detection training

[lfxuan]

The data is synthesized and labeled in the same way, and the wrong image is not damaged after query. Why do the following errors appear in the training process

Traceback (most recent call last):
  File "references/detection/train_tensorflow.py", line 261, in <module>
    main(args)
  File "references/detection/train_tensorflow.py", line 200, in main
    fit_one_epoch(model, train_loader, batch_transforms, optimizer, mb)
  File "references/detection/train_tensorflow.py", line 42, in fit_one_epoch
    train_loss = model(images, targets, training=True)['loss']
  File "***/.pyenv/versions/doctr/lib/python3.7/site-packages/keras/engine/base_layer.py", line 1037, in __call__
    outputs = call_fn(inputs, *args, **kwargs)
  File "***/doctr/doctr/models/detection/differentiable_binarization/tensorflow.py", line 245, in call
    loss = self.compute_loss(logits, thresh_map, target)
  File "***/doctr/doctr/models/detection/differentiable_binarization/tensorflow.py", line 183, in compute_loss
    seg_target, seg_mask, thresh_target, thresh_mask = self.compute_target(target, out_map.shape[:3])
  File "***/doctr/doctr/models/detection/differentiable_binarization/base.py", line 326, in compute_target
    thresh_mask[idx])
  File "***/doctr/doctr/models/detection/differentiable_binarization/base.py", line 251, in draw_thresh_map
    canvas[ymin_valid:ymax_valid + 1, xmin_valid:xmax_valid + 1]
ValueError: could not broadcast input array from shape (3,0) into shape (3,1)

[fg-mindee]

Hi @lfxuan 👋

In order to help you, would you mind providing more details about how you got this error please? 🙏
What do you mean by "the wrong image is not damaged after query" ? What arguments did you pass to the script?

Best

[lfxuan]

Hello! This is the image I made a mistake during training. I guess the coordinates of the image detection box appear negative after processing. What I do now is to change my image margin and try to avoid this situation to solve this problem.errors.zip

[fg-mindee]

Hi @lfxuan,

I think you forgot to upload the image 😅
I'm not sure I understand: did you manage to solve your problem? If yes, could you share how please ? 🙏 If not, can you give more specifics please?

[fg-mindee]

Any update @lfxuan ? 🙏

[lfxuan]

@fg-mindee I'm very sorry to see the message now! Now the image and label files have been uploaded again
error.zip
.

[fg-mindee]

It looks like a text recognition image. While your error seems to point out you were trying to train your model for text detection 🤔

Taking a look at your labels, it looks like you're trying to text a character-level localization model. The models in docTR are designed for word-level localization 😅

So a few options:

• if you are looking for text detection, character-level localization is fully experimental. We haven't tried it so far
• if you are looking for word-level localization, your image is already it. Considering a full page :)