CNN network training accuracy reaches a high point and then drops to zero.

User 2920 | 5/6/2016, 5:17:35 AM


Recently I have been experiencing a strange behavior in graph lab create. I have 121 Classes ( All scanned documents ). On the same data I was able to train a classifier on 128x128 image size which gave me good accuracy. Now I resized my data to 224x224 to see if it will perform better. The classifier training accuracy reaches some high point in the initial iterations and then drops to zero. I have run the experiment multiple times. It shows the same behavior when i increase the number of classes to 122 on 128x128 image size as well. I completely fail to grasp a reason for this. Please have a look at the following snippets.

This is the model that I am using

Please let me know if there is something wrong with my parameters or is it some bug in the library. Thanks


User 5159 | 5/9/2016, 2:53:32 AM

The learning rate is too large. You need to reduce learning rate during training.