These are chat archives for beniz/deepdetect

Jan 2018
Yili Zhao
Jan 08 2018 07:22 UTC
Anyone fine-tuning resnet50 successfully? When fine-tuning resnet50, I still get very low accuracy.
Emmanuel Benazera
Jan 08 2018 08:38 UTC
We finetune it all the time. Check your dataset, try from scratch first. And try on dome well known dataset if you'd like to test your setup first.
Yili Zhao
Jan 08 2018 12:59 UTC
Thanks, and I will follow your suggestion.
Jan 08 2018 15:46 UTC
I've been researching some more on differences in the resize function between OpenCV, skimage, matlab and PIL which seem to account for some divergences between model serve systems. It appears that the PIL resize function did a uint8 conversion during the resize which stripped data away but resulted in some better classification results for my model, while the others have shown some options and differences in the applications of their bi-linear resize with some or no application of anti-aliasing function on their values.
Yili Zhao
Jan 08 2018 22:45 UTC
@beniz If I trained a resnet50 model with caffe directly, how can I use this model in deepdetect? (I still can't train a good model in deepdetect)