":     [-0.05, 1.05]
        },
        "nnx": {
            "indices": [77, 96],
            "labels":  ["X_f", "X_t"],
            "lim":     [-0.05, 1.05]
        },
        "nno": {
            "indices": [0],
            "labels":  ["out"],
            "lim":     [-6.1, 6.1]
        }
    }

and gets me this kind of output https://github.com/snowgoon88/braincraft/blob/dev_graph/data/fig_debug_plot_switcher.png

I am not sure this can find its place in the braincraft repo, but you can play with it (in the dev_graph branch of my forked repo: https://github.com/snowgoon88/braincraft/tree/dev_graph)

warning: I had to slightly alter "player_switcher.py" and "player_switcher_alt.py" to make them work with "debug_plot.py", but not changing at all their model. Added import numpy as np at the beginning of files and make sure that the output Ovector is a 2D numpy array. (@vforch uses simple 1D vector).

[rougier]

Thanks, that might be useful for other participants. Maybe a simple solution would be for the debug argument to be a function to be called instead of a bool. It could default to None and have the current debug as an option and your as a second option.

[snowgoon88]

Hello,
yes, I could do that: use a function as the debug argument.

It could also pave the way to a tool that could be used with args on the command line. Like, for example:

$> python run_player.py --debug_3d player_switcher

# or

$> pyton run_player.py --debug_plot --seed 78 --nb_run 100 player_switcher

# or even

$> python run_player.py --make_graph player_switcher

where player_switcher.py would have a function generate_model() that returns the model and a generate_info() that could be used for debugging. So eventually, no need to copy any braincraft code into player_xxx.py files.

Oh, and by the way, I have also a tool to generate a graph of the network. See for example: where it is quite clear that "my" switcher model could be streamlined :o)

[rougier]

That's a lot of tools! Maybe best would be list them in a pinned issue.