Download Trained Model Error using Pretrained Model


#1

Superb program! As a novice in only a few weeks operating on a stand alone windows 10 pc I was able to use transfer learning on an image categorization model and achieve >99% validation accuracy. However I was surprised that the Inference only showed about 2% correct identifications. Yes, I did retrained the last 6% of the VGG19 model. So I believe that this error has been discussed in this forum, not that I understood the answer given. However, I wanted to mention that when I downloaded the trained model and ran the python template program on my data the same low value of correct identifications was observed. I do not know if this issue was known or not. On closer observation this downloaded program makes the identical errors as the inference program. Is the real issue here that the coefficients that were retrained in the VGG19 model were not used in the downloaded coefficients, substituting in their stead the original VGG19 pretrained coefficients?

PS The inference program works correctly when 0% of the VGG19 model is trainable.

If this is a known issue can you guess as to when a version of the program will be available with a fix? Thank you.


#2

It is a known issue with MxNet backend. The issue seems to be what you said. The trained weights are not getting saved in the scenario when trainable is set.

We plan to investigate/fix this issue in a month timeframe. We will update this thread when a fix is released.


#3

I have just switch to version 2.5.0. The problem with retraining part of VGG19 still seems to persist. Is that correct?