PyTorch introduces "better" feature extraction via symbolic tracing


Thought you all might appreciate this. TorchVision now provides a new method to extract features generated during the forward pass of a neural network. This new technique promises to allow feature extraction without modifying the source code of your given model and by simply tracing the graph of its operations, including intraline sub-operations.

Try it out and see what you think!


Same here. I also use PyCharms on my laptop with OSX, and connect a remote interpreter on my desktop running ubuntu with the 1080ti. I use SSH and the remote deploy feature of PyCharms, this allows me to work remotely and have the desktop workstation headless (no screen). Code lives and is versioned on OSX and synced on ubuntu by pycharms, I start jupyter by SSHing into ubuntu, and I open the web interface of jupyter by accessing the local IP. It’s a bit of setup, but so far the experience has been really great.