Outputting evaluation loss

[amitlisha]

I want to track the training using the validation set loss, but by default it only outputs the accuracy, is there any way to also output the validation loss?

Thanks!

[GreatV]

By default, PaddleOCR primarily outputs validation accuracy, but you can also track the validation loss during training.

Steps to Output Validation Loss:

1. Modify the Training Script:

   • If you are using PaddleOCR’s built-in training, check the code that handles logging and ensure that validation loss is included in the printed/output metrics.

2. Use Logging or Visualization:

   • The document discussions mention that VisualDL is deprecated in PaddleOCR (ppocr WARNING: You are using VisualDL, the VisualDL is deprecated and removed in ppocr!). This means you cannot rely on VisualDL for tracking loss.
   • As an alternative, you can manually log the validation loss using Python’s built-in logging packages (logging) or save it into a file for later visualization.

3. Check Configuration (.yml file):

   • If applicable, modify the .yml config file to ensure validation loss tracking is enabled (check how metrics are computed in the evaluations settings).

4. Custom Callback for Loss Monitoring:

   • If PaddleOCR allows custom logging hooks or callbacks, add a function to print and save validation loss after each evaluation step.

Alternative:

If you need to visualize the loss curve, you can manually extract loss values from logs and plot them using matplotlib or tensorboard.

For further insights, check this discussion:
about visualDL or visualize loss curve

Let me know if you need more details! 🚀

---

Response generated by 🤖 feifei-bot | chatgpt-4o-latest

⚠️ This is an AI-generated response. Please verify any critical information.

[amitlisha]

after digging into the code, it seems that the eval function doesn't compute the loss at all, and it requires a significant change to be able to. Tracking the validation loss is kinda standard, is there something I'm missing?