DataResize part code question

[wangjiyuan9]

Emm,
At here 41~48
I think inputs[k],inputs[(n + "_aug", im)] and inputs[(n, im)] are same???

[Dwawayu]

Hi! I want to remind that Line 41-48 is for resizing cropping, while Line 71-77 is for resizing only. You can use the --no_crop flag to choose which one to use.

Regarding your question of Line 41-48, inputs[(n + "_aug", im)] and inputs[(n, im)] are the same here, both cropped and resized. inputs[k] is only resized without cropping.

Later on, inputs[(n + "_aug", im)] will be further augmented in the color space, so all of them will be different.

By default, inputs[(n + "_aug", im)] is the input of network and inputs[(n, im)] is used for computing the loss. You can use --match_aug flag to compute loss on inputs[(n + "_aug", im)] if you wish.

I hope this answered your question. Please feel free to contact me if you meet any more questions.

[wangjiyuan9]

My previous understanding was superficial, thank you for your detailed explanation.

[wangjiyuan9]

Hi, I think I have a deep understanding of your code.
Actually ,your codes are robust to hold any condition.
But the inputs[k] really don't necessary. Because in the function add_flip_right_inputs you make the new inputs and never consider the inputs[k].
So your code use another useless resize time to create inputs[k]. (the resize code will done over 100k time in training!)