[WIP, Policy] Docstring and refactoring on top of PR 327 #335
Conversation
| self.kernel_sizes = config.get("kernel_sizes", [(3, 3)] * self.n_layers) | ||
| self.strides = config.get("strides", [(1, 1)] * self.n_layers) | ||
| # Environment | ||
| # TODO: rethink whether storing the whole environment is needed | ||
| self.env = env |
There was a problem hiding this comment.
We might not need to store the env most of the time actually. But sometimes one might need to know more about the environment configuration to induce inductive bias to their policy model. In the CNN Policy, we use to access the grid dimension like self.ev.height, self.env.width but is not crucial.
There was a problem hiding this comment.
+1, I think storing the whole env is likely to duplicate a lot of info unnecessarily, especially if you're trying to do multiprocessing stuff down the line.
gflownet/policy/base.py
Outdated
| Base Policy class for GFlowNet policy models. | ||
| """ | ||
|
|
||
| from typing import Union |
There was a problem hiding this comment.
I think you also need to import Tuple here
|
@josephdviviano finalizing this PR will hopefully close both #327 and #335 |
| self.kernel_sizes = config.get("kernel_sizes", [(3, 3)] * self.n_layers) | ||
| self.strides = config.get("strides", [(1, 1)] * self.n_layers) | ||
| # Environment | ||
| # TODO: rethink whether storing the whole environment is needed | ||
| self.env = env |
There was a problem hiding this comment.
+1, I think storing the whole env is likely to duplicate a lot of info unnecessarily, especially if you're trying to do multiprocessing stuff down the line.
josephdviviano
merged commit c9ec03f
into
293-flexible-policy-definition
This PR incorporates docstring and refactoring of the code on top of PR #327
Summary of refactoring:
parse_config()and included its content in__init__()to make the handling of all the attributes of a policy more explicit.instantiate()and combined withmake_*()(e.g.make_mlp,make_cnn) into a single method common to all policies,make_model(), which instantiates and returns the corresponding model.🚧 Work in progress 🚧