Conversation
hyperparameters
![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.
Reviewed commit: b49f7a9473
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".
code/LinearTrain.py
Outdated
| from Functions import LinearConfig, plot_loss, run_batch, test_model, graph_roc_curve | ||
|
|
||
| def linear_train(config, model_pth): | ||
| config = LinearConfig() |
There was a problem hiding this comment.
linear_train
The new sweep in code/Main.py:59-65 mutates config.d_model and config.learning_rate before each call, but linear_train() immediately replaces that object with a fresh LinearConfig. Every non-default run therefore trains the default 32-hidden-unit / 0.1-LR model while the console output and linear_history.txt record it as a different hyperparameter combination, which makes the linear search results unusable.
Useful? React with 👍 / 👎.
code/LogprobTrain.py
Outdated
|
|
||
| def logprob_train(config, model_pth): | ||
|
|
||
| config = LogitConfig() |
There was a problem hiding this comment.
logprob_train
code/Main.py:74-87 iterates over add_conv, d_model, conv_ch, and learning_rate, but logprob_train() discards those values by constructing a new default LogitConfig here. As written, the 60-run search will keep retraining the same no-CNN, 32-dimension, 0.1-LR model instead of the requested variants, so the comparison this commit adds is invalid for every non-default logprob run.
Useful? React with 👍 / 👎.
1 participant
hyperparameters