mindspore/model_zoo/official/cv/maskrcnn

• Contents
• MaskRCNN Description
• Model Architecture
• Dataset
• Environment Requirements
• Quick Start
• Script Description
• • Script and Sample Code
  • Script Parameters
  • • Training Script Parameters
    • Parameters Configuration
  • Training Process
  • • Training
    • Distributed Training
    • Training Result
  • Evaluation Process
  • • Evaluation
    • Evaluation result
• Model Description
• • Performance
  • • Training Performance
    • Evaluation Performance
• Description of Random Situation
• ModelZoo Homepage

Contents

• MaskRCNN Description
• Model Architecture
• Dataset
• Environment Requirements
• Quick Start
• Script Description
  • Script and Sample Code
  • Script Parameters
    • Training Script Parameters
    • Parameters Configuration
  • Training Process
    • Training
    • Distributed Training
    • Training Result
  • Evaluation Process
    • Evaluation
    • Evaluation Result
• Model Description
  • Performance
    • Training Performance
    • Evaluation Performance
• Description of Random Situation
• ModelZoo Homepage

MaskRCNN Description

MaskRCNN is a conceptually simple, flexible, and general framework for object instance segmentation. The approach efficiently detects objects in an image while simultaneously generating a high-quality segmentation mask for each instance. The method, called Mask R-CNN, extends Faster R-CNN by adding a branch for predicting an object mask in
parallel with the existing branch for bounding box recognition. Mask R-CNN is simple to train and adds only a small overhead to Faster R-CNN, running at 5 fps. Moreover, Mask R-CNN is easy to generalize to other tasks, e.g., allowing to estimate human poses in the same framework.
It shows top results in all three tracks of the COCO suite of challenges, including instance segmentation, boundingbox object detection, and person keypoint detection. Without bells and whistles, Mask R-CNN outperforms all existing, single-model entries on every task, including the COCO 2016 challenge winners.

Model Architecture

MaskRCNN is a two-stage target detection network. It extends FasterRCNN by adding a branch for predicting an object mask in parallel with the existing branch for bounding box recognition.This network uses a region proposal network (RPN), which can share the convolution features of the whole image with the detection network, so that the calculation of region proposal is almost cost free. The whole network further combines RPN and mask branch into a network by sharing the convolution features.

Paper: Kaiming He, Georgia Gkioxari, Piotr Dollar and Ross Girshick. "MaskRCNN"

Dataset

• COCO2017 is a popular dataset with bounding-box and pixel-level stuff annotations. These annotations can be used for scene understanding tasks like semantic segmentation, object detection and image captioning. There are 118K/5K images for train/val.

• 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

• Hardware（Ascend）

  • Prepare hardware environment with Ascend 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 Python API

• third-party libraries

pip install Cython
pip install pycocotools
pip install mmcv=0.2.14

Quick Start

1. Download the dataset COCO2017.

2. Change the COCO_ROOT and other settings you need in config.py. The directory structure should look like the follows:

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

   If you use your own dataset to train the network, Select dataset to other when run script.
   Create a txt file to store dataset information organized in the way as shown as following:

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

   Each row is an image annotation split by spaces. The first column is a relative path of image, followed by columns containing box and class information in 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), which can be set in config.py.

3. Execute train script.
   After dataset preparation, you can start training as follows:

   # distributed training
   sh run_distribute_train.sh [RANK_TABLE_FILE] [PRETRAINED_CKPT]
   
   # standalone training
   sh run_standalone_train.sh [PRETRAINED_CKPT]
   

   Note:

   1. To speed up data preprocessing, MindSpore provide a data format named MindRecord, hence the first step is to generate MindRecord files based on COCO2017 dataset before training. The process of converting raw COCO2017 dataset to MindRecord format may take about 4 hours.
   2. For distributed training, a hccl configuration file with JSON format needs to be created in advance.
   3. PRETRAINED_CKPT is a resnet50 checkpoint that trained over ImageNet2012.

4. Execute eval script.
   After training, you can start evaluation as follows:

   # Evaluation
   sh run_eval.sh [VALIDATION_JSON_FILE] [CHECKPOINT_PATH]
   

   Note:

   1. VALIDATION_JSON_FILE is a label json file for evaluation.

Script Description

Script and Sample Code

.
└─MaskRcnn      
  ├─README.md                             # README 
  ├─scripts                               # shell script
    ├─run_standalone_train.sh             # training in standalone mode(1pcs)
    ├─run_distribute_train.sh             # training in parallel mode(8 pcs)
    └─run_eval.sh                         # evaluation
  ├─src
    ├─maskrcnn
      ├─__init__.py
      ├─anchor_generator.py               # generate base bounding box anchors 
      ├─bbox_assign_sample.py             # filter positive and negative bbox for the first stage learning
      ├─bbox_assign_sample_stage2.py      # filter positive and negative bbox for the second stage learning
      ├─mask_rcnn_r50.py                  # main network architecture of maskrcnn
      ├─fpn_neck.py                       # fpn network
      ├─proposal_generator.py             # generate proposals based on feature map
      ├─rcnn_cls.py                       # rcnn bounding box regression branch
      ├─rcnn_mask.py                      # rcnn mask branch
      ├─resnet50.py                       # backbone network
      ├─roi_align.py                      # roi align network
      └─rpn.py                            # reagion proposal network
    ├─config.py                           # network configuration
    ├─dataset.py                          # dataset utils
    ├─lr_schedule.py                      # leanring rate geneatore
    ├─network_define.py                   # network define for maskrcnn   
    └─util.py                             # routine operation
  ├─mindspore_hub_conf.py                 # mindspore hub interface
  ├─eval.py                               # evaluation scripts
  └─train.py                              # training scripts

Script Parameters

Training Script Parameters

# distributed training
Usage: sh run_distribute_train.sh [RANK_TABLE_FILE] [PRETRAINED_MODEL]
 
# standalone training
Usage: sh run_standalone_train.sh [PRETRAINED_MODEL]

Parameters Configuration

"img_width": 1280,          # width of the input images
"img_height": 768,          # height of the input images

# random threshold in data augmentation
"keep_ratio": True,         
"flip_ratio": 0.5,
"photo_ratio": 0.5,
"expand_ratio": 1.0,

"max_instance_count": 128, # max number of bbox for each image
"mask_shape": (28, 28),    # shape of mask in rcnn_mask

# anchor
"feature_shapes": [(192, 320), (96, 160), (48, 80), (24, 40), (12, 20)], # shape of fpn feaure maps
"anchor_scales": [8],                                                    # area of base anchor
"anchor_ratios": [0.5, 1.0, 2.0],                                        # ratio between width of height of base anchors
"anchor_strides": [4, 8, 16, 32, 64],                                    # stride size of each feature map levels
"num_anchors": 3,                                                        # anchor number for each pixel

# resnet
"resnet_block": [3, 4, 6, 3],                                            # block number in each layer
"resnet_in_channels": [64, 256, 512, 1024],                              # in channel size for each layer
"resnet_out_channels": [256, 512, 1024, 2048],                           # out channel size for each layer

# fpn
"fpn_in_channels": [256, 512, 1024, 2048],                               # in channel size for each layer
"fpn_out_channels": 256,                                                 # out channel size for every layer
"fpn_num_outs": 5,                                                       # out feature map size

# rpn
"rpn_in_channels": 256,                                                  # in channel size                     
"rpn_feat_channels": 256,                                                # feature out channel size
"rpn_loss_cls_weight": 1.0,                                              # weight of bbox classification in rpn loss
"rpn_loss_reg_weight": 1.0,                                              # weight of bbox regression in rpn loss
"rpn_cls_out_channels": 1,                                               # classification out channel size 
"rpn_target_means": [0., 0., 0., 0.],                                    # bounding box decode/encode means
"rpn_target_stds": [1.0, 1.0, 1.0, 1.0],                                 # bounding box decode/encode stds

# bbox_assign_sampler
"neg_iou_thr": 0.3,                                                      # negative sample threshold after IOU
"pos_iou_thr": 0.7,                                                      # positive sample threshold after IOU
"min_pos_iou": 0.3,                                                      # minimal positive sample threshold after IOU 
"num_bboxes": 245520,                                                    # total bbox numner
"num_gts": 128,                                                          # total ground truth number
"num_expected_neg": 256,                                                 # negative sample number
"num_expected_pos": 128,                                                 # positive sample number

# proposal
"activate_num_classes": 2,                                               # class number in rpn classification
"use_sigmoid_cls": True,                                                 # whethre use sigmoid as loss function in rpn classification

# roi_alignj
"roi_layer": dict(type='RoIAlign', out_size=7, mask_out_size=14, sample_num=2), # ROIAlign parameters
"roi_align_out_channels": 256,                                                  # ROIAlign out channels size
"roi_align_featmap_strides": [4, 8, 16, 32],                                    # stride size for differnt level of ROIAling feature map 
"roi_align_finest_scale": 56,                                                   # finest scale ofr ROIAlign
"roi_sample_num": 640,                                                          # sample number in ROIAling layer

# bbox_assign_sampler_stage2                                                    # bbox assign sample for the second stage, parameter meaning is similar with bbox_assign_sampler
"neg_iou_thr_stage2": 0.5,
"pos_iou_thr_stage2": 0.5,
"min_pos_iou_stage2": 0.5,
"num_bboxes_stage2": 2000,
"num_expected_pos_stage2": 128,
"num_expected_neg_stage2": 512,
"num_expected_total_stage2": 512,

# rcnn                                                                          # rcnn parameter for the second stage, parameter meaning is similar with fpn
"rcnn_num_layers": 2,                                                           
"rcnn_in_channels": 256,                                                        
"rcnn_fc_out_channels": 1024,
"rcnn_mask_out_channels": 256,
"rcnn_loss_cls_weight": 1,
"rcnn_loss_reg_weight": 1,
"rcnn_loss_mask_fb_weight": 1,
"rcnn_target_means": [0., 0., 0., 0.],
"rcnn_target_stds": [0.1, 0.1, 0.2, 0.2],

# train proposal
"rpn_proposal_nms_across_levels": False,                                      
"rpn_proposal_nms_pre": 2000,                                                  # proposal number before nms in rpn
"rpn_proposal_nms_post": 2000,                                                 # proposal number after nms in rpn
"rpn_proposal_max_num": 2000,                                                  # max proposal number in rpn
"rpn_proposal_nms_thr": 0.7,                                                   # nms threshold for nms in rpn
"rpn_proposal_min_bbox_size": 0,                                               # min size of box in rpn

# test proposal                                                                # part of parameters are similar with train proposal
"rpn_nms_across_levels": False,
"rpn_nms_pre": 1000,                                                          
"rpn_nms_post": 1000,
"rpn_max_num": 1000,
"rpn_nms_thr": 0.7,
"rpn_min_bbox_min_size": 0,
"test_score_thr": 0.05,                                                        # score threshold
"test_iou_thr": 0.5,                                                           # IOU threshold
"test_max_per_img": 100,                                                       # max number of instance
"test_batch_size": 2,                                                          # batch size

"rpn_head_loss_type": "CrossEntropyLoss",                                      # loss type in rpn
"rpn_head_use_sigmoid": True,                                                  # whether use sigmoid or not in rpn
"rpn_head_weight": 1.0,                                                        # rpn head weight in loss
"mask_thr_binary": 0.5,                                                        # mask threshold for in rcnn

# LR
"base_lr": 0.02,                                                               # base learning rate
"base_step": 58633,                                                            # bsae step in lr generator
"total_epoch": 13,                                                             # total epoch in lr generator
"warmup_step": 500,                                                            # warmp up step in lr generator
"warmup_mode": "linear",                                                       # warmp up mode
"warmup_ratio": 1/3.0,                                                         # warpm up ratio
"sgd_momentum": 0.9,                                                           # momentum in optimizer

# train
"batch_size": 2,
"loss_scale": 1,
"momentum": 0.91,
"weight_decay": 1e-4,
"pretrain_epoch_size": 0,                                                      # pretrained epoch size
"epoch_size": 12,                                                              # total epoch size
"save_checkpoint": True,                                                       # whether save checkpoint or not
"save_checkpoint_epochs": 1,                                                   # save checkpoint interval
"keep_checkpoint_max": 12,                                                     # max number of saved checkpoint
"save_checkpoint_path": "./checkpoint",                                        # path of checkpoint

"mindrecord_dir": "/home/maskrcnn/MindRecord_COCO2017_Train",                  # path of mindrecord
"coco_root": "/home/maskrcnn/",                                                # path of coco root dateset
"train_data_type": "train2017",                                                # name of train dataset
"val_data_type": "val2017",                                                    # name of evaluation dataset
"instance_set": "annotations/instances_{}.json",                               # name of annotation
"coco_classes": ('background', 'person', 'bicycle', 'car', 'motorcycle', 'airplane', 'bus',
                 'train', 'truck', 'boat', 'traffic light', 'fire hydrant',
                 'stop sign', 'parking meter', 'bench', 'bird', 'cat', 'dog',
                 'horse', 'sheep', 'cow', 'elephant', 'bear', 'zebra',
                 'giraffe', 'backpack', 'umbrella', 'handbag', 'tie',
                 'suitcase', 'frisbee', 'skis', 'snowboard', 'sports ball',
                 'kite', 'baseball bat', 'baseball glove', 'skateboard',
                 'surfboard', 'tennis racket', 'bottle', 'wine glass', 'cup',
                 'fork', 'knife', 'spoon', 'bowl', 'banana', 'apple',
                 'sandwich', 'orange', 'broccoli', 'carrot', 'hot dog', 'pizza',
                 'donut', 'cake', 'chair', 'couch', 'potted plant', 'bed',
                 'dining table', 'toilet', 'tv', 'laptop', 'mouse', 'remote',
                 'keyboard', 'cell phone', 'microwave', 'oven', 'toaster', 'sink',
                 'refrigerator', 'book', 'clock', 'vase', 'scissors',
                 'teddy bear', 'hair drier', 'toothbrush'),
"num_classes": 81

Training Process

• Set options in config.py, including loss_scale, learning rate and network hyperparameters. Click here for more information about dataset.

Training

• Run run_standalone_train.sh for non-distributed training of MaskRCNN model.

# standalone training
sh run_standalone_train.sh [PRETRAINED_MODEL]

Distributed Training

• Run run_distribute_train.sh for distributed training of Mask model.

sh run_distribute_train.sh [RANK_TABLE_FILE] [PRETRAINED_MODEL]

hccl.json which is specified by RANK_TABLE_FILE is needed when you are running a distribute task. You can generate it by using the hccl_tools.
As for PRETRAINED_MODEL, if not set, the model will be trained from the very beginning. Ready-made pretrained_models are not available now. Stay tuned.
This is processor cores binding operation regarding the device_num and total processor numbers. If you are not expect to do it, remove the operations taskset in scripts/run_distribute_train.sh

Training Result

Training result will be stored in the example path, whose folder name begins with "train" or "train_parallel". You can find checkpoint file together with result like the followings in loss_rankid.log.

# distribute training result(8p)
epoch: 1 step: 7393 ,rpn_loss: 0.10626, rcnn_loss: 0.81592, rpn_cls_loss: 0.05862, rpn_reg_loss: 0.04761, rcnn_cls_loss: 0.32642, rcnn_reg_loss: 0.15503, rcnn_mask_loss: 0.33447, total_loss: 0.92218
epoch: 2 step: 7393 ,rpn_loss: 0.00911, rcnn_loss: 0.34082, rpn_cls_loss: 0.00341, rpn_reg_loss: 0.00571, rcnn_cls_loss: 0.07440, rcnn_reg_loss: 0.05872, rcnn_mask_loss: 0.20764, total_loss: 0.34993
epoch: 3 step: 7393 ,rpn_loss: 0.02087, rcnn_loss: 0.98633, rpn_cls_loss: 0.00665, rpn_reg_loss: 0.01422, rcnn_cls_loss: 0.35913, rcnn_reg_loss: 0.21375, rcnn_mask_loss: 0.41382, total_loss: 1.00720
...
epoch: 10 step: 7393 ,rpn_loss: 0.02122, rcnn_loss: 0.55176, rpn_cls_loss: 0.00620, rpn_reg_loss: 0.01503, rcnn_cls_loss: 0.12708, rcnn_reg_loss: 0.10254, rcnn_mask_loss: 0.32227, total_loss: 0.57298
epoch: 11 step: 7393 ,rpn_loss: 0.03772, rcnn_loss: 0.60791, rpn_cls_loss: 0.03058, rpn_reg_loss: 0.00713, rcnn_cls_loss: 0.23987, rcnn_reg_loss: 0.11743, rcnn_mask_loss: 0.25049, total_loss: 0.64563
epoch: 12 step: 7393 ,rpn_loss: 0.06482, rcnn_loss: 0.47681, rpn_cls_loss: 0.04770, rpn_reg_loss: 0.01709, rcnn_cls_loss: 0.16492, rcnn_reg_loss: 0.04990, rcnn_mask_loss: 0.26196, total_loss: 0.54163

Evaluation Process

Evaluation

• Run run_eval.sh for evaluation.

# infer
sh run_eval.sh [VALIDATION_ANN_FILE_JSON] [CHECKPOINT_PATH]

As for the COCO2017 dataset, VALIDATION_ANN_FILE_JSON is refer to the annotations/instances_val2017.json in the dataset directory.
checkpoint can be produced and saved in training process, whose folder name begins with "train/checkpoint" or "train_parallel*/checkpoint".

Evaluation result

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

Evaluate annotation type *bbox*
Accumulating evaluation results...
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.376
 Average Precision  (AP) @[ IoU=0.50      | area=   all | maxDets=100 ] = 0.598
 Average Precision  (AP) @[ IoU=0.75      | area=   all | maxDets=100 ] = 0.405
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.239
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.414
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.475
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=  1 ] = 0.311
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets= 10 ] = 0.500
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.528
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.371
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.572
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.653

Evaluate annotation type *segm*
Accumulating evaluation results...
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.326
 Average Precision  (AP) @[ IoU=0.50      | area=   all | maxDets=100 ] = 0.553
 Average Precision  (AP) @[ IoU=0.75      | area=   all | maxDets=100 ] = 0.344
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.169
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.356
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.462
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=  1 ] = 0.278
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets= 10 ] = 0.426
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.445
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.294
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.484
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.558

Model Description

Performance

Training Performance

Parameters	MaskRCNN
Model Version	V1
Resource	Ascend 910; CPU 2.60GHz, 192cores; Memory, 755G
uploaded Date	08/01/2020 (month/day/year)
MindSpore Version	0.6.0-alpha
Dataset	COCO2017
Training Parameters	epoch=12, batch_size = 2
Optimizer	SGD
Loss Function	Softmax Cross Entropy ,Sigmoid Cross Entropy,SmoothL1Loss
Speed	1pc: 250 ms/step; 8pcs: 260 ms/step
Total time	1pc: 52 hours; 8pcs: 6.6 hours
Parameters (M)	280
Scripts	https://gitee.com/mindspore/mindspore/tree/master/model_zoo/official/cv/maskrcnn

Evaluation Performance

Parameters	MaskRCNN
Model Version	V1
Resource	Ascend 910
Uploaded Date	08/01/2020 (month/day/year)
MindSpore Version	0.6.0-alpha
Dataset	COCO2017
batch_size	2
outputs	mAP
Accuracy	IoU=0.50:0.95 32.4%
Model for inference	254M (.ckpt file)

Description of Random Situation

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

ModelZoo Homepage

Please check the official homepage.