Adan: Adaptive Nesterov Momentum Algorithm for Faster Optimizing Deep Models

This is an official PyTorch implementation of Adan. See the paper here. If you find our adan helpful or heuristic to your projects, please cite this paper and also star this repository. Thanks!

@article{xie2022adan,
  title={Adan: Adaptive Nesterov Momentum Algorithm for Faster Optimizing Deep Models},
  author={Xie, Xingyu and Zhou, Pan and Li, Huan and Lin, Zhouchen and Yan, Shuicheng},
  journal={arXiv preprint arXiv:2208.06677},
  year={2022}
}

News

• 🔥 🔥 🔥FusedAdan with less memory footprint is released.
• Adan is supported in the lasted version of Timm.
• Results on large language models, like GPT2, are released.
• Adan is chosen as the default optimizer in the text-to-3D DreamFusion Project. See more results here.
• TF's implementation (third party) refers to DenisVorotyntsev/Adan.
• JAX's version (third party) is implemented and also supported in Deepmind/optax.
• Adan is supported in the MMClassification of the OpenMMLab project. The user can find the log and example of using Adan to train ViT-B here. The results of the detection tasks are coming soon.

Installation

python3 -m pip install git+https://github.com/sail-sg/Adan.git

FusedAdan is installed by default. If you want to use the original Adan, please install it by:

git clone https://github.com/sail-sg/Adan.git
cd Adan
python3 setup.py install --unfused

A brief comparison of peak memory and wall duration for the optimizer is as follows. The duration time is the total time of 200 optimizer.step(). We further compare Adam and FusedAdan in great detail on GPT-2. See more results here.

Usage

For your convenience to use Adan, we briefly provide some intuitive instructions below, then provide some general experimental tips, and finally provide more details (e.g., specific commands and hyper-parameters) for each experiment in the paper.

1) Two steps to use Adan

Step 1. add Adan-dependent hyper-parameters by adding the following hyper-parameters to the config:

parser.add_argument('--max-grad-norm', type=float, default=0.0, help='if the l2 norm is large than this hyper-parameter, then we clip the gradient  (default: 0.0, no gradient clip)')
parser.add_argument('--weight-decay', type=float, default=0.02,  help='weight decay, similar one used in AdamW (default: 0.02)')
parser.add_argument('--opt-eps', default=None, type=float, metavar='EPSILON', help='optimizer epsilon to avoid the bad case where second-order moment is zero (default: None, use opt default 1e-8 in adan)')
parser.add_argument('--opt-betas', default=None, type=float, nargs='+', metavar='BETA', help='optimizer betas in Adan (default: None, use opt default [0.98, 0.92, 0.99] in Adan)')
parser.add_argument('--no-prox', action='store_true', default=False, help='whether perform weight decay like AdamW (default=False)')

opt-betas: To keep consistent with our usage habits, the $\beta$'s in the paper are actually the $(1-\beta)$'s in the code.

foreach (bool): If True, Adan would use torch._foreach implementation. It is faster but uses slightly more memory.

no-prox: It determines the update rule of parameters with weight decay. By default, Adan updates the parameters in the way presented in Algorithm 1 in the paper:

$$\boldsymbol{\theta}_{k+1} = ( 1+\lambda \eta)^{-1}\left[\boldsymbol{\theta}_k - \boldsymbol{\eta}_k \circ (\mathbf{m}_k+(1-{\color{blue}\beta_2})\mathbf{v}k)\right],$$

But one also can update the parameter like Adamw:

$$\boldsymbol{\theta}_{k+1} = ( 1-\lambda \eta)\boldsymbol{\theta}_k - \boldsymbol{\eta}_k \circ (\mathbf{m}_k+(1-{\color{blue}\beta_2})\mathbf{v}_k).$$ In all experiments, we set no-prox=False in our paper.

Step 2. create the Adan optimizer as follows. In this step, we can directly replace the vanilla optimizer by using the following command:

from adan import Adan
optimizer = Adan(param, lr=args.lr, weight_decay=args.weight_decay, betas=args.opt_betas, eps = args.opt_eps, max_grad_norm=args.max_grad_norm, no_prox=args.no_prox)

2) Tips for Experiments

• To make Adan simple, in all experiments except Table 12 in the paper, we do not use the restart strategy in Adan. But Table 12 shows that the restart strategy can further slightly improve the performance of Adan.
• Adan often allows one to use a large peak learning rate which often fails other optimizers, e.g., Adam and AdamW. For example, in all experiments except for the MAE pre-training and LSTM, the learning rate used by Adan is 5-10 times larger than that in Adam/AdamW.
• Adan is relatively robust to beta1, beta2, and beta3, especially for beta2. If you want better performance, you can first tune beta3 and then beta1.
• Interestingly, we found that weight_decay = 0.02 is suitable for all experiments in our paper.
• Adan has a slightly higher GPU memory cost than Adam/AdamW on a single node. However, this problem can be solved using the ZeroRedundancyOptimizer, which shares optimizer states across distributed data-parallel processes to reduce per-process memory footprint. Specifically, when using the ZeroRedundancyOptimizer on more than two GPUs, Adan and Adam consume almost the same amount of memory.

3) More extra detailed steps&results

Please refer to the following links for detailed steps. In these detailed steps, we even include the docker images for reproducibility.

• Instruction for ViTs, ResNets, and ConvNext.
• Instruction for MAE.
• Instruction for BERT.
• Instruction for Transformer-XL.
• Instruction for GPT2
• Resutls for text-to-3D DreamFusion.

Model Zoo

Results on vision tasks

For your convenience to use Adan, we provide the configs and log files for the experiments on ImageNet-1k.

Results on NLP tasks

BERT-base

We give the configs and log files of the BERT-base model pre-trained on the Bookcorpus and Wikipedia datasets and fine-tuned on GLUE tasks. Note that we provide the config, log file, and detailed instructions for BERT-base in the folder ./NLP/BERT.

For fine-tuning on GLUE-Task, see the total batch size in their corresponding configure files.

Transformer-XL-base

We provide the config and log for Transformer-XL-base trained on the WikiText-103 dataset. The total batch size for this experiment is 60*4.

Results on Large Language Models

GPT2-345m

We provide the config and log for GPT2-345m pre-trained on the dataset that comes from BigCode and evaluated on the HumanEval dataset by zero-shot learning. HumanEval is used to measure functional correctness for synthesizing programs from docstrings. It consists of 164 original programming problems, assessing language comprehension, algorithms, and simple mathematics, with some comparable to simple software interview questions. We set Temperature = 0.8 during evaluation.

Adan obtains comparable results with only half cost.

Results on Diffusion Models

We show the results of the text-to-3D task supported by the DreamFusion Project. More visualization results could be founded here. Examples generated from text prompt Sydney opera house, aerial view with Adam and Adan: