Yet Another Pytorch Distributed MobileNetV2-based Networks Implementation

This repo reproduces some MobileNetV2-based networks which reach similar or even higher performance than the original reported results.

Model Zoo

All of the following models are trained by MnasNet training schedule.

	FLOPs	Parameters	Top-1 Report	Top-1	Top-1 Calib
MobileNetV2	300M	3.4M	72.0	73.6	-
Proxyless (mobile)	320M	4.1M	74.6	74.9	75.1
SinglePath	334M	4.4M	75.0	75.0	75.1
AtomNAS-A	258M	3.9M	-	-	74.6
AtomNAS-B	326M	4.4M	-	-	75.5
AtomNAS-C	360M	4.7M	-	-	75.9
AtomNAS-A+	260M	4.7M	-	-	76.3
AtomNAS-B+	329M	5.5M	-	-	77.2
AtomNAS-C+	363M	5.9M	-	-	77.6
AutoNL-S	267M	4.4M	-	-	76.5
AutoNL-L1	353M	5.6M	-	-	77.5

AtomNAS series comes from our ICLR 2020 paper AtomNAS: Fine-Grained End-to-End Neural Architecture Search whose code resides in AtomNAS.

AutoNL series comes from our CVPR 2020 paper Neural Architecture Search for Lightweight Non-Local Networks, and the original tf version resides in AutoNL.

Pretrained Models could be downloaded from onedrive

¹ A bit different with tf implementation because of lacking Auto-Augmentation.

Setup

Distributed Training

Set the following ENV variable:

$DATA_ROOT: Path to data root
$METIS_WORKER_0_HOST: IP address of worker 0
$METIS_WORKER_0_PORT: Port used for initializing distributed environment
$METIS_TASK_INDEX: Index of task
$ARNOLD_WORKER_NUM: Number of workers
$ARNOLD_WORKER_GPU: Number of GPUs (NOTE: should exactly match local GPU numbers with `CUDA_VISIBLE_DEVICES `)
$ARNOLD_OUTPUT: Output directory

Non-Distributed Training (Not Recommend)

Set the following ENV variable:

$DATA_ROOT: Path to data root
$ARNOLD_WORKER_GPU: Number of GPUs (NOTE: should exactly match local GPU numbers with `CUDA_VISIBLE_DEVICES `)
$ARNOLD_OUTPUT: Output directory

Training

Take mobilenet v2 as example, for distributed training:

bash scripts/run.sh ./apps/mobilenet/mobilenet_v2_mnas.yml

For non-distributed training:

bash scripts/run_non_distributed_no_copy.sh ./apps/mobilenet/mobilenet_v2_mnas.yml

Testing

For MobileNet, Proxyless, SinglePathNAS, AutoNL:

TRAIN_CONFIG=$(realpath {{train_config_path}}) ATOMNAS_VAL=True bash scripts/run.sh apps/eval/eval.yml --pretrained {{ckpt_path}}

For AtomNAS:

FILE=$(realpath {{log_dir_path}}) checkpoint=ckpt ATOMNAS_VAL=True bash scripts/run.sh apps/eval/eval_shrink.yml

For AtomNAS+:

TRAIN_CONFIG=$(realpath {{train_config_path}}) ATOMNAS_VAL=True bash scripts/run.sh apps/eval/eval_se.yml --pretrained {{ckpt_path}}

Related Info

1. Requirements

   • See requirements.txt

2. Environment

   • The code is developed using python 3. NVIDIA GPUs are needed. The code is developed and tested using 4 servers with 32 NVIDIA V100 GPU cards. Other platforms or GPU cards are not fully tested.

3. Dataset

   • Prepare ImageNet data following pytorch example.
   • Optional: Generate lmdb dataset by utils/lmdb_dataset.py. If not, please overwrite dataset:imagenet1k_lmdb in yaml to dataset:imagenet1k.
   • The directory structure of $DATA_ROOT should look like this:

     ${DATA_ROOT}
     ├── imagenet
     └── imagenet_lmdb
     

4. Miscellaneous

   • The codebase is a general ImageNet training framework using yaml config with several extension under apps dir, based on PyTorch.
     • YAML config with additional features
       • ${ENV} in yaml config.
       • _include for hierachy config.
       • _default key for overwriting.
       • xxx.yyy.zzz for partial overwriting.
     • --{{opt}} {{new_val}} for command line overwriting.

Acknowledgment

This repo is based on slimmable_networks and benefits from the following projects

• apex
• Efficient-PyTorch

Thanks the contributors of these repos!

Citation

If you find this work or code is helpful in your research, please cite:

@inproceedings{
    mei2020atomnas,
    title={Atom{NAS}}: Fine-Grained End-to-End Neural Architecture Search},
    author={Jieru Mei and Yingwei Li and Xiaochen Lian and Xiaojie Jin and Linjie Yang and Alan Yuille and Jianchao Yang},
    booktitle={International Conference on Learning Representations},
    year={2020},
}