torchexplorer
torchexplorer copied to clipboard
Interactively inspect module inputs, outputs, parameters, and gradients.
Right now, all i/o gradients are logged using the grad norm, and parameter gradients are logged using the raw gradient. It would be nice to make this configurable.
This is pretty tricky to implement so probably not a priority for now. In the scenario where the root module's `forward` method is invoked multiple times before the loss `backward`,...
An edge into a module should have just one small number above it with a rectangular background showing which i/o index it is.
This would be a big new feature. For each element in the batch the user should supply a tooltip text and point color. There should then be a 2d view...
Right now, in order to ensure that all submodules have input gradients / output gradients, a dummy tensor is added to the inputs of the `watch`'d module. This adds new...
Makes it easier to quickly compare a particular set of nodes across different runs. Also should do something similar with the standalone.
I am following the tutorial from Andrej K. building gpt2 from scratch. I thought it would be a good idea to visualize his GPT2 model using torchexplorer. This what I...