Hi,
I’m tried to build model(transfer learning) with ipython notebook.
Actually, I’m not the expert of the DL ground.
I’m the embedded engineer who want to test models on the board.
I understand that Fine-tuning in transfer learning usually fixes the front layer of the pre-trained model and updates the parameters of the back layer from scratch.
However, the code was different from what I knew.
The Fine-tuning epoch was only 10.
I don’t know how the training cycle and learning rate I chose in visual simple mode affect.
Please explain.
@dansitu can you answer here?
In our transfer learning scheme we run two training cycles. The main one trains the final layer of the model from scratch with a high learning rate. The “fine tuning” cycle then uses a smaller learning rate to update a subset of the earlier layers. All of this is configurable in the script via FINE_TUNE_EPOCHS and FINE_TUNE_PERCENTAGE.
Note that right now transfer learning doesn’t work in an iPython Notebook since we don’t provide the model weights. I’m working on a fix for this currently and it should be working by the end of the week!
2 Likes
Thank you for your reply @dansitu!! 
2 Likes