mindspore/model_zoo/official/cv/dpn

• Contents
• Description
• Model Architecture
• Dataset
• Features
• • Mixed Precision
• Environment Requirements
• Quick Start
• • Dataset Preparation
  • Running
• Script Description
• • Script and Sample Code
  • Script Parameters
  • Training Process
  • • Training
    • • Running on Ascend
    • Distributed Training
    • • Running on Ascend
  • Evaluation Process
  • • Running on Ascend
• Model Description
• • Performance
  • • Accuracy
    • • DPN92 (Training)
    • Efficiency
    • • DPN92
• Description of Random Situation
• ModelZoo Homepage

Contents

• Description
• Model Architecture
• Dataset
• Features
  • Mixed Precision
• Environment Requirements
• Quick Start
  • Dataset Preparation
  • Running
• Script Description
  • Script and Sample Code
  • Script Parameters
  • Training Process
    • Training
      • Running on Ascend
    • Distributed Training
      • Running on Ascend
  • Evaluation Process
    • Running on Ascend
• Model Description
  • Performance
    • Accuracy
      • DPN92 (Training)
    • Efficiency
      • DPN92
• Description of Random Situation
• ModelZoo Homepage

Description

Dual Path Network (DPN) is a convolution-based neural network for the task of image classification. It combines the advantage of both ResNeXt and DenseNet to get higher accuracy. More detail about this model can be found in:

Yunpeng Chen, Jianan Li, Huaxin Xiao, Xiaojie Jin, Shuicheng Yan, Jiashi Feng. "Dual Path Networks" (NIPS17).

This repository contains a Mindspore implementation of DPNs based upon cypw's original MXNet implementation (https://github.com/cypw/DPNs). The training and validating scripts are also included, and the validation results with cypw’s pretrained weights are shown in the Results section.

Model Architecture

The overall network architecture of DPN is show below:

Link

Dataset

All the models in this repository are trained and validated on ImageNet-1K. The models can achieve the results with the configurations of the dataset preprocessing as follow:

• For the training dataset:

  • Range (min, max) of the respective size of the original size to be cropped is (0.08, 1.0)

  • Range (min, max) of aspect ratio to be cropped is (0.75, 1.333)

  • The size of input images is reshaped to (width = 224, height = 224)

  • Probability of random horizontal flip is 50%

  • In normalization, the mean is (255*0.485, 255*0.456, 255*0.406) and the standard deviation is (255*0.229, 255*0.224, 255*0.225)

• For the evaluation dataset:

  • Input size of images is 224*224 (Resize to 256*256 then crops images at the center)
  • In normalization, the mean is (255*0.485, 255*0.456, 255*0.406) and the standard deviation is (255*0.229, 255*0.224, 255*0.225)

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

To run the python scripts in the repository, you need to prepare the environment as follow:

• Hardware
  • Prepare hardware environment with Ascend or GPU processor.
• Python and dependencies
  • Python3.7
  • Mindspore 1.1.0
• For more information, please check the resources below：
  • MindSpore tutorials
  • MindSpore Python API

Quick Start

Dataset Preparation

The DPN models use ImageNet-1K dataset to train and validate in this repository. Download the dataset from ImageNet.org. You can place them anywhere and tell the scripts where they are when running.

Running

To train the DPNs, run the shell script scripts/train_standalone.sh with the format below:

sh scripts/train_standalone.sh [device_id] [dataset_dir] [ckpt_path_to_save] [eval_each_epoch] [pretrained_ckpt(optional)]

To validate the DPNs, run the shell script scripts/eval.sh with the format below:

sh scripts/eval.sh [device_id] [dataset_dir] [pretrained_ckpt]

Script Description

Script and Sample Code

The structure of the files in this repository is shown below.

└─ mindspore-dpns
    ├─ scripts
    │   ├─ eval.sh                // launch ascend standalone evaluation
    │   ├─ train_distributed.sh   // launch ascend distributed training
    │   └─ train_standalone.sh    // launch ascend standalone training
    ├─ src
    │   ├─ config.py              // network and running config
    │   ├─ crossentropy.py        // loss function
    │   ├─ dpn.py                 // dpns implementation
    │   ├─ imagenet_dataset.py    // dataset processor and provider
    │   └─ lr_scheduler.py        // dpn learning rate scheduler
    ├─ eval.py                    // evaluation script
    ├─ train.py                   // training script
    ├─ export.py                  // export model
    └─ README.md                  // descriptions about this repository

Script Parameters

Parameters for both training and evaluation can be set in src/config.py

• Configurations for DPN92 with ImageNet-1K dataset

# model config
config.image_size = (224,224)               # inpute image size
config.num_classes = 1000                   # dataset class number
config.backbone =  'dpn92'                  # backbone network
config.is_save_on_master = True

# parallel config
config.num_parallel_workers = 4             # number of workers to read the data
config.rank = 0                             # local rank of distributed
config.group_size =  1                      # group size of distributed

# training config
config.batch_size =  32                     # batch_size
config.global_step =  0                     # start step of learning rate
config.epoch_size = 180                     # epoch_size
config.loss_scale_num =  1024               # loss scale
# optimizer config
config.momentum =  0.9                      # momentum (SGD)
config.weight_decay =  1e-4                 # weight_decay (SGD)
# learning rate config
config.lr_schedule = 'warmup'               # learning rate schedule
config.lr_init = 0.01                       # init learning rate
config.lr_max = 0.1                         # max learning rate
config.factor =  0.1                        # factor of lr to drop
config.epoch_number_to_drop = [5,15]        # learing rate will drop after these epochs
config.warmup_epochs = 5                    # warmup epochs in learning rate schedule

# dataset config
config.dataset = "imagenet-1K"              # dataset
config.label_smooth = False                 # label_smooth
config.label_smooth_factor = 0.0            # label_smooth_factor

# parameter save config
config.keep_checkpoint_max = 3              # only keep the last keep_checkpoint_max checkpoint

Training Process

Training

Running on Ascend

Run scripts/train_standalone.sh to train the model standalone. The usage of the script is:

sh scripts/train_standalone.sh [device_id] [dataset_dir] [ckpt_path_to_save] [eval_each_epoch] [pretrained_ckpt(optional)]

For example, you can run the shell command below to launch the training procedure.

sh scripts/train_standalone.sh 0 /data/dataset/imagenet/ scripts/pretrian/ 0

If eval_each_epoch is 1, it will evaluate after each epoch and save the parameters with the max accuracy. But in this case, the time of one epoch will be longer.

If eval_each_epoch is 0, it will save parameters every some epochs instead of evaluating in the training process.

The script will run training in the background, you can view the results through the file train_log.txt as follows (eval_each_epoch = 0):

epoch: 1 step: 40036, loss is 3.6232593
epoch time: 10048893.336 ms, per step time: 250.996 ms
epoch: 2 step: 40036, loss is 3.200775
epoch time: 9306154.456 ms, per step time: 232.445 ms
...

or as follows (eval_each_epoch = 1):

epoch: 1 step: 40036, loss is 3.6232593
epoch time: 10048893.336 ms, per step time: 250.996 ms
Save the maximum accuracy checkpoint,the accuracy is 0.2629158669225848
...

The model checkpoint will be saved into [ckpt_path_to_save].

Distributed Training

Running on Ascend

Run scripts/train_distributed.sh to train the model distributed. The usage of the script is:

sh scripts/train_distributed.sh [rank_table] [dataset_dir] [ckpt_path_to_save]  [rank_size] [eval_each_epoch] [pretrained_ckpt(optional)]

For example, you can run the shell command below to launch the training procedure.

sh scripts/train_distributed.sh /home/rank_table.json /data/dataset/imagenet/ ../scripts 8 0 ../pretrain/dpn92.ckpt

The above shell script will run distribute training in the background. You can view the results through the file train_parallel[X]/log.txt as follows:

epoch: 1 step 5004, loss is 4.5680037
epoch time: 2312519.441 ms, per step time: 462.134 ms
epoch: 2 step 5004, loss is 2.964888
Epoch time: 1350398.913 ms, per step time: 369.864 ms
...

The model checkpoint will be saved into [ckpt_path_to_save].

Evaluation Process

Running on Ascend

Run scripts/eval.sh to evaluate the model with one Ascend processor. The usage of the script is:

sh scripts/eval.sh [device_id] [dataset_dir] [pretrained_ckpt]

For example, you can run the shell command below to launch the validation procedure.

sh scripts/eval.sh 0 /data/dataset/imagenet/ pretrain/dpn-180_5004.ckpt

The above shell script will run evaluation in the background. You can view the results through the file eval_log.txt. The result will be achieved as follows:

Evaluation result: {'top_5_accuracy': 0.9449223751600512, 'top_1_accuracy': 0.7911731754161332}.
DPN evaluate success!

Model Description

Performance

The evaluation of model performance is divided into two parts: accuracy and efficiency. The part of accuracy shows the accuracy of the model in classifying images on ImageNet-1K dataset, and it can be evaluated by top-k measure. The part of efficiency reveals the time cost by model training on ImageNet-1K.

All results are validated at image size of 224x224. The dataset preprocessing and training configurations are shown in Dataset section.

Accuracy

DPN92 (Training)

Parameters	Ascend
Model Version	DPN92 (Train)
Resource	Ascend 910; OS Euler2.8
Uploaded Date	12/20/2020 (month/day/year)
MindSpore Version	1.1.0
Dataset	ImageNet-1K
epochs	180
outputs	probability
train performance	Top1:78.91%; Top5:94.53%

Efficiency

DPN92

Parameters	Ascend
Model Version	DPN92
Resource	Ascend 910; OS Euler2.8
Uploaded Date	12/20/2020 (month/day/year)
MindSpore Version	1.1.0
Dataset	ImageNet-1K
batch_size	32
outputs	probability
speed	1pc:233 ms/step;8pc:240 ms/step

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.