I was playing with the mnist example just now and wanted to print the first training example. So after the line

training_data = training_set[0].view(-1, 1, 28, 28).div(255)

I wrote

print(training_data[0])

Printing took ~33 seconds on my mbp. print(training_data[0].numpy()) is instantaneous.

Oddly,

a = torch.FloatTensor(1, 28,28)
print(a)

which prints a tensor of the same shape an type is fast.