Changelog¶

v1.2.1180 (Jan 07, 2022)¶

Add limit parameter to BCN.evaluate. BCN.run_wp limits the training set evaluation to 10,000 samples instead of 60,000, in the interest of training times.
Add webhook parameter to BCN.run_wp.
Add Results.wp_layers to keep track of which layers were perturbed.
Add padding, width, and depth kwargs to the Fault class.
Add fault argument to BCN.run_epoch and BCN.run_epochs.
Add activation argument to BCN and BCNLayer.
Add BCN.confusion for computing interclass confusion matrices.

Change Results class
- When loading from a file, update Results.__dict__ instead of entirely replacing it, in the interest of backward- and forward-compatibility.
- Remove Results.best_epoch.
- Remove Results.best_valid_loss.
- Remove Results.version.
- Add best, representing the index of the maximum F-1 score (instead of minimum validation loss).
- Add versions, a set containing all the versions that the model was trained under.
- Add devices, a set containing all the devices the model was trained on, to help put training times into perspective.
- Add step, the number of weight perturbation steps
Change BCN network matrices
- Harness the power of 4D batch tensors and vectorization for (marginally?) improved training times!
- Encode the used device in the local filename to prevent device-related errors.

Fix local networks filenames weren’t specific (e.g. it now writes uniform.NextToNN instead of just NextToNN).

See the migration page for more detailed info.

Add this changelog.
Add bcn.BCN.default_weights_filename and bcn.BCN.default_results_filename.
Add a from_path parameter to bcn.BCN.train that allows the user to specify a results directory path, from which the model should look for and load its weights and results, instead of explicitly specifying the two file paths via from_weights and from_results.