/VRDL_HW3_instance_segmentation

Primary LanguageJupyter Notebook

VRDL_HW3_instance_segmentation

Objective

  • Instance segmentation task to detect and segment all the nuclei.

  • Data:
    • Nuclear segmentation dataset contains 24 training images with 14,598 nuclear and 6 test images with 2,360 nuclear.
      • train: 20 images, 12188 nuclear
      • val: 3 images, 2410 nuclear
      • test: 6 images, 2360 nuclear

Environment

  • Python 3.7
  • Pytorch 1.6.0
  • CUDA 10.2

Reproducing Submission

To reproduce, do the following steps:

  1. Installation
  2. Data Preparation
  3. Select Config file
  4. Download Pretrained Model
  5. Training
  6. Testing
  7. Results
  8. Report
  9. Reference

Install Packages

  • install pytorch from https://pytorch.org/get-started/previous-versions/

  • install mmdetection

    # quick install:
# install pytorch correspond to your cuda version
conda install pytorch==1.6.0 torchvision==0.7.0 cudatoolkit=10.2 -c pytorch
pip install openmim
mim install mmdet

  • install dependencies

    • scikit-image
    • tqdm
    • apex (optional for swin transformer)

Data Preparation

  • Recommend Way

    1. Download dataset which has been preprocessed. Google Drive. The data should be as follow.
    data
└── coco
    ├── annotations
    │   ├── train_coco.json
    │   ├── val_coco.json
    │   └── test_coco.json
    ├── train2017/train_pic
    │   ├── 1.png
    │   ├── 2.png
    │   └── ...
    ├── val2017/val_pic
    │   ├── 1.png
    │   ├── 2.png
    │   └── ...
    └── test2017/test
        ├── 1.png
        ├── 2.png
        └── ...

  • Or, Prepare Data From Scratch

    1. Download the given dataset from Google Drive.

    2. Split train data into "train" and "val" folder. (python my_data_preprocess/data_process.py)

    3. Run command python my_data_preprocess/data_process.py to create coco format input data.

    • The result will be like:
    - dataset/
  ├── train_coco.json
  ├── val_coco.json
  ├── test_coco.json
  ├── train_pic 
  ├── val_pic
  │   ├── 1.png
  │   └── ...    
  └── test
      ├── 1.png
      └── ...
    1. Reorganize data as Data Preparation

Select Config file

  • cascade mask rcnn: config_cascade_mask_rcnn_50.py
  • swin mask rcnn: config_swinT.py
  • detectoRS: config_detectoRS.py

Download Pretrained Model

Training

  • train model with pretrained model
# cd mmdetection
python tools/train.py config_cascade_mask_rcnn_50.py

Edit "load_from" "data" in config_cascade_mask_rcnn_50.py

Testing

  • Inference test data
python my_inference.py
  • The result will be stored in "./answer.json".
  • Format of "./answer.json" is same as COCO results.

Results

  • Results:
    • val result:
      • bbox_mAP: 0.061
      • segm_mAP: 0.060
    • Test mAP: 0.233557 (41)

Interesting finding

  • Using mmdetection tips:
    1. Pick a config file in ./configs/...
    2. Run it anyway.
    3. Then you can get the complete config file.

Report

https://hackmd.io/@Bmj6Z_QbTMy769jUvLGShA/VRDL_HW3

Reference