Inference giving 45 characters output.

[AnishTadepalli]

Hi, I am training the model on captcha images, the model is getting upto 99.4% accuracy, and I can see the ground truth and prediction characters are same in length. But when I am using inference, I am getting 45 characters list as output. I have tried lot of things but still same type of output. Predictions: ['g', '8', '8', '8', '8', '8', '8', '8', 'n', 'f', 'w', 'w', '7', 'e', 'e', '8', 'w', 'x', 'w', 'x', 'w', 'b', 'f', 'f', 'f', 'w', '3', 'p', 'p', 'p', '3', '3', '3', '3', '3', '3', '3', '3', '3', '3', 'p', 'p', 'p', 'p', '7'] and the characters in the image are: 2b827 This is my config: defaults:

• override hydra/job_logging: custom

processing: device: cpu image_width: 180 image_height: 50

training: lr: 3e-4 batch_size: 8 num_workers: 4 num_epochs: 100

bools: DISPLAY_ONLY_WRONG_PREDICTIONS: true VIEW_INFERENCE_WHILE_TRAINING: true SAVE_CHECKPOINTS: true

paths: dataset_dir: ./dataset save_model_as: ./logs/crnn.pth

model: use_attention: true use_ctc: true gray_scale: true dims: 256

[GabrielDornelles]

Can you provide me your inference.py file?

[AnishTadepalli]

Yes, please fine the below attachment.

inference.txt

[GabrielDornelles]

Please rewrite your classes list to the same that was trained, its printed in the terminal as the training starts. The first element in the list MUST be "∅".

I see you also changed a few things in the image processing (mean and std, grayscale convertion), so you need to ensure that you have done this at training code too.

[AnishTadepalli]

Hi, as you mentioned I have modified the training code (mean,std blank char), but still getting the same type of output. I am adding the modified files. train.txt data_loading.txt inference.txt

[AnishTadepalli]

Hi, thanks for your time and responses. Actually at the end of the training I am getting the error as " load state dict Expected state_dict to be dict-like, got <class 'NoneType'>.",because I am giving the training.epochs=25,5. So if the best accuracy is achieved at epoch 18 in experiment 2, without saving the model I am continuing the training loop so the model weights are not loading into the model. I did small modification in train.py. now I am getting proper predictions. train.txt