What is the structure of torch dataset?

问题:

I am beginning to use torch 7 and I want to make my dataset for classification. I've already made pixel images and corresponding labels. However, I do not know how to feed those data to the torch. I read some codes from others and found out that they are using the dataset whose extension is '.t7' and I think it is a tensor type. Is it right? And I wonder how I can convert my pixel images(actually, I made them with Matlab by using MNIST dataset) into t7 extension compatible to the torch. There must be structure of dataset in the t7 format but I cannot find it (also for the labels too).

To sum up, I have pixel images and labels and want to convert those to t7 format compatible to the torch.

Thanks in advance!

回答1:

The datasets '.t7' are tables of labeled Tensors. For example the following lua code :

if (not paths.filep("cifar10torchsmall.zip")) then
    os.execute('wget -c https://s3.amazonaws.com/torch7/data/cifar10torchsmall.zip')
    os.execute('unzip cifar10torchsmall.zip')
end
Readed_t7 = torch.load('cifar10-train.t7')
print(Readed_t7)

Will return through itorch :

{
  data : ByteTensor - size: 10000x3x32x32
  label : ByteTensor - size: 10000
}

Which means the file contains a table of two ByteTensor one labeled "data" and the other one labeled "label".

To answer your question, you should first read your images (with torchx for example : https://github.com/nicholas-leonard/torchx/blob/master/README.md ) then put them in a table with your Tensor of label. The following code is just a draft to help you out. It considers the case where : there are two classes, all your images are in the same folder and are ordered through those classes.

require 'torchx';

--Read all your dataset (the chosen extension is png)
files = paths.indexdir("/Path/to/your/images/", 'png', true)
data1 = {}
for i=1,files:size() do
   local img1 = image.load(files:filename(i),3)
   table.insert(data1, img1)
end

--Create the table of label according to 
label1 = {}
for i=1, #data1 do
    if i <= number_of_images_of_the_first_class then
        label1[i] = 1
    else
        label1[i] = 2
    end
end

--Reshape the tables to Tensors
label = torch.Tensor(label1)
data = torch.Tensor(#data1,3,16,16)
for i=1, #data1 do
    data[i] = data1[i]
end

--Create the table to save
Data_to_Write = { data = data, label = label }

--Save the table in the /tmp
torch.save("/tmp/Saved_Data.t7", Data_to_Write)

It should be possible to make a less hideous code but this one details all the steps and works with torch 7 and Jupyter 5.0.0 .

Hope it helps.

Regards