Introducing: The Pytorch Profiling hook! Saves a trace, that can be interactively viewed via the chrome://tracing tool. Take a look at the attached pictures.
Still on the todo list:
[ ] Add warning that traces get huge fast! Solution: `--num_steps 20``
[ ] Add example interpretation (?)
Image 1: Some 15 training steps. Note the validation steps every 2^n steps.
Image 2: Zoom to about one step. Left: Forward pass, middle: gradient calculation, right: gradient application.
Image 3: Some Conv - Batch-Norm - ReLU layers
This needs to be documented, otherwise it is just another awesome feature which nobody will ever know about and we will kick it out again after some time.
Introducing: The Pytorch Profiling hook! Saves a trace, that can be interactively viewed via the
chrome://tracing
tool. Take a look at the attached pictures.Still on the todo list:
Image 1: Some 15 training steps. Note the validation steps every 2^n steps. Image 2: Zoom to about one step. Left: Forward pass, middle: gradient calculation, right: gradient application. Image 3: Some Conv - Batch-Norm - ReLU layers