protoseg

Prototyped Segmentation

Install

Virtual environment

virtualenv -p python3 venv
source venv/bin/activate
pip install -r requirements.txt
python3 train.py configs/all.yml

Pip

The following package should be installed

pip install git+https://github.com/chriamue/pytorch-semseg
pip install gluoncv,mxboard, tensorboardx

then protoseg can be installed

pip install git+https://github.com/chriamue/protoseg

Workflow

Data

Images should be copied to ./data folder. Images for training to train folder, the masks to train_masks folder. Images for validation to val folder, the masks to val_masks folder. Images for testing into test folder.

Config

Every run is stored in a config file like: config

Backends

There should exist multiple backends.

gluoncv
pytorch-semseg

gluoncv:
  backend: gluoncv_backend
  backbone: resnet50

ptsemseg_unet:
  backend: ptsemseg_backend
  backbone: unet
  classes: 2
  width: 572
  height: 572
  mask_width: 388
  mask_height: 388
  orig_width: 768
  orig_height: 768
  gray_img: True

ptsemseg_segnet:
  backend: ptsemseg_backend
  backbone: segnet
  classes: 2
  width: 512
  height: 512

Augmentation

There are several options for augmentation. You can find a full list of augmentators in the docs of imgaug. Each augmentator will be executed with probability 0.5 .

Shape

Shape augmentation will be executed on image and mask.

  shape_augmentation:
    - Affine:
        rotate: -15
    - Affine:
        rotate: 30
    - Affine:
        scale: [0.8, 2.5]
    - Fliplr:
        p: 1.0

Image

Image augmentation will be executed only on the image.

  img_augmentation:
    - GaussianBlur:
        sigma: [0, 0.5]
    - AdditiveGaussianNoise:
        scale: 50

Filters

Filters are functions to preprocess images. The functions have to consume an image first and then some parameters and have to return the processed image. Parameters can be given as list or as dictionary.

  filters:
    - 'cv2.Canny': [100,200]
    - 'protoseg.filters.morphological.erosion':
        kernelw: 5
        kernelh: 5
        iterations: 1

Metrices

There are some metrices available to measure on validation data.

  metrices:
    - 'pixel_accuracy': 'protoseg.metrices.accuracy.pixel_accuracy'
    - 'mean_accuracy': 'protoseg.metrices.accuracy.mean_accuracy'
    - 'dice': 'protoseg.metrices.dice.dice'
    - 'jaccard': 'protoseg.metrices.jaccard.jaccard'
    - 'iou': 'protoseg.metrices.iou.iou'

Report

A report as PDF file of the results can be created.

Running

python3 train.py configs/ptsemseg_segnet.yml

produces following images and pdf report file after training.

Hyperparameteropt

Hyperparameter trains multiple times with multiple configurations and tries to find

hyperparameters which produce the best loss.

Learnrate will be variated by default, but every other parameter in the config file can

be variated too.

An example configuration looks like:

ptsemseg_segnet:
  flit: False
  filters:
    - 'cv2.Canny': [100,200]
  hyperparamopt:
    - flip: [True, False]
    - filters: [['protoseg.filters.canny.addcanny': [100,200]],['protoseg.filters.morphological.opening': [5,5,1]]]

A report will be generated as pdf file.

Kaggle Competition Data

In folder ./scripts is ultrasound-nerve-segmentation.py which should be run as

python3 ./scripts/ultrasound-nerve-segmentation.py /path/to/competition-data data/

The script extracts competition images and copies them to the data folder.

Usefull links

A 2017 Guide to Semantic Segmentation with Deep Learning Satellite Image Segmentation: a Workflow with U-Net

enyalien / protoseg