Submission to ICML 2022

What's Changed

• Simple by @ernoult in #2
• Fixes PL Implementation, reproduces CIFAR10 performance by @amoudgl in #4
• Add Vanilla DTP and Target Propagation models by @lebrice in #5
• Use one Optimizer per feedback layer by @lebrice in #6
• Add Callback that compares DTP and Backprop grads by @lebrice in #7
• Add code / scripts for running HPO Sweeps by @lebrice in #9
• Fix for #12 by @lebrice in #13
• Add ImageNet dataset and models by @amoudgl in #14
• Fix unit tests, minor refactor of network classes by @lebrice in #16
• Improve optimizer config and logging by @amoudgl in #17
• Add 3x std normalization for CIFAR by @amoudgl in #25
• Fix #24, remove track_grad_norm flag by @lebrice in #29
• Merge LeNet by @scspinney in #26
• Add top5 metric by @amoudgl in #30
• Fix a small bug with the logged net_type by @lebrice in #32
• Bug fix in layer distances and angles logging by @amoudgl in #35
• Add back the logging of weight distances and angle by @lebrice in #37
• Standardize CIFAR10 normalization by @amoudgl in #36
• Fix angle computation by @amoudgl in #40
• Add theorems tests by @amoudgl in #41
• Update backprop hparams with tuned values by @amoudgl in #42
• Typing improvements + Add Meulemans to figures by @lebrice in #43

Full Changelog: https://github.com/ernoult/scalingDTP/commits/ICML2022