mindspore/model_zoo/official/cv/resnext50

• Contents
• ResNeXt50 Description
• Model architecture
• Dataset
• Features
• • Mixed Precision
• Environment Requirements
• Script description
• • Script and sample code
  • Script Parameters
  • Training Process
  • • • Usage
      • Launch
  • Evaluation Process
  • • Usage
    • • Launch
      • Result
• Model description
• • Performance
  • • Training Performance
    • • Inference Performance
• Description of Random Situation
• ModelZoo Homepage

Contents

• ResNeXt50 Description
• Model Architecture
• Dataset
• Features
  • Mixed Precision
• Environment Requirements
• Quick Start
• Script Description
  • Script and Sample Code
  • Script Parameters
  • Training Process
  • Evaluation Process
• Model Description
  • Performance
    • Training Performance
    • Inference Performance
• Description of Random Situation
• ModelZoo Homepage

ResNeXt50 Description

ResNeXt is a simple, highly modularized network architecture for image classification. It designs results in a homogeneous, multi-branch architecture that has only a few hyper-parameters to set in ResNeXt. This strategy exposes a new dimension, which we call “cardinality” (the size of the set of transformations), as an essential factor in addition to the dimensions of depth and width.

Paper: Xie S, Girshick R, Dollár, Piotr, et al. Aggregated Residual Transformations for Deep Neural Networks. 2016.

Model architecture

The overall network architecture of ResNeXt 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

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）
  • Prepare hardware environment with Ascend or GPU processor. If you want to try Ascend , please send the application form to ascend@huawei.com. Once approved, you can get the resources.
• Framework
  • MindSpore
• For more information, please check the resources below：
  • MindSpore tutorials
  • MindSpore API

Script description

Script and sample code

.
└─resnext50      
  ├─README.md
  ├─scripts      
    ├─run_standalone_train.sh         # launch standalone training for ascend(1p)
    ├─run_distribute_train.sh         # launch distributed training for ascend(8p)
    ├─run_standalone_train_for_gpu.sh # launch standalone training for gpu(1p)
    ├─run_distribute_train_for_gpu.sh # launch distributed training for gpu(8p)
    └─run_eval.sh                     # launch evaluating
  ├─src
    ├─backbone
      ├─_init_.py                     # initalize
      ├─resnet.py                     # resnext50 backbone
    ├─utils
      ├─_init_.py                     # initalize
      ├─cunstom_op.py                 # network operation
      ├─logging.py                    # print log
      ├─optimizers_init_.py           # get parameters
      ├─sampler.py                    # distributed sampler
      ├─var_init_.py                  # calculate gain value
    ├─_init_.py                       # initalize
    ├─config.py                       # parameter configuration
    ├─crossentropy.py                 # CrossEntropy loss function
    ├─dataset.py                      # data preprocessing
    ├─head.py                         # commom head
    ├─image_classification.py         # get resnet
    ├─linear_warmup.py                # linear warmup learning rate
    ├─warmup_cosine_annealing.py      # learning rate each step
    ├─warmup_step_lr.py               # warmup step learning rate
  ├─eval.py                           # eval net
  └─train.py                          # train net
  

Script Parameters

Parameters for both training and evaluating can be set in config.py.

"image_height": '224,224'                 # image size
"num_classes": 1000,                      # dataset class number
"per_batch_size": 128,                    # batch size of input tensor
"lr": 0.05,                               # base learning rate
"lr_scheduler": 'cosine_annealing',       # learning rate mode
"lr_epochs": '30,60,90,120',              # epoch of lr changing
"lr_gamma": 0.1,                          # decrease lr by a factor of exponential lr_scheduler
"eta_min": 0,                             # eta_min in cosine_annealing scheduler
"T_max": 150,                             # T-max in cosine_annealing scheduler
"max_epoch": 150,                         # max epoch num to train the model
"backbone": 'resnext50',                  # backbone metwork
"warmup_epochs" : 1,                      # warmup epoch
"weight_decay": 0.0001,                   # weight decay
"momentum": 0.9,                          # momentum
"is_dynamic_loss_scale": 0,               # dynamic loss scale
"loss_scale": 1024,                       # loss scale
"label_smooth": 1,                        # label_smooth
"label_smooth_factor": 0.1,               # label_smooth_factor
"ckpt_interval": 2000,                    # ckpt_interval
"ckpt_path": 'outputs/',                  # checkpoint save location
"is_save_on_master": 1,
"rank": 0,                                # local rank of distributed
"group_size": 1                           # world size of distributed

Training Process

Usage

You can start training by python script:

python train.py --data_dir ~/imagenet/train/ --platform Ascend --is_distributed 0

or shell stript:

Ascend: 
    # distribute training example(8p)
    sh run_distribute_train.sh RANK_TABLE_FILE DATA_PATH
    # standalone training
    sh run_standalone_train.sh DEVICE_ID DATA_PATH
GPU:
    # distribute training example(8p)
    sh run_distribute_train_for_gpu.sh DATA_PATH
    # standalone training
    sh run_standalone_train_for_gpu.sh DEVICE_ID DATA_PATH

Launch

# distributed training example(8p) for Ascend
sh scripts/run_distribute_train.sh RANK_TABLE_FILE /dataset/train
# standalone training example for Ascend
sh scripts/run_standalone_train.sh 0 /dataset/train

# distributed training example(8p) for GPU
sh scripts/run_distribute_train_for_gpu.sh /dataset/train
# standalone training example for GPU
sh scripts/run_standalone_train_for_gpu.sh 0 /dataset/train

You can find checkpoint file together with result in log.

Evaluation Process

Usage

You can start training by python script:

python eval.py --data_dir ~/imagenet/val/ --platform Ascend --pretrained resnext.ckpt

or shell stript:

# Evaluation
sh run_eval.sh DEVICE_ID DATA_PATH PRETRAINED_CKPT_PATH PLATFORM

PLATFORM is Ascend or GPU, default is Ascend.

Launch

# Evaluation with checkpoint
sh scripts/run_eval.sh 0 /opt/npu/datasets/classification/val /resnext50_100.ckpt Ascend

Result

Evaluation result will be stored in the scripts path. Under this, you can find result like the followings in log.

acc=78.16%(TOP1)
acc=93.88%(TOP5)

Model description

Performance

Training Performance

Parameters	ResNeXt50
Resource	Ascend 910, cpu:2.60GHz 56cores, memory:314G	NV SMX2 V100-32G
uploaded Date	06/30/2020	07/23/2020
MindSpore Version	0.5.0	0.6.0
Dataset	ImageNet	ImageNet
Training Parameters	src/config.py	src/config.py
Optimizer	Momentum	Momentum
Loss Function	SoftmaxCrossEntropy	SoftmaxCrossEntropy
Loss	1.76592	1.8965
Accuracy	78%(TOP1)	77.8%(TOP1)
Total time	7.8 h 8ps	21.5 h 8ps
Checkpoint for Fine tuning	192 M(.ckpt file)	192 M(.ckpt file)

Inference Performance

Parameters
Resource	Huawei 910	NV SMX2 V100-32G	Huawei 310
uploaded Date	06/30/2020	07/23/2020	07/23/2020
MindSpore Version	0.5.0	0.6.0	0.6.0
Dataset	ImageNet, 1.2W	ImageNet, 1.2W	ImageNet, 1.2W
batch_size	1	1	1
outputs	probability	probability	probability
Accuracy	acc=78.16%(TOP1)	acc=78.05%(TOP1)

Description of Random Situation

In dataset.py, we set the seed inside “create_dataset" function. We also use random seed in train.py.

ModelZoo Homepage

Please check the official homepage.