Loss_l : inf in training ?

[transcendentsky]

Hello, first of all, thanks for your code releasing.
I got the training loss inf, acutally loss_l = inf, i use your original code (only fixed some bug), but i don't know why i got inf.
Parameters: lr:0.004, batchsize:32, base_model:vgg_reducedfc.pth
GPU: 1080ti

Any comments will be appreciated.
Thanks very much!

[GOATmessi8]

@transcendentsky You may look at this pull for details.

[transcendentsky]

I have tried this way, but the problem still exists.

[GOATmessi8]

@transcendentsky Once you got the inf loss, what is your next batch loss? Does it still inf? I actually also meet some inf loss in training the COCO, but it seems OK for final convergence.

[transcendentsky]

I tracked the code and find the problem comes to prior (prior_box.py) , the code s_k = self.min_sizes[k]/self.image_size get 0 because these are two integers. I compared it with code in https://github.com/amdegroot/ssd.pytorch , and There should have a from __future__ import division on the top and code goes right.
Cause I used py2.7 , in py3 there's no problem maybe.

[GOATmessi8]

@transcendentsky The current code has no been tested in Python2.