pytorch: Convert a tuple of FloatTensors into a numpy array

Question

I just got a torchvision.datasets object with MNIST data

train_dataset= dsets.MNIST(root='./data',train=True,transform=transforms.ToTensor(),download=True)

I want to convert this tuple into a set of numpy arrays of shape 60000x28x28 and labels of 60000

I know that the form that the data is provided, can be directly applied into a pytorch neuron in order to be used as training data, but I would like to convert this data into numpy arrays.

the first thing I did was to divide the tuples of (data,labels) with zip(*train_dataset)

data,labels = zip(*train_dataset)

labels is easy to convert into a numpy array, however I have not been able to convert "data" into a numpy array the way I would like. When I try to convert all of the data into numpy.array like

data[:].numpy()

I receive an error telling me 'tuple' object has no attribute 'numpy'. and if I convert this data[:] into numpy, just the first dimension (the 60000 data) is converted into numpy.array but the rest remains as Tensors.

I can convert all of the data using a loop (actually I did it) and it works, however it is really slow. Has anyone done any conversion like this before?

Thanks.

Answer 1

Maybe something like this will work for you:

train_dataset.train_data.numpy()  #contains (60000, 28, 28) numpy array
train_dataset.train_labels.numpy() # contains labels

Answer 2

It seems like there are two tensor in data. You need to check if the variable data only contains one tensor, then the command tensor.numpy() should work fine.