120 points by dl_wizard 6 months ago flag hide 15 comments
user1 6 months ago next
Great article! Visualizing neural networks in real-time is an amazing way to understand how they're learning.
author 6 months ago next
Thank you! I had a lot of fun building this tool and I'm glad you found it informative.
user2 6 months ago prev next
I'm having trouble getting this to work on my GPU. Any suggestions?
helpful_user 6 months ago next
Make sure you've got the latest version of TensorFlow and your GPU drivers. You might also try dropping down to a lower TensorFlow version if your GPU isn't super powerful.
user3 6 months ago prev next
This is such an awesome tool. Great job!
author 6 months ago next
Thanks user3, I'm happy to hear you found it useful!
user4 6 months ago prev next
How did you go about visualizing the real-time aspect? I'm curious about the implementation.
author 6 months ago next
I used TensorBoard's event file format and recorded the training data to a file at every step. Then I read from the file and displayed the graphs using D3.js. I'm happy to go into more detail if you're interested.
user5 6 months ago prev next
Do you have any tips for working with large neural networks?
experienced_user 6 months ago next
Try using gradient checkpointing to reduce the amount of memory required. It's a technique where you save only a subset of intermediate results while training, rather than the whole computational graph. Also, consider using sparse representations for weights and connections where possible.