About the batchsize and epoch

[YSN1011]

Hello, thank you for your work.Your code only has places for iteration, but not for batchsize and epoch.So how are these two parameters set? Thank you very much.

[Foruck]

For each iteration, we fed one image to the model, along with 15 augmented GT pairs and at most 60 Neg pairs (if there were enough). We didn't use epoch to measure the training process.

[YSN1011]

But in the code, the shape of fc7_H is (16, 1024)，and the sp is (46, 1024).So I want to ask what does 16 and 46 mean?

[Foruck]

For TIN_HICO.py, the shapes of fc7_H and sp mentioned ((16, 1024) and (46, 1024)) were probably wrong, since fc7_H should have a size of (num_pos_neg, 2048) and sp should have a size of (num_pos_neg, 5408), while fc7_O should have a size of (num_pos, 2048). So I assume that you probably mean the shape of fc7_O is (16, 2048), and the shape of sp is (46, 5408). Here, 16 means 16 augmented GT pairs are attached to the trained image, and 46 means there are 46 pairs attached to the image at all, including 16 GT pairs and 30 Neg pairs.

[YSN1011]

Thank you for your answer.I made a mistake because I changed the dimensions of both to 1024.But (?, 2048), (?,5408 )is the shape of the tensor that I'm going to put out directly, and then I'm going to put it into numpy form, and the outputs are (16,2048) and (46,5408).

[YSN1011]

GT pair is given directly by the label, right? So how does Neg pair come about?

[Foruck]

Yep. And if you output the shape in another iteration, you will find their shapes vary.

[Foruck]

GT pair is given directly by the label, right? So how does Neg pair come about?

You could refer to #25 and #30 for this question.

[YSN1011]

Thank you very much. Your answer helped me a lot