Not strictly related, but I've been wanting a 3D debugger for ML models for damn near a year now. It seems like a slam-dunk idea. Stack images like https://bat

Not strictly related, but I've been wanting a 3D debugger for ML models for damn near a year now. It seems like a slam-dunk idea. Stack images like https://battle.shawwn.com/sdb/visualizations/2020-09-16-117m-gpt-revnet/index.html https://battle.shawwn.com/sdb/visualizations/2020-09-16-117m... in 3D space, like a cube. Put them in order of layers. Ditto for activations. The spatial structure is important. Right now there's nothing, not even tensorboard, that preserves spatial structure. For example, look at these two images: https://battle.shawwn.com/sdb/visualizations/2020-09-16-117m-gpt-revnet/model_g_h6_attn_c_attn_w.png https://battle.shawwn.com/sdb/visualizations/2020-09-16-117m... https://battle.shawwn.com/sdb/visualizations/2020-09-16-117m-gpt-revnet/model_g_h6_attn_c_attn_b.png https://battle.shawwn.com/sdb/visualizations/2020-09-16-117m... You're looking at weights, followed by bias. You can see visually that the bias lines up with the weights. If you tried visualizing that with tensorboard, it'd shrink the images to squares with padding. Which would be fine if you could rearrange squares so that one is above the other, but you can't. Meaning you lose the ability to correlate visually what's going on. Someone please make an ML doom blaster so that I can at least reset layers of my ML model by shooting them.