SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2411.02853 · NeurIPS · 2024

ADOPT: Modified Adam Can Converge with Any β 2 with the Optimal Rate

Yutaka Matsuo, Shohei Taniguchi, Yusuke Iwasawa, Masahiro Suzuki, Yuta Oshima, Keno Harada, Gouki Minegishi, Seong Jeong, Go Nagahara, Tomoshi Iiyama

arXiv · PDF · Open in the Atlas

Code that ran

We lifted 23 functions out of this paper's own repositories and ran 11 of them in a sandbox. "Ran" means the function executed on a synthesized input and returned a value. It is not a reproduction of the paper's results.

RepositoryRoleRan
UiPath/torchvision canonical 9 of 18
iShohei220/adopt canonical 2 of 5
FunctionStatusWhere it lives
accuracy Ran iShohei220/adopt/imagenet/utils.py
code served (permissive licence) · get_code("0d95e28fac98427f")
channel_shuffle Ran UiPath/torchvision/torchvision/models/shufflenetv2.py
code served (permissive licence) · get_code("cf7081fc34608ea3")
conv1x1 Ran UiPath/torchvision/torchvision/models/resnet.py
code served (permissive licence) · get_code("2a80220dabcb742a")
conv3x3 Ran UiPath/torchvision/torchvision/models/resnet.py
code served (permissive licence) · get_code("600ff2c45e0de056")
densenet121 Ran UiPath/torchvision/torchvision/models/densenet.py
code served (permissive licence) · get_code("9eb4b3b03ed7a1c9")
densenet169 Ran UiPath/torchvision/torchvision/models/densenet.py
code served (permissive licence) · get_code("858b62a0718acf3c")
googlenet Ran UiPath/torchvision/torchvision/models/googlenet.py
code served (permissive licence) · get_code("cc86d582f7c75a6d")
mnasnet0_75 Ran UiPath/torchvision/torchvision/models/mnasnet.py
code served (permissive licence) · get_code("561d4224329b8802")
mnasnet1_0 Ran UiPath/torchvision/torchvision/models/mnasnet.py
code served (permissive licence) · get_code("541eb8cf9b978129")
mobilenet_v2 Ran UiPath/torchvision/torchvision/models/mobilenetv2.py
code served (permissive licence) · get_code("57f30bbef481b9c8")
store_model_weights Ran iShohei220/adopt/imagenet/utils.py
code served (permissive licence) · get_code("d92101f375faa8a6")
alexnet Not yet run UiPath/torchvision/torchvision/models/alexnet.py
code served (permissive licence) · get_code("8ccdf1f64321d691")
average_checkpoints Not yet run iShohei220/adopt/imagenet/utils.py
code served (permissive licence) · get_code("4eca7815147e1708")
densenet161 Not yet run UiPath/torchvision/torchvision/models/densenet.py
code served (permissive licence) · get_code("e945c7af558f65a9")
get_args_parser Not yet run iShohei220/adopt/imagenet/train_quantization.py
code served (permissive licence) · get_code("e36bcfb158ae8f9c")
get_module Not yet run iShohei220/adopt/imagenet/presets.py
code served (permissive licence) · get_code("b2ea55703711dd12")
inception_v3 Not yet run UiPath/torchvision/torchvision/models/inception.py
code served (permissive licence) · get_code("d5f1aa4870d7ad8b")
load_meta_file Not yet run UiPath/torchvision/torchvision/datasets/imagenet.py
code served (permissive licence) · get_code("fd93b84b5a89ac58")
mnasnet0_5 Not yet run UiPath/torchvision/torchvision/models/mnasnet.py
code served (permissive licence) · get_code("095cb1dfd64f93c1")
parse_devkit_archive Not yet run UiPath/torchvision/torchvision/datasets/imagenet.py
code served (permissive licence) · get_code("d523a6fcfaf250c7")
resnet18 Not yet run UiPath/torchvision/torchvision/models/resnet.py
code served (permissive licence) · get_code("6af00ec71cf439d8")
shufflenet_v2_x0_5 Not yet run UiPath/torchvision/torchvision/models/shufflenetv2.py
code served (permissive licence) · get_code("3edb47a32cd6a773")
shufflenet_v2_x1_0 Not yet run UiPath/torchvision/torchvision/models/shufflenetv2.py
code served (permissive licence) · get_code("97931550315207a2")

Repositories linked to this paper

Some links come from the archived Papers with Code dataset (CC BY-SA 4.0): attribution and licence.

Abstract

Adam is one of the most popular optimization algorithms in deep learning. However, it is known that Adam does not converge in theory unless choosing a hyperparameter, i.e., β 2 , in a problem-dependent manner. There have been many attempts to fix the non-convergence (e.g., AMSGrad), but they require an impractical assumption that the gradient noise is uniformly bounded. In this paper, we propose a new adaptive gradient method named ADOPT, which achieves the optimal convergence rate of O(1/ √ T ) with any choice of β 2 without depending on the bounded noise assumption. ADOPT addresses the non-convergence issue of Adam by removing the current gradient from the second moment estimate and changing the order of the momentum update and the normalization by the second moment estimate. We also conduct intensive numerical experiments, and verify that our ADOPT achieves superior results compared to Adam and its variants across a wide range of tasks, including image classification, generative modeling, natural language processing, and deep reinforcement learning. The implementation is available at https://github.com/iShohei220/adopt.

For agents

The same record, over MCP at https://syntology.ai/mcp:

get_harvested_code_for_paper("2411.02853")
get_code_for_paper("2411.02853")
have("2411.02853")

Connect an agent — have() is free.