Something strange in loss calculation

[Angel-Jia]

Hi,
I am trying to reproduce the siamfc net with pytorch.
I have built the whole net, but I encounter some difficulty in building loss function.
I always get a loss around 0.693 with the formula in paper and it seems that no matter what the score you get, you will get 0.693 in the end. For example:

import numpy as np

score = np.random.randint(-5,5,size=(1,5,5))
label = np.ones((1,5,5))

for i in range(1, 4):
    for j in range(1,4):
        label[0, i, j] = -1

weight = np.ones((1,5,5))
indexP = np.where(label == 1)
indexF = np.where(label == -1)

weight[indexP] = 0.5 / indexP[0].shape[0]
weight[indexF] = 0.5 / indexF[0].shape[0]

label = weight * label
print(np.sum(np.log(1+np.exp(-label*score)))/25)

you will get a value of 0.693. You can try any other score or label, initializing it with random number and you will get 0.693 again.

I wonder is there something not mentioned in paper?

[jvlmdr]

I do not think that this is an issue with this repository.

Without looking at your code, I can tell you that 0.693 is -ln(0.5), which is the loss that you expect to get with random initialization in a balanced 2-class classification problem. Your code seems to be computing the loss for a random score map? This is like averaging over 25 independent classification problems, so it's not that surprising that the loss is roughly constant across samples.