mindspore/model_zoo/official/cv/deeplabv3

• Contents
• DeepLabV3 Description
• • Description
• Model Architecture
• Dataset
• Features
• • Mixed Precision
• Environment Requirements
• Quick Start
• Script Description
• • Script and Sample Code
  • Script Parameters
  • Training Process
  • • Usage
    • • Running on Ascend
    • Result
  • Evaluation Process
  • • Usage
    • • Running on Ascend
    • Result
    • • Training accuracy
• Model Description
• • [Performance](#contents
  • • Evaluation Performance
  • Inference Performance
• Description of Random Situation
• ModelZoo Homepage

Contents

• DeepLabV3 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
    • Evaluation Performance
• Description of Random Situation
• ModelZoo Homepage

DeepLabV3 Description

Description

DeepLab is a series of image semantic segmentation models, DeepLabV3 improves significantly over previous versions. Two keypoints of DeepLabV3: Its multi-grid atrous convolution makes it better to deal with segmenting objects at multiple scales, and augmented ASPP makes image-level features available to capture long range information.
This repository provides a script and recipe to DeepLabV3 model and achieve state-of-the-art performance.

Refer to this paper for network details.
Chen L C, Papandreou G, Schroff F, et al. Rethinking atrous convolution for semantic image segmentation[J]. arXiv preprint arXiv:1706.05587, 2017.

Model Architecture

Resnet101 as backbone, atrous convolution for dense feature extraction.

Dataset

Pascal VOC datasets and Semantic Boundaries Dataset

• Download segmentation dataset.

• Prepare the training data list file. The list file saves the relative path to image and annotation pairs. Lines are like:

JPEGImages/00001.jpg SegmentationClassGray/00001.png
JPEGImages/00002.jpg SegmentationClassGray/00002.png
JPEGImages/00003.jpg SegmentationClassGray/00003.png
JPEGImages/00004.jpg SegmentationClassGray/00004.png
......

• Configure and run build_data.sh to convert dataset to mindrecords. Arguments in scripts/build_data.sh:

--data_root                 root path of training data
--data_lst                  list of training data(prepared above)
--dst_path                  where mindrecords are saved
--num_shards                number of shards of the mindrecords
--shuffle                   shuffle or not

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 types, 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）

• Prepare hardware environment with Ascend. 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

• Install python packages in requirements.txt

• Generate config json file for 8pcs training

  # From the root of this project
  cd src/tools/
  python3 get_multicards_json.py 10.111.*.*
  # 10.111.*.* is the computer's ip address.
  

Quick Start

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

• Running on Ascend

Based on original DeepLabV3 paper, we reproduce two training experiments on vocaug (also as trainaug) dataset and evaluate on voc val dataset.

For single device training, please config parameters, training script is:

run_standalone_train.sh

For 8 devices training, training steps are as follows:

1. Train s16 with vocaug dataset, finetuning from resnet101 pretrained model, script is:

run_distribute_train_s16_r1.sh

2. Train s8 with vocaug dataset, finetuning from model in previous step, training script is:

run_distribute_train_s8_r1.sh

3. Train s8 with voctrain dataset, finetuning from model in pervious step, training script is:

run_distribute_train_s8_r2.sh

For evaluation, evaluating steps are as follows:

1. Eval s16 with voc val dataset, eval script is:

run_eval_s16.sh

2. Eval s8 with voc val dataset, eval script is:

run_eval_s8.sh

3. Eval s8 multiscale with voc val dataset, eval script is:

run_eval_s8_multiscale.sh

4. Eval s8 multiscale and flip with voc val dataset, eval script is:

run_eval_s8_multiscale_flip.sh

Script Description

Script and Sample Code

.
└──deeplabv3
  ├── README.md
  ├── scripts
    ├── build_data.sh                             # convert raw data to mindrecord dataset
    ├── run_distribute_train_s16_r1.sh            # launch ascend distributed training(8 pcs) with vocaug dataset in s16 structure
    ├── run_distribute_train_s8_r1.sh             # launch ascend distributed training(8 pcs) with vocaug dataset in s8 structure
    ├── run_distribute_train_s8_r2.sh             # launch ascend distributed training(8 pcs) with voctrain dataset in s8 structure
    ├── run_eval_s16.sh                           # launch ascend evaluation in s16 structure
    ├── run_eval_s8.sh                            # launch ascend evaluation in s8 structure
    ├── run_eval_s8_multiscale.sh                 # launch ascend evaluation with multiscale in s8 structure
    ├── run_eval_s8_multiscale_filp.sh            # launch ascend evaluation with multiscale and filp in s8 structure
    ├── run_standalone_train.sh                   # launch ascend standalone training(1 pc)
  ├── src
    ├── data
        ├── dataset.py                            # mindrecord data generator
        ├── build_seg_data.py                     # data preprocessing
    ├── loss
       ├── loss.py                                # loss definition for deeplabv3
    ├── nets
       ├── deeplab_v3
          ├── deeplab_v3.py                       # DeepLabV3 network structure
       ├── net_factory.py                         # set S16 and S8 structures
    ├── tools
       ├── get_multicards_json.py                 # get rank table file
    └── utils
       └── learning_rates.py                      # generate learning rate
  ├── eval.py                                     # eval net
  ├── train.py                                    # train net
  └── requirements.txt                            # requirements file

Script Parameters

Default configuration

"data_file":"/PATH/TO/MINDRECORD_NAME"            # dataset path
"train_epochs":300                                # total epochs
"batch_size":32                                   # batch size of input tensor
"crop_size":513                                   # crop size
"base_lr":0.08                                    # initial learning rate
"lr_type":cos                                     # decay mode for generating learning rate
"min_scale":0.5                                   # minimum scale of data argumentation
"max_scale":2.0                                   # maximum scale of data argumentation
"ignore_label":255                                # ignore label
"num_classes":21                                  # number of classes
"model":deeplab_v3_s16                            # select model
"ckpt_pre_trained":"/PATH/TO/PRETRAIN_MODEL"      # path to load pretrain checkpoint
"is_distributed":                                 # distributed training, it will be True if the parameter is set
"save_steps":410                                  # steps interval for saving
"keep_checkpoint_max":200                         # max checkpoint for saving

Training Process

Usage

Running on Ascend

Based on original DeepLabV3 paper, we reproduce two training experiments on vocaug (also as trainaug) dataset and evaluate on voc val dataset.

For single device training, please config parameters, training script is as follows:

# run_standalone_train.sh
python ${train_code_path}/train.py --data_file=/PATH/TO/MINDRECORD_NAME  \
                    --train_dir=${train_path}/ckpt  \
                    --train_epochs=200  \
                    --batch_size=32  \
                    --crop_size=513  \
                    --base_lr=0.015  \
                    --lr_type=cos  \
                    --min_scale=0.5  \
                    --max_scale=2.0  \
                    --ignore_label=255  \
                    --num_classes=21  \
                    --model=deeplab_v3_s16  \
                    --ckpt_pre_trained=/PATH/TO/PRETRAIN_MODEL  \
                    --save_steps=1500  \
                    --keep_checkpoint_max=200 >log 2>&1 &

For 8 devices training, training steps are as follows:

1. Train s16 with vocaug dataset, finetuning from resnet101 pretrained model, script is as follows:

# run_distribute_train_s16_r1.sh
for((i=0;i<=$RANK_SIZE-1;i++));
do
    export RANK_ID=${i}
    export DEVICE_ID=$((i + RANK_START_ID))
    echo 'start rank='${i}', device id='${DEVICE_ID}'...'
    mkdir ${train_path}/device${DEVICE_ID}
    cd ${train_path}/device${DEVICE_ID} || exit
    python ${train_code_path}/train.py --train_dir=${train_path}/ckpt  \
                                               --data_file=/PATH/TO/MINDRECORD_NAME  \
                                               --train_epochs=300  \
                                               --batch_size=32  \
                                               --crop_size=513  \
                                               --base_lr=0.08  \
                                               --lr_type=cos  \
                                               --min_scale=0.5  \
                                               --max_scale=2.0  \
                                               --ignore_label=255  \
                                               --num_classes=21  \
                                               --model=deeplab_v3_s16  \
                                               --ckpt_pre_trained=/PATH/TO/PRETRAIN_MODEL  \
                                               --is_distributed  \
                                               --save_steps=410  \
                                               --keep_checkpoint_max=200 >log 2>&1 &
done

2. Train s8 with vocaug dataset, finetuning from model in previous step, training script is as follows:

# run_distribute_train_s8_r1.sh
for((i=0;i<=$RANK_SIZE-1;i++));
do
    export RANK_ID=${i}
    export DEVICE_ID=$((i + RANK_START_ID))
    echo 'start rank='${i}', device id='${DEVICE_ID}'...'
    mkdir ${train_path}/device${DEVICE_ID}
    cd ${train_path}/device${DEVICE_ID} || exit
    python ${train_code_path}/train.py --train_dir=${train_path}/ckpt  \
                                               --data_file=/PATH/TO/MINDRECORD_NAME  \
                                               --train_epochs=800  \
                                               --batch_size=16  \
                                               --crop_size=513  \
                                               --base_lr=0.02  \
                                               --lr_type=cos  \
                                               --min_scale=0.5  \
                                               --max_scale=2.0  \
                                               --ignore_label=255  \
                                               --num_classes=21  \
                                               --model=deeplab_v3_s8  \
                                               --loss_scale=2048  \
                                               --ckpt_pre_trained=/PATH/TO/PRETRAIN_MODEL  \
                                               --is_distributed  \
                                               --save_steps=820  \
                                               --keep_checkpoint_max=200 >log 2>&1 &
done

3. Train s8 with voctrain dataset, finetuning from model in pervious step, training script is as follows:

# run_distribute_train_s8_r2.sh
for((i=0;i<=$RANK_SIZE-1;i++));
do
    export RANK_ID=${i}
    export DEVICE_ID=$((i + RANK_START_ID))
    echo 'start rank='${i}', device id='${DEVICE_ID}'...'
    mkdir ${train_path}/device${DEVICE_ID}
    cd ${train_path}/device${DEVICE_ID} || exit
    python ${train_code_path}/train.py --train_dir=${train_path}/ckpt  \
                                               --data_file=/PATH/TO/MINDRECORD_NAME  \
                                               --train_epochs=300  \
                                               --batch_size=16  \
                                               --crop_size=513  \
                                               --base_lr=0.008  \
                                               --lr_type=cos  \
                                               --min_scale=0.5  \
                                               --max_scale=2.0  \
                                               --ignore_label=255  \
                                               --num_classes=21  \
                                               --model=deeplab_v3_s8  \
                                               --loss_scale=2048  \
                                               --ckpt_pre_trained=/PATH/TO/PRETRAIN_MODEL  \
                                               --is_distributed  \
                                               --save_steps=110  \
                                               --keep_checkpoint_max=200 >log 2>&1 &
done

Result

• Training vocaug in s16 structure

# distribute training result(8p)
epoch: 1 step: 41, loss is 0.8319108
epoch time: 213856.477 ms, per step time: 5216.012 ms
epoch: 2 step: 41, loss is 0.46052963
epoch time: 21233.183 ms, per step time: 517.883 ms
epoch: 3 step: 41, loss is 0.45012417
epoch time: 21231.951 ms, per step time: 517.852 ms
epoch: 4 step: 41, loss is 0.30687785
epoch time: 21199.911 ms, per step time: 517.071 ms
epoch: 5 step: 41, loss is 0.22769661
epoch time: 21240.281 ms, per step time: 518.056 ms
epoch: 6 step: 41, loss is 0.25470978
...

• Training vocaug in s8 structure

# distribute training result(8p)
epoch: 1 step: 82, loss is 0.024167
epoch time: 322663.456 ms, per step time: 3934.920 ms
epoch: 2 step: 82, loss is 0.019832281
epoch time: 43107.238 ms, per step time: 525.698 ms
epoch: 3 step: 82, loss is 0.021008959
epoch time: 43109.519 ms, per step time: 525.726 ms
epoch: 4 step: 82, loss is 0.01912349
epoch time: 43177.287 ms, per step time: 526.552 ms
epoch: 5 step: 82, loss is 0.022886964
epoch time: 43095.915 ms, per step time: 525.560 ms
epoch: 6 step: 82, loss is 0.018708453
epoch time: 43107.458 ms per step time: 525.701 ms
...

• Training voctrain in s8 structure

# distribute training result(8p)
epoch: 1 step: 11, loss is 0.00554624
epoch time: 199412.913 ms, per step time: 18128.447 ms
epoch: 2 step: 11, loss is 0.007181881
epoch time: 6119.375 ms, per step time: 556.307 ms
epoch: 3 step: 11, loss is 0.004980865
epoch time: 5996.978 ms, per step time: 545.180 ms
epoch: 4 step: 11, loss is 0.0047651967
epoch time: 5987.412 ms, per step time: 544.310 ms
epoch: 5 step: 11, loss is 0.006262637
epoch time: 5956.682 ms, per step time: 541.517 ms
epoch: 6 step: 11, loss is 0.0060750707
epoch time: 5962.164 ms, per step time: 542.015 ms
...

Evaluation Process

Usage

Running on Ascend

Configure checkpoint with --ckpt_path and dataset path. Then run script, mIOU will be printed in eval_path/eval_log.

./run_eval_s16.sh                     # test s16
./run_eval_s8.sh                      # test s8
./run_eval_s8_multiscale.sh           # test s8 + multiscale
./run_eval_s8_multiscale_flip.sh      # test s8 + multiscale + flip

Example of test script is as follows:

python ${train_code_path}/eval.py --data_root=/PATH/TO/DATA  \
                    --data_lst=/PATH/TO/DATA_lst.txt  \
                    --batch_size=16  \
                    --crop_size=513  \
                    --ignore_label=255  \
                    --num_classes=21  \
                    --model=deeplab_v3_s8  \
                    --scales=0.5  \
                    --scales=0.75  \
                    --scales=1.0  \
                    --scales=1.25  \
                    --scales=1.75  \
                    --flip  \
                    --freeze_bn  \
                    --ckpt_path=/PATH/TO/PRETRAIN_MODEL >${eval_path}/eval_log 2>&1 &

Result

Our result were obtained by running the applicable training script. To achieve the same results, follow the steps in the Quick Start Guide.

Training accuracy

Network	OS=16	OS=8	MS	Flip	mIOU	mIOU in paper
deeplab_v3	√				77.37	77.21
deeplab_v3		√			78.84	78.51
deeplab_v3		√	√		79.70	79.45
deeplab_v3		√	√	√	79.89	79.77

Note: There OS is output stride, and MS is multiscale.

Model Description

[Performance](#contents

Evaluation Performance

Parameters	Ascend 910
Model Version	DeepLabV3
Resource	Ascend 910
Uploaded Date	09/04/2020 (month/day/year)
MindSpore Version	0.7.0-alpha
Dataset	PASCAL VOC2012 + SBD
Training Parameters	epoch = 300, batch_size = 32 (s16_r1) epoch = 800, batch_size = 16 (s8_r1) epoch = 300, batch_size = 16 (s8_r2)
Optimizer	Momentum
Loss Function	Softmax Cross Entropy
Outputs	probability
Loss	0.0065883575
Speed	60 ms/step（1pc, s16） 480 ms/step（8pcs, s16） 244 ms/step (8pcs, s8)
Total time	8pcs: 706 mins
Parameters (M)	58.2
Checkpoint for Fine tuning	443M (.ckpt file)
Model for inference	223M (.air file)
Scripts	Link

Inference Performance

Parameters	Ascend
Model Version	DeepLabV3 V1
Resource	Ascend 910
Uploaded Date	09/04/2020 (month/day/year)
MindSpore Version	0.7.0-alpha
Dataset	VOC datasets
batch_size	32 (s16); 16 (s8)
outputs	probability
Accuracy	8pcs: s16: 77.37 s8: 78.84% s8_multiscale: 79.70% s8_Flip: 79.89%
Model for inference	443M (.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.

ModelZoo Homepage

Please check the official homepage.