如何使用 tf-keras-vis 正确实现损失 function？

Question

I am trying to understand the following snippet of code demonstrated in this example of tf-keras-vis use.

I am trying to create saliency maps for a project I have been working on. The snippet I am having issues with is as follows:

Define necessary functions

Define Loss functions

You MUST define loss function that return target scores. Here, it returns the scores corresponding Goldfish, Bear, Assault Rifle.

# The 'output' variable refer to the output of the model,
# so, in this case, `output` shape is `(3, 1000)` i.e., (samples, classes).
def loss(output):
    # 1 is the imagenet index corresponding to Goldfish, 294 to Bear and 413 to Assault Rifle.
    return (output[0][1], output[1][294], output[2][413])

How is this exactly calculating loss? I need to understand this so I can modify it to apply it to my own model. If I simply type:

def loss(output):
    return output

Something resembling a saliency map is output but I need to understand what is going on with the original code piece

Answer 1

From how I understand the code in the linked example, the "loss" function simply returns the.network outputs that you're interested in analyzing.

In the sample code:

def loss(output):
    # 1 is the imagenet index corresponding to Goldfish, 294 to Bear and 413 to Assault Rifle.
    return (output[0][1], output[1][294], output[2][413])

the "loss" is set up to generate three saliency images, one per input image (notethat it could be more). For the first one, since the image is a goldfish image, you return the output matching the goldfish label ( output[0][1] , where image 0 is the goldfish one and label 1 is goldfish) and similarly you return the outputs of class "Bear" and "Assault Rifle" for the other two sample images.

So, provided that you know what labels the sample images you're feeding have, just return the model's output for those labels.