How to use cifar-10 like single label dataset to evaluate this model? about cvpr17-dvsq

[bl0]

Hi, you can use this model in single label just same with the multilabel dataset without any change.

[bl0]

I will close this issue. Feel free to reopen it if u also have any question.

[Retrospection]

Thanks！

[Retrospection]

Excuse me, If I want to make this net to be a image feature extractor, that means, if I want to give an image and take a feature from this network without the whole bunch of config params, what should I do?

[bl0]

What do you mean by "feature"? the 4096-dimensional feature(for fc7) or 64-dimensional feature(for fc8)?
If you are familiar with Tensorflow, simply run the graph and fetch the fc7(or fc8) as the feature(line 580 in net_val.py may help you).

[Retrospection]

yep, the "feature" I mean is just the fc8, but to use the net_val.py, I need to prepare the images in to dataset, make batchs to feed into the network, make whole bunch of config params. If I dont want to prepare such dataset, and dont want to prepare such config, what can I do? For example , If I just want to load the net param from .npy and simply use the skeleton of this network to make an inference to get the fc8, how can I achieve this? Thanks for your kindness

[bl0]

You can delete the useless part and only leave the skeleton. Then you need not prepare config and just need to:

1. load the pre-trained model into the model.
2. read the image from whatever place.
3. run the graph to get the feature.

[Retrospection]

Thanks for your great kindless

[bl0]

不客气😀