Replies: 2 comments 3 replies
-
|
Hi @lfxuan, |
Beta Was this translation helpful? Give feedback.
-
|
Hello! @charlesmindee I tried every model, in which the loss of master model hovered around 3, the loss of SAR resnet31 model hovered around 4, and the loss of crnn model hovered around 17. |
Beta Was this translation helpful? Give feedback.
-
|
Hello again @lfxuan, are you using the tensorflow backend or the torch backend ? |
Beta Was this translation helpful? Give feedback.
-
|
@charlesmindee I use tensorflow backend and vacob to reset to the required characters, and add training data for training. The following training settings and loss output: Namespace(arch='master', batch_size=80, epochs=30, input_size=40, lr=0.001, name=None, pretrained=False, resume=None, show_samples=False, test_only=False, train_path='/doctr/train_data/rec_data/train_set_million', val_path='/doctr/train_data/rec_data/valid_set_million', vocab='la', wb=False, workers=None) Validation loss decreased inf --> 5.04232: saving state... |
Beta Was this translation helpful? Give feedback.
-
|
This is mainly because the training data set is too complex. Using a simple data set to train a pre training model and lower learning rate training is helpful to train the model. |
Beta Was this translation helpful? Give feedback.
-
When the pre training model is not loaded, ten thousand 4-character verification code images are synthesized to train the master model. The loss remains 3 unchanged. When the pre training model is loaded for training, it can be reduced to 0.1. How to solve the problem of loss non convergence without loading the pre training model?
Beta Was this translation helpful? Give feedback.
All reactions