search

githubharald / SimpleHTR

Handwritten Text Recognition (HTR) system implemented with TensorFlow.
https://towardsdatascience.com/2326a3487cd5
MIT License
1.99k stars 894 forks source link

not able to train on different data set using the pretrained model #132

Closed ileniTudor closed 2 years ago

ileniTudor commented 2 years ago

Thank you for sharing the work with us. I'm trying to generate a new dataset with image size 128x32, but after loading the model 'checkpoints-33', and run self.trainBatch, I have dimensionality issue.

`tensorflow.python.framework.errors_impl.InvalidArgumentError: Conv2DCustomBackpropFilter: filter and out_backprop must have the same out_depth [[node gradients/convolution_grad/Conv2DBackpropFilter (defined at D:\work\ocr-prescription\python\ocrPrescription\SimpleHTR\src\model.py:51) ]]

Errors may have originated from an input operation. Input Source operations connected to node gradients/convolution_grad/Conv2DBackpropFilter:`

It is about the image dimension size? What is the train dimension size?

githubharald commented 2 years ago

Hi, image shape is 128x32 for training in single-word mode, and 256x32 in multi-words mode. But any other shape with width 4*k (with k some int) and height 32 should also work. So I guess something went wrong, e.g. while loading the images. I suggest to check what the shape of the tensor fed to the NN during training looks like.