Closed jamesrobertwilliams closed 1 year ago
you should modify kernel size of gdc layer to 14x14. https://github.com/cavalleria/cavaface/blob/822651f0e6d4d08df5441922acead39dc5375103/backbone/mobilefacenet.py#L190
@cavalleria thank you EVER so much for your fast reply and your awesome pkg. I did think of doubling the kernel size but I was not sure about the implications to GDC of using an even sized kernel and also a large kenel. Anyways, it seems to train now :) Thank you again.
@cavalleria
firstly thank you for this awesome repo. I would like to train mobilefacenet with an input size of 224x224 and not 112x112.
Thus, I would like to modify the source here: https://github.com/cavalleria/cavaface/blob/master/backbone/mobilefacenet.py
Am i correct in saying that all I need to do is modify the Linear layer here: https://github.com/cavalleria/cavaface/blob/822651f0e6d4d08df5441922acead39dc5375103/backbone/mobilefacenet.py#L193
to adjust for the increased size of the ouput from
conv_6_dw
?Thank you!