mindspore/model_zoo/official/cv/mobilenetv2

Contents

• MobileNetV2 Description
• Model Architecture
• Dataset
• Features
  • Mixed Precision
• Environment Requirements
• Script Description
  • Script and Sample Code
    • Training Process
    • Evaluation Process
    • Inference Process
      • Export MindIR
      • Infer on Ascend310
      • result
• Model Description
  • Performance
    • Training Performance
    • Evaluation Performance
    • Inference Performance
• Description of Random Situation
• ModelZoo Homepage

MobileNetV2 Description

MobileNetV2 is tuned to mobile phone CPUs through a combination of hardware- aware network architecture search (NAS) complemented by the NetAdapt algorithm and then subsequently improved through novel architecture advances.Nov 20, 2019.

Paper Howard, Andrew, Mark Sandler, Grace Chu, Liang-Chieh Chen, Bo Chen, Mingxing Tan, Weijun Wang et al. "Searching for MobileNetV2." In Proceedings of the IEEE International Conference on Computer Vision, pp. 1314-1324. 2019.

Model architecture

The overall network architecture of MobileNetV2 is show below:

Link

Dataset

Dataset used: imagenet

• Dataset size: ~125G, 1.2W colorful images in 1000 classes
  • Train: 120G, 1.2W images
  • Test: 5G, 50000 images
• Data format: RGB images.
  • Note: Data will be processed in src/dataset.py

Features

Mixed Precision(Ascend)

The mixed precision training method accelerates the deep learning neural network training process by using both the single-precision and half-precision data formats, and maintains the network precision achieved by the single-precision training at the same time. Mixed precision training can accelerate the computation process, reduce memory usage, and enable a larger model or batch size to be trained on specific hardware. For FP16 operators, if the input data type is FP32, the backend of MindSpore will automatically handle it with reduced precision. Users could check the reduced-precision operators by enabling INFO log and then searching ‘reduce precision’.

Environment Requirements

• Hardware（Ascend/GPU/CPU）
  • Prepare hardware environment with Ascend, GPU or CPU processor.
• Framework
  • MindSpore
• For more information, please check the resources below：
  • MindSpore Tutorials
  • MindSpore Python API

Script description

Script and sample code

├── MobileNetV2
  ├── README.md                  # descriptions about MobileNetV2
  ├── ascend310_infer            # application for 310 inference
  ├── scripts
  │   ├──run_infer_310.sh        # shell script for 310 infer
  │   ├──run_train.sh            # shell script for train, fine_tune or incremental  learn with CPU, GPU or Ascend
  │   ├──run_eval.sh             # shell script for evaluation with CPU, GPU or Ascend
  │   ├──cache_util.sh           # a collection of helper functions to manage cache
  │   ├──run_train_nfs_cache.sh  # shell script for train with NFS dataset and leverage caching service for better performance
  ├── src
  │   ├──aipp.cfg                # aipp config
  │   ├──args.py                 # parse args
  │   ├──config.py               # parameter configuration
  │   ├──dataset.py              # creating dataset
  │   ├──lr_generator.py         # learning rate config
  │   ├──mobilenetV2.py          # MobileNetV2 architecture
  │   ├──models.py               # contain define_net and Loss, Monitor
  │   ├──utils.py                # utils to load ckpt_file for fine tune or incremental learn
  ├── train.py                   # training script
  ├── eval.py                    # evaluation script
  ├── export.py                  # export mindir script
  ├── mindspore_hub_conf.py      #  mindspore hub interface
  ├── postprocess.py             # postprocess script

Training process

Usage

You can start training using python or shell scripts. The usage of shell scripts as follows:

• Ascend: sh run_train.sh Ascend [DEVICE_NUM] [VISIABLE_DEVICES(0,1,2,3,4,5,6,7)] [RANK_TABLE_FILE] [DATASET_PATH] [CKPT_PATH] [FREEZE_LAYER] [FILTER_HEAD]
• GPU: sh run_trian.sh GPU [DEVICE_NUM] [VISIABLE_DEVICES(0,1,2,3,4,5,6,7)] [DATASET_PATH] [CKPT_PATH] [FREEZE_LAYER] [FILTER_HEAD]
• CPU: sh run_trian.sh CPU [DATASET_PATH] [CKPT_PATH] [FREEZE_LAYER] [FILTER_HEAD]

DATASET_PATH is the train dataset path. We use ImageFolderDataset as default dataset, which is a source dataset that reads images from a tree of directories. The directory structure is as follows, and you should use DATASET_PATH=dataset/train for training and DATASET_PATH=dataset/val for evaluation:

        └─dataset
            └─train
              ├─class1
                ├─0001.jpg
                ......
                └─xxxx.jpg
              ......
              ├─classx
                ├─0001.jpg
                ......
                └─xxxx.jpg
            └─val
              ├─class1
                ├─0001.jpg
                ......
                └─xxxx.jpg
              ......
              ├─classx
                ├─0001.jpg
                ......
                └─xxxx.jpg

CKPT_PATH FREEZE_LAYER and FILTER_HEAD are optional, when set CKPT_PATH, FREEZE_LAYER must be set. FREEZE_LAYER should be in ["none", "backbone"], and if you set FREEZE_LAYER="backbone", the parameter in backbone will be freezed when training and the parameter in head will not be load from checkpoint. if FILTER_HEAD=True, the parameter in head will not be load from checkpoint.

RANK_TABLE_FILE is HCCL configuration file when running on Ascend. The common restrictions on using the distributed service are as follows. For details, see the HCCL documentation.

• In a single-node system, a cluster of 1, 2, 4, or 8 devices is supported. In a multi-node system, a cluster of 8 x N devices is supported.
• Each host has four devices numbered 0 to 3 and four devices numbered 4 to 7 deployed on two different networks. During training of 2 or 4 devices, the devices must be connected and clusters cannot be created across networks.

Launch

# training example
  python:
      Ascend: python train.py --platform Ascend --dataset_path [TRAIN_DATASET_PATH]
      GPU: python train.py --platform GPU --dataset_path [TRAIN_DATASET_PATH]
      CPU: python train.py --platform CPU --dataset_path [TRAIN_DATASET_PATH]

  shell:
      Ascend: sh run_train.sh Ascend 8 0,1,2,3,4,5,6,7 hccl_config.json [TRAIN_DATASET_PATH]
      GPU: sh run_train.sh GPU 8 0,1,2,3,4,5,6,7 [TRAIN_DATASET_PATH]
      CPU: sh run_train.sh CPU [TRAIN_DATASET_PATH]

# fine tune whole network example
  python:
      Ascend: python train.py --platform Ascend --dataset_path [TRAIN_DATASET_PATH] --pretrain_ckpt [CKPT_PATH] --freeze_layer none --filter_head True
      GPU: python train.py --platform GPU --dataset_path [TRAIN_DATASET_PATH] --pretrain_ckpt [CKPT_PATH] --freeze_layer none --filter_head True
      CPU: python train.py --platform CPU --dataset_path [TRAIN_DATASET_PATH] --pretrain_ckpt [CKPT_PATH] --freeze_layer none --filter_head True

  shell:
      Ascend: sh run_train.sh Ascend 8 0,1,2,3,4,5,6,7 hccl_config.json [TRAIN_DATASET_PATH]  [CKPT_PATH] none True
      GPU: sh run_train.sh GPU 8 0,1,2,3,4,5,6,7 [TRAIN_DATASET_PATH] [CKPT_PATH] none True
      CPU: sh run_train.sh CPU [TRAIN_DATASET_PATH] [CKPT_PATH] none True

# fine tune full connected layers example
  python:
      Ascend: python --platform Ascend train.py --dataset_path [TRAIN_DATASET_PATH]--pretrain_ckpt [CKPT_PATH] --freeze_layer backbone
      GPU: python --platform GPU train.py --dataset_path [TRAIN_DATASET_PATH] --pretrain_ckpt [CKPT_PATH] --freeze_layer backbone
      CPU: python --platform CPU train.py --dataset_path [TRAIN_DATASET_PATH] --pretrain_ckpt [CKPT_PATH] --freeze_layer backbone

  shell:
      Ascend: sh run_train.sh Ascend 8 0,1,2,3,4,5,6,7 hccl_config.json [TRAIN_DATASET_PATH] [CKPT_PATH] backbone
      GPU: sh run_train.sh GPU 8 0,1,2,3,4,5,6,7 [TRAIN_DATASET_PATH] [CKPT_PATH] backbone
      CPU: sh run_train.sh CPU [TRAIN_DATASET_PATH] [CKPT_PATH] backbone

Result

Training result will be stored in the example path. Checkpoints will be stored at . /checkpoint by default, and training log will be redirected to ./train.log like followings with the platform CPU and GPU, will be wrote to ./train/rank*/log*.log with the platform Ascend .

epoch: [  0/200], step:[  624/  625], loss:[5.258/5.258], time:[140412.236], lr:[0.100]
epoch time: 140522.500, per step time: 224.836, avg loss: 5.258
epoch: [  1/200], step:[  624/  625], loss:[3.917/3.917], time:[138221.250], lr:[0.200]
epoch time: 138331.250, per step time: 221.330, avg loss: 3.917

Evaluation process

Usage

You can start training using python or shell scripts.If the train method is train or fine tune, should not input the [CHECKPOINT_PATH] The usage of shell scripts as follows:

• Ascend: sh run_eval.sh Ascend [DATASET_PATH] [CHECKPOINT_PATH]
• GPU: sh run_eval.sh GPU [DATASET_PATH] [CHECKPOINT_PATH]
• CPU: sh run_eval.sh CPU [DATASET_PATH] [BACKBONE_CKPT_PATH]

Launch

# eval example
  python:
      Ascend: python eval.py --platform Ascend --dataset_path [VAL_DATASET_PATH] --pretrain_ckpt ./ckpt_0/mobilenetv2_15.ckpt
      GPU: python eval.py --platform GPU --dataset_path [VAL_DATASET_PATH] --pretrain_ckpt ./ckpt_0/mobilenetv2_15.ckpt
      CPU: python eval.py --platform CPU --dataset_path [VAL_DATASET_PATH] --pretrain_ckpt ./ckpt_0/mobilenetv2_15.ckpt

  shell:
      Ascend: sh run_eval.sh Ascend [VAL_DATASET_PATH] ./checkpoint/mobilenetv2_head_15.ckpt
      GPU: sh run_eval.sh GPU [VAL_DATASET_PATH] ./checkpoint/mobilenetv2_head_15.ckpt
      CPU: sh run_eval.sh CPU [VAL_DATASET_PATH] ./checkpoint/mobilenetv2_head_15.ckpt

checkpoint can be produced in training process.

Result

Inference result will be stored in the example path, you can find result like the followings in eval.log.

result: {'acc': 0.71976314102564111} ckpt=./ckpt_0/mobilenet-200_625.ckpt

Training with dataset on NFS

You can use script run_train_nfs_cache.sh for running training with a dataset located on Network File System (NFS). By default, a standalone cache server will be started to cache all images in tensor format in memory to improve performance.

Please refer to Training Process for the usage of this shell script.

# training with NFS dataset example
Ascend: sh run_train_nfs_cache.sh Ascend 8 0,1,2,3,4,5,6,7 hccl_config.json [TRAIN_DATASET_PATH]
GPU: sh run_train_nfs_cache.sh GPU 8 0,1,2,3,4,5,6,7 [TRAIN_DATASET_PATH]
CPU: sh run_train_nfs_cache.sh CPU [TRAIN_DATASET_PATH]

With cache enabled, a standalone cache server will be started in the background to cache the dataset in memory. However, Please make sure the dataset fits in memory (around 120GB of memory is required for caching ImageNet train dataset). Users can choose to shutdown the cache server after training or leave it alone for future usage.

Inference process

Export MindIR

python export.py --platform [PLATFORM] --ckpt_file [CKPT_PATH] --file_format [EXPORT_FORMAT]

The ckpt_file parameter is required, EXPORT_FORMAT should be in ["AIR", "MINDIR"]

Infer on Ascend310

Before performing inference, the mindir file must be exported by export.py script. We only provide an example of inference using MINDIR model. Current batch_size can only be set to 1.

# Ascend310 inference
bash run_infer_310.sh [MINDIR_PATH] [DATA_PATH] [LABEL_PATH] [DVPP] [DEVICE_ID]

• LABEL_PATH label.txt path. Write a py script to sort the category under the dataset, map the file names under the categories and category sort values,Such as[file name : sort value], and write the mapping results to the labe.txt file.
• DVPP is mandatory, and must choose from ["DVPP", "CPU"], it's case-insensitive.The size of the picture that MobilenetV2 performs inference is [224, 224], the DVPP hardware limits the width of divisible by 16, and the height is divisible by 2. The network conforms to the standard, and the network can pre-process the image through DVPP.
• DEVICE_ID is optional, default value is 0.

result

Inference result is saved in current path, you can find result like this in acc.log file.

'Accuracy': 0.71654

Model description

Performance

Training Performance

Parameters	MobilenetV2
Model Version	V1	V1
Resource	Ascend 910; cpu 2.60GHz, 192cores; memory 755G; OS Euler2.8	NV SMX2 V100-32G
uploaded Date	05/06/2020	05/06/2020
MindSpore Version	0.3.0	0.3.0
Dataset	ImageNet	ImageNet
Training Parameters	src/config.py	src/config.py
Optimizer	Momentum	Momentum
Loss Function	SoftmaxCrossEntropy	SoftmaxCrossEntropy
outputs	probability	probability
Loss	1.908	1.913
Accuracy	ACC1[71.78%]	ACC1[71.08%]
Total time	753 min	845 min
Params (M)	3.3 M	3.3 M
Checkpoint for Fine tuning	27.3 M	27.3 M
Scripts	Link

Inference Performance

Parameters	Ascend
Model Version	MobilenetV2
Resource	Ascend 310; CentOS 3.10
Uploaded Date	11/05/2021 (month/day/year)
MindSpore Version	1.2.0
Dataset	ImageNet
batch_size	1
outputs	Accuracy
Accuracy	Accuracy=0.71654
Model for inference	27.3M(.ckpt file)

Description of Random Situation

In train.py, we set the seed which is used by numpy.random, mindspore.common.Initializer, mindspore.ops.composite.random_ops and mindspore.nn.probability.distribution.

ModelZoo Homepage

Please check the official homepage.