soeque1 edited a comment on issue #13647: Update lip reading example URL: https://github.com/apache/incubator-mxnet/pull/13647#issuecomment-462132474 @aaronmarkham I checked the file you uploaded. It was nice as intended. @thomelane @seujung and I are working together for this example. (1) After we did some experiments, we decided to use default learning rate. (2) We checked the prediction, decoded using beam search. It takes too long time to train this model. The main reason is the decode part (beam search) on validation data (def infer_batch). Actually, we do not need decode all the validation examples during training, so we skip this or check only one mini batch example. To speed up, we only check the decoded result using infer.py not main.py (train). But it helps to understand how good the result is. (3) Although the loss is still decreasing, I attach the pre-trained model. [Model](https://github.com/soeque1/temp_files/files/2848870/epoches_81_loss_15.7157.zip) [Sample](https://user-images.githubusercontent.com/11376047/52533982-d7227680-2d7e-11e9-9f18-c15b952faf0e.png) You can get the result. ``` python infer.py model_path='checkpoint/epoches_81_loss_15.7157' ``` Or You can resume training. ``` python main.py model_path='checkpoint/epoches_81_loss_15.7157' ```
---------------------------------------------------------------- This is an automated message from the Apache Git Service. To respond to the message, please log on GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: [email protected] With regards, Apache Git Services
