The Inference Wrapper essentially interconnects the inference method with the Inference Quantizer (and optional preprocessing procedure). It is especially useful when your inference model's prediction format is not compatible with the Explanation Model. In this example, a setup where this component proves necessary is presented.
Let's assume we have an inference model which returns segmentation masks for a given dataset. The model assigns each pixel a value between 0 and 255. The higher the value, the more likely it belongs to a certain segment to be classified in the image.
In that case, the model will output a matrix with the same shape as the original image. In order to use the CEM explanation model though, we have to somehow transform the prediction result into a class score vector, typically seen in classification problems. Furthermore, that transformation function has to be continuous and sensitive to changes in the image.
We will consider the BinaryConfidenceMethod as InferenceQuantizer (BinaryConfidenceMethod documentation) which will transform our segmentation mask into a binary classification representation.
The first index of the vector represents the probability score of a class being absent in the image. The second index of the vector represents the probability score of a certain class being present.
- The prediction output
>>> prediction = predict(image)
>>> prediction
[[164 112 225 253 208 156 52 65 199 126]
[211 17 213 122 71 188 102 97 159 3]
[147 190 2 10 164 13 71 40 239 47]
[ 77 42 8 67 57 180 101 40 30 59]
[ 72 188 15 135 34 91 159 208 109 153]
[ 66 237 248 248 225 40 118 85 92 64]
[ 86 56 50 119 62 162 11 247 239 78]
[178 67 127 119 58 183 85 146 245 147]
[ 5 85 144 50 233 73 118 74 191 220]
[ 42 159 51 79 15 70 118 15 135 248]]- After quantization
>>> quantizer = BinaryConfidenceMethod()
>>> quantizer(prediction)
[ 0.09772549 -0.09772549]Now that we have setup the quantizer, we have to connect the entities through the InferenceWrapper. The resulting instance will then produce the desired output format as prediction.
>>> inference_wrapper = maxi.InferenceWrapper(inference_model=predict, quantizer=quantizer)
>>> inference_wrapper(image)
[ 0.09772549 -0.09772549]Eventually, we can parse the explanation method conforming InferenceWrapper to the ExplanationGenerator.run() method.
# 'cem' is an instance of the 'ExplanationGenerator'
>>> result = cem.run(image=image, inference_call=inference_wrapper)