Is precompute = 'True' biased towards ILSVCR?


(Raghu) #1

HI All,

I have heard precompute=‘True’ is more biased towards data from ILSVCR dataset. What if i have to train on a new Image data-set like Statelite , Medical Imaging or More complex data sets from Kaggle.

Will precompute=‘True’ is helpful in that case and to what extent?


(Matthijs) #2

Yes, usually the pretrained networks are trained on the ImageNet (ILSVRC) dataset, and so they’ve learned things about detecting photos.

You might still be able to use such a pretrained network on different kinds of images (like medical images), as the network has already learned something, which arguably is better than a random initialization. But you’ll probably have to retrain all the layers – i.e. you can’t simply add a classifier on top of the pretrained layers and only train the classifier.