search

sushant097 / Handwritten-Line-Text-Recognition-using-Deep-Learning-with-Tensorflow

Use Convolutional Recurrent Neural Network to recognize the Handwritten line text image without pre segmentation into words or characters. Use CTC loss Function to train.
Apache License 2.0
292 stars 122 forks source link

Pretrained model in CNN layer #9

Closed praveenraj2018 closed 4 years ago

praveenraj2018 commented 4 years ago

Hi Sushant Great work. Appreciate.

I have a few questions.

1) Have you ever tried adding a pre-trained model such as VGG, Inception or ResNet at the CNN layer? If so, how was results? What you suggest on adding one?

2) I have read about MD-LSTM and there are results that gives better improvement when used over BiLSTMs. Any experience or experiments on this work?

3) By the way, have you ever attempted to run the code under GPU or TPU?

thanks Raj

sushant097 commented 4 years ago

I haven't use the pre-trained model yet. But hope that it may or not provide a better result. MD-LSTM can be great as it directly recognizes the whole paragraph without pre-segmentation into lines here but training is difficult and needs lots of tricks on it. And I have run the code on GPU on the Cloud. You can use a small GPU with a small batch of sizes for training.

praveenraj2018 commented 4 years ago

Thanks Sushant. It was useful.