Repository navigation
Add the AdaNorm optimizer - #552
Open
shaneraphel wants to merge 6 commits into
Open
shaneraphel wants to merge 6 commits into
shaneraphel wants to merge 6 commits into
Conversation
get_trace averaged the Hessian diagonal over the spatial dims only for 4D kernels. Conv1d (3D) and Conv3d (5D) weights left tmp_output unbound. Average over dims 2..ndim-1 instead; 4D is unchanged.
Each mode factor of an order-k tensor enters to the power -1/(2k): -1/4 per side for matrices. The code used -1/k, twice the published exponent. The 7 pre-existing test_optimizer Shampoo failures are a state-dict precision issue on this torch version and fail identically without this change.
warmup_rate (default 1e-6) replaces the hard-coded warm-up slope and min_step_size (default 1e-2) the floor. Old checkpoints without the new group keys fall back to the old values. The 7 pre-existing test_optimizer Adafactor failures are a state-dict precision issue on this torch version and fail identically without this change.
Second-order optimizer with a diagonal Hessian estimate (SophiaG). Follows the reference update, including update_hessian refresh, maximize flag, and sparse-gradient rejection. CUDA-graph capturable execution is not ported. Wired into the test harness lists and README.
Layerwise-gradient-statistics optimizer with K mutual projections of the gradient and its running average. Modernizes the reference in-place add_ calls, keeps eps out of the state, and holds K/xi in the group defaults. Wired into the test harness lists and README.
Gradient-norm-corrected Adam with AMSGrad, decoupled/fixed decay, and an RAdam-style rectified update. Modernizes the reference in-place ops, keeps eps out of the state, guards the zero-norm correction, and holds all flags in the group defaults. Wired into the harness lists + README.
This branch has not been deployed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Adds
AdaNorm, which rescales each gradient by its running norm average before the Adam update, from https://arxiv.org/abs/2210.06364.The port follows https://github.com/shivram1987/AdaNorm with four fixes:
add_(scalar, tensor)calls use the pre-1.5 torch signature and fail on current torch; the port usesadd_(tensor, alpha=...)/addcmul_epsintoexp_avg_varin place (twice per step), soepsaccumulates into the state; the port computes the denom out of placegrad * avg / 0; the correction is now skipped when the current norm is zeroweight_decouple/fixed_decay/rectifymove from print-guarded instance attributes into the group defaults (with__setstate__fallbacks), so checkpoints round-tripFixes #499.
Test plan
pytest tests/test_adanorm.py: 5 passed — one step matches a hand-computed reference, zero gradient stays finite, rectify takes the SGD path early and the Adam path later, both decoupled variants move,reset()clearspytest tests/test_optimizer_with_nn.py -k AdaNorm: 1 passedpytest tests/test_param_validation.py -k AdaNorm: 2 passedtests/test_optimizer.py -k AdaNorm: 7 failures in the state-dict round-trip assertion, the same pre-existing torch-2.11 harness issue documented in Add the Sophia optimizer #550