Conversation
Updated: methods/catalog/__init__.py To do: Need to updated experiments/experimental_setup.yaml, experiments/run_experiment.py
which is Python 3.7 / PyTorch 1.x compatible
|
@codex review |
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| # Load trained Transformer | ||
| self._transformer = self._load_transformer(self._params["model_path"]) | ||
|
|
||
| # Store device | ||
| self._device = torch.device("cuda" if torch.cuda.is_available() else "cpu") | ||
| self._transformer = self._transformer.to(self._device) | ||
| self._transformer.eval() | ||
|
|
||
|
|
||
| def _load_transformer(self, model_path: str) -> GenReTransformer: | ||
| """ | ||
| Load trained GenRe Transformer. | ||
|
|
||
| Args: | ||
| model_path: Path to model checkpoint (relative to genre/ directory) | ||
|
|
||
| Returns: | ||
| Loaded GenRe Transformer model | ||
| """ | ||
| # Construct absolute path ⭐ | ||
| # model_path is relative to methods/catalog/genre/ | ||
| base_path = Path(__file__).parent | ||
|
|
||
| # If path doesn't start with "saved_models/", prepend it | ||
| if not model_path.startswith("saved_models/"): | ||
| model_path = f"saved_models/{model_path}" |
There was a problem hiding this comment.
The new GenRe recourse method always loads a single checkpoint from saved_models/genre_transformer.pth (lines 126‑151) and never uses the data_name that run_experiment injects into the hyperparameters. As a result, any CLI invocation that runs --recourse_method genre on a dataset whose feature dimensionality differs from the COMPAS checkpoint will crash when self._transformer.generate receives inputs of the wrong size (the linear embeddings were constructed with checkpoint['n_features']). The method should either select a dataset-specific model based on data_name or raise a clear error before loading; otherwise the experiments fail for every dataset except the one the checkpoint was trained on.
Useful? React with 👍 / 👎.
2 participants
No description provided.