ttAugment 0.3.3

Test Time Augmentations

TTAugment

Perform Augmentation during Inference and aggregate the results of all the applied augmentation to create a final output

Installation

pip install git+https://github.com/cypherics/ttAugment.git#egg=ttAugment

Supported Augmentation

Library supports all color, blur and contrast transformation provided by imgaug along with custom Geometric Transformation.

1. Mirror : Crop an image to transform_dimension and mirror pixels to match the size of network_dimension
2. CropScale : Crop an image to transform_dimension and rescale the image to match the size of network_dimension
3. NoAugment : Keep the input unchanged
4. Crop : Crop an image to transform_dimension
5. Rot : Rotate an Image
6. FlipHorizontal
7. FlipVertical

Parameters

How to use when test image is much bigger than what my models needs as input, Don't worry the library has it covered it will generate fragments according to the specified dimension, so you can run inference on the desired dimension, and get the output as per the original test image.

• image_dimension - What is the size of my input image

    image_dimension=(2, 1500, 1500, 3) 
  

• transformer - list of augmentations to perform transform_dimension - specifies what dimension is the network expecting the input to be in, if less than image_dimension, the library will generate smaller fragment of size transform_dimension for inference and apply transformation over it

    transformers=[
    {
        "name": "CLAHE",
        "transform_dimension": (2, 1000, 1000, 3),
    },
    ],
  

  • Dealing with parameters during Scaling transformation, two transformation perform scaling on the test images For Scaling transformation transform_dimension and network_dimension are mandatory parameters

      transformers=[
      {
      "name": "Mirror",
      "transform_dimension": (2, 800, 800, 3),
      "network_dimension": (2, 1000, 1000, 3)
      },
      
      {
      "name": "ScaleCrop",
      "transform_dimension": (2, 800, 800, 3),
      "network_dimension": (2, 1000, 1000, 3)
      }
      ],
    

    The network_dimension parameter informs the library to override the network input and crop the image to transform_dimension and rescale it to network_dimension to get it as per network requirement

    And again the library will merge all the fragments to form the final output of image_dimension

  • For using Rot - Rotate add "param": angle as an additional argument

If the test image has the same dimension to what the network expects, in that case just remove the transform_dimension param.

Inference

Define tta object

tta = Segmentation.populate_color(
        image_dimension=(2, 1500, 1500, 3),
         transformers=transformers) # transfromer as defined in parameters

Calling the generator

Input image is required to be a 4d numpy array of shape (batch, height, width, channels)

for image list(loop over all the images): 
    for transformation in tta.run():
        # Apply forward transfromation
        forward_image = tta.forward(transformation, image=image)

        # Apply normalization
        # Convert input to framework specific type
        # Perform inference
        inferred_image = model.predict(forward_image)

        # make sure to convert the inferred_image to 4d numpy array [batch, height, width, classes]
        reversed_image = tta.reverse(transformation, inferred_image)

        # Add the reversed image to transformation
        tta.update(transformation, reversed_image)
    
    # Access the output
    output = tta.transformations.output