what accuracy to expect fine-tuning AlexNet (bvlc_reference_caffenet)?

[Carlos]

Hello,

This is a question for practiotioners with some hands-on experience training CNNs both from scratch and fine-tuning.

I am planning on fine-tuning AlexNet (bvlc_reference_caffenet) to classify on aprox. 30 classes with about 5-10.000 examples per class.

The fine-tuning guide on Flickr data (http://caffe.berkeleyvision.org/gathered/examples/finetune_flickr_style.html) is a bit dissapointing in that it only gets to an accuracy of 39%.

My question is: is this the accuracy to expect by fine-tuning AlexNet instead of training it from scratch?

As AlexNet was trained with about 1.000 examples per class (1.000 classes), do I have examples enough per class as to train AlexNet from scratch? If not, what would be an approximate number?

Thanks,

Carlos