Repository navigation
Mnist why 32x32 and not 28x28 #48
Description
Hello there,
I don't understand why the dataset MNIST from Torch Tutorial dataset MNIST is made of 32x32 images instead of 28x28 as in the original dataset.
Do I have to crop the images using narrow ?
Thanks
Activity
chengzhishi commentedon Aug 19, 2019
Here is an explanation I found A Tensorflow Tutorial:
The MNIST dataset contains vectorized images of 28X28. Therefore we define a new function to reshape each batch of MNIST images to 28X28 and then resize to 32X32. The reason of resizing to 32X32 is to make it a power of two and therefore we can easily use the stride of 2 for downsampling and upsampling.
Hope this helps!