Any plans for releasing your datasets and feature vectors?

technicolor-research / subic

Tensorflow implementation of a supervised approach to learn highly compressed image representations

Other

27 stars 4 forks source link

Any plans for releasing your datasets and feature vectors? #3

Open msharmavikram opened 6 years ago

msharmavikram commented 6 years ago

As pointed out in readme, I don't see any link for the dataset or feature vectors. Where can I download it from?

himalayajain commented 6 years ago

Hello, Here is a temporary link until we provide a permanent one.

ibadiu commented 5 years ago

Hi, Thanks for sharing the dataset related files. Could you please detail the extraction of the feature files (ImgNetVal_vgg128_x.npy)? Are these extracted from the pretrained VGG-128 network and, if so, is this the fc7 layer? Or is this a 128-dimensions layer obtained after SUBIC training? Also, concerning SUBIC training, do you use a model conversion tool to upload the pretrained VGG-128 Caffe model in TensorFlow? Thanks!

himalayajain commented 5 years ago

Hi,

ImgNetVal_vgg128_x.npy is the uncompressed fc7 feature from VGG-128 network. I think, I used https://github.com/ethereon/caffe-tensorflow to convert. If you are using the code, you might want to use subic_imagenet_m8_k256_be1_me1_lr1e-4_5k.npy provided in the shared folder. It is a trained SuBiC model with layers till fc7 are frozen to the pretrained VGG-128 (so no need to convert).

SikaStar commented 5 years ago

Hi, Where is your ".prototxt" file of your caffemodel?