caffe visualization with t-SNE

[Mahfujur Rahman]

I am trying to implement this code. I extracted fc7 features. But I am not sure how to get 4096-dimensional CNN codes for the 50,000 images (as .mat file). Can you help in this regards?

http://cs.stanford.edu/people/karpathy/cnnembed/

[Khanh Chuong Le]

Hi there,

AlexNet's fc7 layer has 4096 features (=CNN codes) so all you have to do is to forward propagate all the 50,000 images and for each image store it's CNN code. You will basically get a matrix of the type (50000x4096) and feed this matrix to t-SNE.

If you use other network networks like ResNet then CNN-code is probably conv5_3 with 2048 dimensions (maybe you need to flatten the tensor 1x1x2048 to a vector 2048). And again for all those 50,000 images forward propagate and store their CNN-codes. You will get a matrix of the shape (50000x2048).

Regards

ML

[Mahfujur Rahman]

I extracted features using the following command:


./build/tools/extract_features.bin  /models/bvlc_reference_caffenet/bvlc_reference_caffenet.caffemodel examples/_temp/imagenet_val.prototxt fc7  examples/_temp/features 10 leveldb


Am I right ?

[Mahfujur Rahman]

In this code they provided one .mat file. How can I create that mat file? The mat file is attached.