• Contents
• • SSD Description
  • Model Architecture
  • Dataset
  • Environment Requirements
  • Quick Start
  • • Run the scripts
  • Script Description
  • • Script and Sample Code
    • Script Parameters
    • Training Process
    • • Training on Ascend
    • Evaluation Process
    • • Evaluation on Ascend
  • Export Process
  • • Export
  • Inference Process
  • • Inference
    • Performance
  • Export MindIR
  • Description of Random Situation
  • ModelZoo Homepage

Contents

• Contents
  • SSD Description
  • Model Architecture
  • Dataset
  • Environment Requirements
  • Quick Start
    • Prepare the model
    • Run the scripts
  • Script Description
    • Script and Sample Code
    • Script Parameters
    • Training Process
      • Training on Ascend
    • Evaluation Process
      • Evaluation on Ascend
      • Performance
    • Export Process
      • Export
    • Inference Process
      • Inference
  • Description of Random Situation
  • ModelZoo Homepage

SSD Description

SSD discretizes the output space of bounding boxes into a set of default boxes over different aspect ratios and scales per feature map location. At prediction time, the network generates scores for the presence of each object category in each default box and produces adjustments to the box to better match the object shape.Additionally, the network combines predictions from multiple feature maps with different resolutions to naturally handle objects of various sizes.

Paper: Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, Alexander C. Berg.European Conference on Computer Vision (ECCV), 2016 (In press).

Model Architecture

The SSD approach is based on a feed-forward convolutional network that produces a fixed-size collection of bounding boxes and scores for the presence of object class instances in those boxes, followed by a non-maximum suppression step to produce the final detections. The early network layers are based on a standard architecture used for high quality image classification, which is called the base network. Then add auxiliary structure to the network to produce detections.

Dataset

Note that you can run the scripts based on the dataset mentioned in original paper or widely used in relevant domain/network architecture. In the following sections, we will introduce how to run the scripts using the related dataset below.

Dataset used: COCO2017

• Dataset size：19G
  • Train：18G，118000 images
  • Val：1G，5000 images
  • Annotations：241M，instances，captions，person_keypoints etc
• Data format：image and json files
  • Note：Data will be processed in dataset.py

Environment Requirements

• Install MindSpore.

• Download the dataset COCO2017.

• We use COCO2017 as training dataset in this example by default, and you can also use your own datasets.
  First, install Cython ,pycocotool and opencv to process data and to get evaluation result.

  pip install Cython
  
  pip install pycocotools
  
  pip install opencv-python
  

  1. If coco dataset is used. Select dataset to coco when run script.

     Change the coco_root and other settings you need in src/config.py. The directory structure is as follows:

     .
     └─coco_dataset
       ├─annotations
         ├─instance_train2017.json
         └─instance_val2017.json
       ├─val2017
       └─train2017
     

  2. If VOC dataset is used. Select dataset to voc when run script.
     Change classes, num_classes, voc_json and voc_root in src/config.py. voc_json is the path of json file with coco format for evaluation, voc_root is the path of VOC dataset, the directory structure is as follows:

     .
     └─voc_dataset
       └─train
         ├─0001.jpg
         └─0001.xml
         ...
         ├─xxxx.jpg
         └─xxxx.xml
       └─eval
         ├─0001.jpg
         └─0001.xml
         ...
         ├─xxxx.jpg
         └─xxxx.xml
     

  3. If your own dataset is used. Select dataset to other when run script.
     Organize the dataset information into a TXT file, each row in the file is as follows:

     train2017/0000001.jpg 0,259,401,459,7 35,28,324,201,2 0,30,59,80,2
     

     Each row is an image annotation which split by space, the first column is a relative path of image, the others are box and class infomations of the format [xmin,ymin,xmax,ymax,class]. We read image from an image path joined by the image_dir(dataset directory) and the relative path in anno_path(the TXT file path), image_dir and anno_path are setting in src/config.py.

Quick Start

Run the scripts

After installing MindSpore via the official website, you can start training and evaluation as follows:

• running on Ascend

# distributed training on Ascend
bash run_distribute_train.sh [DEVICE_NUM] [EPOCH_SIZE] [LR] [DATASET] [RANK_TABLE_FILE]

# training on single NPU
bash run_standalone_train.sh

# run eval on Ascend
bash run_eval.sh [DATASET] [CHECKPOINT_PATH] [DEVICE_ID]

• Run on docker

Build docker images(Change version to the one you actually used)

# build docker
docker build -t ssd:20.1.0 . --build-arg FROM_IMAGE_NAME=ascend-mindspore-arm:20.1.0

Create a container layer over the created image and start it

# start docker
bash scripts/docker_start.sh ssd:20.1.0 [DATA_DIR] [MODEL_DIR]

Then you can run everything just like on ascend.

Script Description

Script and Sample Code

.
└─ cv
  └─ ssd
    ├─ README.md                      # descriptions about SSD
    ├─ ascend310_infer                # application for 310 inference
    ├─ scripts
      ├─ run_distribute_train.sh      # shell script for distributed on ascend
      ├─ run_eval.sh                  # shell script for eval on ascend
      └─ run_infer_310.sh             # shell script for 310 inference
    ├─ src
      ├─ __init__.py                  # init file
      ├─ box_utils.py                 # bbox utils
      ├─ eval_utils.py                # metrics utils
      ├─ config.py                    # total config
      ├─ dataset.py                   # create dataset and process dataset
      ├─ init_params.py               # parameters utils
      ├─ lr_schedule.py               # learning ratio generator
      └─ ssd.py                       # ssd architecture
    ├─ eval.py                        # eval scripts
    ├─ train.py                       # train scripts
    ├─ export.py                      # export mindir script
    ├─ postprogress.py                # post process for 310 inference
    └─ mindspore_hub_conf.py          # mindspore hub interface

Script Parameters

Major parameters in train.py and config.py as follows:

  "device_num": 1                                  # Use device nums
  "lr": 0.05                                       # Learning rate init value
  "dataset": coco                                  # Dataset name
  "epoch_size": 500                                # Epoch size
  "batch_size": 32                                 # Batch size of input tensor
  "pre_trained": None                              # Pretrained checkpoint file path
  "pre_trained_epoch_size": 0                      # Pretrained epoch size
  "save_checkpoint_epochs": 10                     # The epoch interval between two checkpoints. By default, the checkpoint will be saved per 10 epochs
  "loss_scale": 1024                               # Loss scale
  "filter_weight": False                           # Load parameters in head layer or not. If the class numbers of train dataset is different from the class numbers in pre_trained checkpoint, please set True.
  "freeze_layer": "none"                           # Freeze the backbone parameters or not, support none and backbone.

  "class_num": 81                                  # Dataset class number
  "image_shape": [300, 300]                        # Image height and width used as input to the model
  "mindrecord_dir": "/data/MindRecord_COCO"        # MindRecord path
  "coco_root": "/data/coco2017"                    # COCO2017 dataset path
  "voc_root": "/data/voc_dataset"                  # VOC original dataset path
  "voc_json": "annotations/voc_instances_val.json" # is the path of json file with coco format for evaluation
  "image_dir": ""                                  # Other dataset image path, if coco or voc used, it will be useless
  "anno_path": ""                                  # Other dataset annotation path, if coco or voc used, it will be useless

Training Process

To train the model, run train.py. If the mindrecord_dir is empty, it will generate mindrecord files by coco_root(coco dataset), voc_root(voc dataset) or image_dir and anno_path(own dataset). Note if mindrecord_dir isn't empty, it will use mindrecord_dir instead of raw images.

Training on Ascend

• Distribute mode

    bash run_distribute_train.sh [DEVICE_NUM] [EPOCH_SIZE] [LR] [DATASET] [RANK_TABLE_FILE] [PRE_TRAINED](optional) [PRE_TRAINED_EPOCH_SIZE](optional)

We need five or seven parameters for this scripts.

• DEVICE_NUM: the device number for distributed train.
• EPOCH_NUM: epoch num for distributed train.
• LR: learning rate init value for distributed train.
• DATASET：the dataset mode for distributed train.
• RANK_TABLE_FILE : the path of rank_table.json, it is better to use absolute path.
• PRE_TRAINED : the path of pretrained checkpoint file, it is better to use absolute path.
• PRE_TRAINED_EPOCH_SIZE : the epoch num of pretrained.

Training result will be stored in the current path, whose folder name begins with "LOG". Under this, you can find checkpoint file together with result like the followings in log

epoch: 1 step: 458, loss is 3.1681802
epoch time: 228752.4654865265, per step time: 499.4595316299705
epoch: 2 step: 458, loss is 2.8847265
epoch time: 38912.93382644653, per step time: 84.96273761232868
epoch: 3 step: 458, loss is 2.8398118
epoch time: 38769.184827804565, per step time: 84.64887516987896
...

epoch: 498 step: 458, loss is 0.70908034
epoch time: 38771.079778671265, per step time: 84.65301261718616
epoch: 499 step: 458, loss is 0.7974688
epoch time: 38787.413120269775, per step time: 84.68867493508685
epoch: 500 step: 458, loss is 0.5548882
epoch time: 39064.8467540741, per step time: 85.29442522723602

Evaluation Process

Evaluation on Ascend

bash run_eval.sh [DATASET] [CHECKPOINT_PATH] [DEVICE_ID]

We need two parameters for this scripts.

• DATASET：the dataset mode of evaluation dataset.
• CHECKPOINT_PATH: the absolute path for checkpoint file.
• DEVICE_ID: the device id for eval.

checkpoint can be produced in training process.

Inference result will be stored in the example path, whose folder name begins with "eval". Under this, you can find result like the followings in log.

Average Precision  (AP) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.327
Average Precision  (AP) @[ IoU=0.50      | area=   all | maxDets=100 ] = 0.474
Average Precision  (AP) @[ IoU=0.75      | area=   all | maxDets=100 ] = 0.358
Average Precision  (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.120
Average Precision  (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.350
Average Precision  (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.459
Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=  1 ] = 0.315
Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets= 10 ] = 0.489
Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.511
Average Recall     (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.208
Average Recall     (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.557
Average Recall     (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.689

========================================

mAP: 0.32719216721918915

Export Process

Export

python export.py --ckpt_file [CKPT_PATH] --device_target [DEVICE_TARGET] --file_format[EXPORT_FORMAT]

EXPORT_FORMAT should be in ["AIR", "MINDIR"]

Inference Process

Inference

Before performing inference, we need to export model first. Air model can only be exported in Ascend 910 environment, mindir model can be exported in any environment.
Current batch_ Size can only be set to 1.

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

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

Average Precision  (AP) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.327
Average Precision  (AP) @[ IoU=0.50      | area=   all | maxDets=100 ] = 0.475
Average Precision  (AP) @[ IoU=0.75      | area=   all | maxDets=100 ] = 0.358
Average Precision  (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.115
Average Precision  (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.353
Average Precision  (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.455
Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=  1 ] = 0.314
Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets= 10 ] = 0.485
Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.509
Average Recall     (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.200
Average Recall     (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.554
Average Recall     (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.692

mAP: 0.3266651054070853

Performance

参数	Ascend
模型版本	SSD resnet50
资源	Ascend 910
上传日期	2020-03-29
MindSpore版本	1.1.0
数据集	COCO2017
mAP	IoU=0.50: 32.7%
模型大小	281M（.ckpt文件）

Export MindIR

python export.py --ckpt_file [CKPT_PATH] --file_name [FILE_NAME] --file_format [FILE_FORMAT]

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

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.