search

myinxd / MCRGNet

Morphological Classification of Radio Galaxy Network
MIT License
4 stars 1 forks source link

cae-pretrain-demo doesn't work #1

Open nottu opened 5 years ago

nottu commented 5 years ago

It seems the demo is older than the latest revision of ConvAE and hasn't been updated to take the changes into account

myinxd commented 5 years ago

Hey. I have run the notebook that you sent to me. Seems that your code works on my data. I guess there might be some problems happened during your data preprocessing procedure. Would you like to share me your data preprocessing code? Besides, I will also check and modify the ConvAE code if possible.

Sorry again for the inconvenience. :)

nottu commented 5 years ago

I reverted my branch to commit "41a68b6" and now the notebooks work (git checkout -b old-state 41a68b6) there must be something that broke the notebooks. I can take a look if you want and make a pull request once I fix the issue.

nottu commented 5 years ago

I just removed the "tf.nn.dropout" layers and all is fine now. I'm not sure what the actual problem is, since dropout should help...

myinxd commented 5 years ago

Thanks for your reply. I have also run into this problem before. I tried to comment out the dropout lines then the network worked. In my view, the dropout could lead to a divided-by-zero error, which causes the nan to the network parameters. Maybe the TensorBoard could be used to visualize the training process to check the nan issue...