mindspore/model_zoo/official/cv/unet3d

• Contents
• • Unet Description
  • Model Architecture
  • Dataset
  • Environment Requirements
  • Quick Start
  • Script Description
  • • Script and Sample Code
    • Script Parameters
  • Training Process
  • • Training
    • • running on Ascend
      • Distributed Training
  • Evaluation Process
  • • Evaluation
  • Model Description
  • • Performance
    • • Evaluation Performance
      • Inference Performance
• Description of Random Situation
• • ModelZoo Homepage

Contents

• Contents
  • Unet Description
  • Model Architecture
  • Dataset
  • Environment Requirements
  • Quick Start
  • Script Description
    • Script and Sample Code
    • Script Parameters
  • Training Process
    • Training
      • running on Ascend
      • Distributed Training
  • Evaluation Process
    • Evaluation
  • Model Description
    • Performance
      • Evaluation Performance
      • Inference Performance
  • Description of Random Situation
  • ModelZoo Homepage

Unet Description

Unet3D model is widely used for 3D medical image segmentation. The construct of Unet3D network is similar to the Unet, the main difference is that Unet3D use 3D operations like Conv3D while Unet is anentirely 2D architecture. To know more information about Unet3D network, you can read the original paper Unet3D: Learning Dense VolumetricSegmentation from Sparse Annotation.

Model Architecture

Unet3D model is created based on the previous Unet(2D), which includes an encoder part and a decoder part. The encoder part is used to analyze the whole picture and extract and analyze features, while the decoder part is to generate a segmented block image. In this model, we also add residual block in the base block to improve the network.

Dataset

Dataset used: LUNA16

• Description: The data is to automatically detect the location of nodules from volumetric CT images. 888 CT scans from LIDC-IDRI database are provided. The complete dataset is divided into 10 subsets that should be used for the 10-fold cross-validation. All subsets are available as compressed zip files.

• Dataset size：888

  • Train：878 images
  • Test：10 images

• Data format：zip

  • Note：Data will be processed in convert_nifti.py

Environment Requirements

• Hardware（Ascend）
  • Prepare hardware environment with Ascend processor.
• Framework
  • MindSpore
• For more information, please check the resources below：
  • MindSpore Tutorials
  • MindSpore Python API

Quick Start

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

• Select the network and dataset to use

Convert dataset into mifti format.
python ./src/convert_nifti.py --input_path=/path/to/input_image/ --output_path=/path/to/output_image/

Refer to src/config.py. We support some parameter configurations for quick start.

• Run on Ascend

# run training example
python train.py --data_url=/path/to/data/ --seg_url=/path/to/segment/ > train.log 2>&1 &

# run distributed training example
bash scripts/run_distribute_train.sh [RANK_TABLE_FILE] [IMAGE_PATH] [SEG_PATH]

# run evaluation example
python eval.py --data_url=/path/to/data/ --seg_url=/path/to/segment/ --ckpt_path=/path/to/checkpoint/ > eval.log 2>&1 &

Script Description

Script and Sample Code

.
└─unet3d
  ├── README.md                       // descriptions about Unet3D
  ├── scripts
  │   ├──run_disribute_train.sh       // shell script for distributed on Ascend
  │   ├──run_standalone_train.sh      // shell script for standalone on Ascend
  │   ├──run_standalone_eval.sh       // shell script for evaluation on Ascend
  ├── src
  │   ├──config.py                    // parameter configuration
  │   ├──dataset.py                   // creating dataset
  │   ├──lr_schedule.py               // learning rate scheduler
  │   ├──transform.py                 // handle dataset
  │   ├──convert_nifti.py             // convert dataset
  │   ├──loss.py                      // loss
  │   ├──conv.py                      // conv components
  │   ├──utils.py                     // General components (callback function)
  │   ├──unet3d_model.py              // Unet3D model
  │   ├──unet3d_parts.py              // Unet3D part
  ├── train.py                        // training script
  ├── eval.py                         // evaluation script

Script Parameters

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

• config for Unet3d, luna16 dataset

  'model': 'Unet3d',                  # model name
  'lr': 0.0005,                       # learning rate
  'epochs': 10,                       # total training epochs when run 1p
  'batchsize': 1,                     # training batch size
  "warmup_step": 120,                 # warmp up step in lr generator
  "warmup_ratio": 0.3,                # warpm up ratio
  'num_classes': 4,                   # the number of classes in the dataset
  'in_channels': 1,                   # the number of channels
  'keep_checkpoint_max': 5,           # only keep the last keep_checkpoint_max checkpoint
  'loss_scale': 256.0,                # loss scale
  'roi_size': [224, 224, 96],         # random roi size
  'overlap': 0.25,                    # overlap rate
  'min_val': -500,                    # intersity original range min
  'max_val': 1000,                    # intersity original range max
  'upper_limit': 5                    # upper limit of num_classes
  'lower_limit': 3                    # lower limit of num_classes

Training Process

Training

running on Ascend

python train.py --data_url=/path/to/data/ -seg_url=/path/to/segment/ > train.log 2>&1 &

The python command above will run in the background, you can view the results through the file train.log.

After training, you'll get some checkpoint files under the script folder by default. The loss value will be achieved as follows:

epoch: 1 step: 878, loss is 0.55011123
epoch time: 1443410.353 ms, per step time: 1688.199 ms
epoch: 2 step: 878, loss is 0.58278626
epoch time: 1172136.839 ms, per step time: 1370.920 ms
epoch: 3 step: 878, loss is 0.43625978
epoch time: 1135890.834 ms, per step time: 1328.537 ms
epoch: 4 step: 878, loss is 0.06556784
epoch time: 1180467.795 ms, per step time: 1380.664 ms

Distributed Training

Notes:
RANK_TABLE_FILE can refer to Link , and the device_ip can be got as Link. For large models like InceptionV4, it's better to export an external environment variable export HCCL_CONNECT_TIMEOUT=600 to extend hccl connection checking time from the default 120 seconds to 600 seconds. Otherwise, the connection could be timeout since compiling time increases with the growth of model size.

bash scripts/run_distribute_train.sh [RANK_TABLE_FILE] [IMAGE_PATH] [SEG_PATH]

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

epoch: 1 step: 110, loss is 0.8294426
epoch time: 468891.643 ms, per step time: 4382.165 ms
epoch: 2 step: 110, loss is 0.58278626
epoch time: 165469.201 ms, per step time: 1546.441 ms
epoch: 3 step: 110, loss is 0.43625978
epoch time: 158915.771 ms, per step time: 1485.194 ms
...
epoch: 9 step: 110, loss is 0.016280059
epoch time: 172815.179 ms, per step time: 1615.095 ms
epoch: 10 step: 110, loss is 0.020185348
epoch time: 140476.520 ms, per step time: 1312.865 ms

Evaluation Process

Evaluation

• evaluation on dataset when running on Ascend

Before running the command below, please check the checkpoint path used for evaluation. Please set the checkpoint path to be the absolute full path, e.g., "username/unet3d/Unet3d-10_110.ckpt".

python eval.py --data_url=/path/to/data/ --seg_url=/path/to/segment/ --ckpt_path=/path/to/checkpoint/ > eval.log 2>&1 &

The above python command will run in the background. You can view the results through the file "eval.log". The accuracy of the test dataset will be as follows:

# grep "eval average dice is:" eval.log
eval average dice is 0.9502010010453671

Model Description

Performance

Evaluation Performance

Parameters	Ascend
Model Version	Unet3D
Resource	Ascend 910; CPU 2.60GHz，192cores；Memory，755G
uploaded Date	03/18/2021 (month/day/year)
MindSpore Version	1.2.0
Dataset	LUNA16
Training Parameters	epoch = 10, batch_size = 1
Optimizer	Adam
Loss Function	SoftmaxCrossEntropyWithLogits
Speed	8pcs: 1795ms/step
Total time	8pcs: 0.62hours
Parameters (M)	34
Scripts	unet3d script

Inference Performance

Parameters	Ascend
Model Version	Unet3D
Resource	Ascend 910
Uploaded Date	03/18/2021 (month/day/year)
MindSpore Version	1.2.0
Dataset	LUNA16
batch_size	1
Dice	dice = 0.9502
Model for inference	56M(.ckpt file)

Description of Random Situation

We set seed to 1 in train.py.

ModelZoo Homepage

Please check the official homepage.