search

mbhenaff / spectral-lib

45 stars 18 forks source link

Format of the dataset #3

Open Conchylicultor opened 8 years ago

Conchylicultor commented 8 years ago

Hi, I have some trouble understanding the input format from the dataset, more specifically, what does the variables self.train_set and self.test_set should contains. How are the laplacians and signals stored ? From what I understand, the dataset which are loaded from "datasources.lua" correspond to your own modified version of the dataset because I didn't find any of the '.th' files you reference. Without data examples, it's difficult to understand it.

Is it possible to have more information on the dataset format ?

Thank you very much

x2mas commented 8 years ago

"the dataset which are loaded correspond to your own modified version of the dataset" Same problem/question here. The datasets used are loaded from directories on your own pc, could we have an insight of what the datasets look like? If you could upload a part of a commonly used dataset, as used in the experiments you set, that would be great.

Thanks in advance

mbhenaff commented 8 years ago

Hi,

I didn't upload the datasets to git because they are too big, but you can download MNIST and CIFAR-10 using the following scripts: https://github.com/torch/tutorials/tree/master/A_datasets.

The datasets are 3D tensors of the form [nSamples x nChannels x nFeatures]. MNIST has 1 channel, CIFAR has 3 (corresponding to R, G, B channels).

You can look at this script to see how to build the graph Laplacian: https://github.com/mbhenaff/spectral-lib/blob/master/mresgraph/build_laplacian.m

Best,

-Mikael

On Thu, Mar 24, 2016 at 5:38 AM, x2mas notifications@github.com wrote:

"the dataset which are loaded correspond to your own modified version of the dataset" Same problem/question here. The datasets used are loaded from directories on your own pc, could we have an insight of what the datasets look like? If you could upload a part of a commonly used dataset, as used in the experiments you set, that would be great.

Thanks in advance

— You are receiving this because you are subscribed to this thread. Reply to this email directly or view it on GitHub https://github.com/mbhenaff/spectral-lib/issues/3#issuecomment-200756108