I’m still confuse on how the model can predict encodings of images on test time that only requires a single image while to train the model you require a batch of triplets and not just a single image. When I tried using the original training architecture for outputting the encoding of a single image it gave me an input error. Am I suppose to save the trained parameters and transfer it to another model with only one input image (since the parameters of the other three different models are identical) to output the encoding ?
Please explain your question.
At test time, you will generate the values for f(test_image) = f_test vector using the trained network, then calculate similarity function d(f1, f2) between f_test vector and vector for each employee in the database. Whichever employee generates the smallest value for this function, is the most similar employee as the test_image. That would be your prediction.