Adding temporary diagnostic metrics

Sometimes you want to quickly log a value from your loss function for debugging purposes without setting up a full MetricWrapper. You can do this by adding the value to the dictionary returned by your loss function and configuring log_strs in Harness.

1. Update your loss function

In your loss_fn (e.g., in FlexMDMLoss.loss_fn), add the desired metric to the returned dictionary. Make sure to .detach() it to avoid memory leaks.

```python:flexmdm_correction/loss_flexmdm_correction.py def loss_fn(self, batch, ...): # ... your loss computation ...

    loss = ...
    my_debug_metric = ...

    return {
        "loss": loss,
        "unmask_loss": unmask_loss.detach(),
        "insertion_loss": insertion_loss.detach(),
        "my_metric": my_debug_metric.detach(),  # Add your new metric here
    }

```

2. Configure log_strs

In your experiment configuration (e.g., experiment/star_easy_flexmdm.yaml), add an entry to lightning_module.log (which maps to log_strs in Harness). The keys are the names you want to see in the logger (e.g., Weights & Biases or TensorBoard), and the values are the keys in the dictionary returned by your loss function. ```yaml

experiment/star_easy_flexmdm_correction.yaml

lightning_module: log: "train/loss": "loss" "train/unmask_loss": "unmask_loss" "train/insertion_loss": "insertion_loss" "train/my_custom_name": "my_metric" # Map "my_metric" from loss_dict to "train/my_custom_name" ```

How it works in Harness

The Harness class (in lib/xlm-core/src/xlm/harness.py) iterates over self.log_strs during the training step and logs each entry:

python:821:832:lib/xlm-core/src/xlm/harness.py if stage == "train": for k, v in self.log_strs.items(): self.log( k, loss_dict[v].detach(), on_step=True, on_epoch=False, prog_bar=True, sync_dist=False, rank_zero_only=True, logger=True, add_dataloader_idx=False, )

By default, prog_bar=True is set, so these metrics will also appear in your progress bar during training.