Error creating new variables using as_networkx()

[EloyAnguiano]

Environment

• Grid2op version: 1.7.0
• Stable Baselines 3 version: 1.4.0
• System: Ubuntu 18.04 LTS
• Python 3.9.12

Bug description

I'm trying to add some features to the space using the as_networkx method but sometimes there are some problems with that conversion.

Command line

# python script.py

Code snippet

import torch as th
from stable_baselines3 import SAC
from stable_baselines3.common.torch_layers import BaseFeaturesExtractor
import gym
from grid2op import make
from grid2op.gym_compat import BoxGymActSpace, GymEnv
from copy import deepcopy
import networkx as nx
from gym.spaces import Box
import numpy as np
import torch.nn as nn

def build_matrix(obs, key):
    if isinstance(obs, nx.Graph):
        graph = obs
    else:
        graph = obs.as_networkx()

    matrix = nx.attr_matrix(graph, edge_attr=key)[0]
    return matrix

class FeatureExtractor(BaseFeaturesExtractor):
    def __init__(self, observation_space:  gym.spaces.Dict, features_dim: int, device):
        super(FeatureExtractor, self).__init__(
            observation_space, features_dim)

        self.device = device

        space = observation_space["rho_matrix"]
        n = space.shape[0] * space.shape[1]

        self.last_dense = nn.Sequential(
            nn.Linear(n, self.features_dim),
            nn.ReLU())


    def forward(self, observations: gym.spaces.Dict) -> th.Tensor:
        return self.last_dense.forward(observations["rho_matrix"].flatten(1))

POLICY = {  "features_extractor_class": FeatureExtractor,
            "features_extractor_kwargs": {
                "features_dim": 256,
                "device": "cpu",
            },
            "activation_fn": th.nn.ReLU}
if __name__ == "__main__":

    grid2op_env = make("rte_case5_example", test=True)

    env = GymEnv(grid2op_env)  # create the gym environment

    # CHANGING ACTION SPACE
    env.action_space = BoxGymActSpace(env.init_env.action_space)

    # CHANGING OBSERVATION SPACE
    # Adding vars
    grid2op_obs = grid2op_env.reset()

    # Adding vars
    graph = deepcopy(grid2op_obs).as_networkx()

    # graph_vars = {}
    graph_vars = {"rho":(-1000, 1000)}

    for k in graph_vars.keys():
        shape_ = build_matrix(graph, k).shape
        env.observation_space.add_key(f"{k}_matrix",
                                        lambda obs: build_matrix(obs, k),
                                        Box(shape=shape_, 
                                            low=np.full(shape_, graph_vars[k][0]),
                                            high=np.full(shape_, graph_vars[k][1])))

    agent = SAC("MultiInputPolicy",env, policy_kwargs=POLICY)
    agent.learn(total_timesteps=1e5)

Current output

ValueError: zero-size array to reduction operation maximum which has no identity

[BDonnot]

Hello,

Thanks for notifying us of this bug.

It's rather easy to reproduce with your code snippet. Thanks. Hopefully i'll be able to fix it for the next release.