edafa 0.1

Test Time Augmentation (TTA) wrapper for computer vision tasks: segmentation, super-resolution, ... etc.

Edafa

Edafa is a simple wrapper that implements Test Time Augmentations (TTA) on images for computer vision problems like: segmentation, super-resolution, Pansharpening, etc. TTAs guarantees better results in most of the tasks.

Installation

PyPI pacakage will be avaialble soon

Getting started

The easiest way to get up and running is to follow this notebook showing an example on using Edafa to improve segmentation score on PASCAL VOC dataset.

How to use Edafa

The whole process can be done in 4 steps:

1. Import BasePredictor abstract class

from edafa import BasePredictor

2. Inherit BasePredictor to your own class and implement the main 3 functions
   • preprocess(self,img) : Implement preprocessing needed after reading image from disk (e.g. normalization, .. )
   • postprocess(self,pred) : Implement postprocessing needed after predicting an image (e.g. clipping, argmax, .. )
   • predict_patches(self,patches) : The function where your model takes image patches and return their prediction

class myPredictor(BasePredictor):
    def __init__(self,model,*args,**kwargs):
        super().__init__(*args,**kwargs)
        self.model = model

    def preprocess(self,img):
        return normalize(img)

    def postprocess(self,pred):
        return clipping(pred)

    def predict_patches(self,patches):
        return self.model.predict(patches)

3. Create an instance of you class

p = myPredictor(model,patch_size,model_output_channels,conf_file_path)

4. Call predict_dir() to run the prediction process

p.predict_dir(in_dir,out_dir,overlap=0,extension='.png')

Configuration file

Configuration file is a json file containing two pieces of information

1. Augmentations to apply (augs). Supported augmentations:
   • NO : No augmentation
   • ROT90 : Rotate 90 degrees
   • ROT180 : Rotate 180 degrees
   • ROT270 : Rotate 270 degrees
   • FLIP_UD : Flip upside-down
   • FLIP_LR : Flip left-right
2. Combination of the results (mean). Supported mean types:
   • ARITHMETIC : Arithmetic mean
   • GEOMETRIC : Geometric mean

Example of a conf file

{
"augs":["NO",
"FLIP_UD",
"FLIP_LR"],
"mean":"ARITHMETIC"
}