ValueError: math domain error

[biesseck]

I'm trying to train YOLO on my own dataset, but in the 1st epoch, after 14 batches, it raises the following exception:

Traceback (most recent call last):
.
.
.
File "pytorch_custom_yolo_training/utils/utils.py", line 239, in build_targets
    tw[b, best_n, gj, gi] = math.log(gw / anchors[best_n][0] + 1e-16)
ValueError: math domain error

My images have the size 98x81, which is smaller than the default size (width=416, height=416) set in the .cfg file. Could it be a cause of the problem? Or is that related to an invalid pixel value of my images? I thought the image size wasn't the problem because it runs some batches normally before raising the error, so I believe the script resizes the input images to the default size or do some padding to fit it into the input layer. Is it right?

I've also tried reduce batch_size from 16 to 8, but got the same error. Any idea?