I am following your notebook convert_to_npz to create the training, validation, and test datasets, and my target arrays y_train, y_valid, y_test have shapes of the form (number, ) with a blank in the second slot of the shape, which is how the shapes appear in your notebook convert_to_npz. However, in your CRNN_model notebook, the imported datasets' y-arrays have shapes (n, 8), like they are one-hot encoded, so I made this guess and trained the model using the same code.
My issue is that in doing this the CRNN model has terrible results (around 13% accuracy or so on both train and validation). When using the data downloaded from the google drive, the issue goes away and I can recreate your results (~50% accuracy on validation). Can you explain further how you made your data? I must be doing something wrong, but I cannot find where our methods have differed.
I am following your notebook
convert_to_npzto create the training, validation, and test datasets, and my target arraysy_train,y_valid,y_testhave shapes of the form(number, )with a blank in the second slot of the shape, which is how the shapes appear in your notebookconvert_to_npz. However, in yourCRNN_modelnotebook, the imported datasets' y-arrays have shapes(n, 8), like they are one-hot encoded, so I made this guess and trained the model using the same code.My issue is that in doing this the CRNN model has terrible results (around 13% accuracy or so on both train and validation). When using the data downloaded from the google drive, the issue goes away and I can recreate your results (~50% accuracy on validation). Can you explain further how you made your data? I must be doing something wrong, but I cannot find where our methods have differed.