bug in 2_supervised/1_data.lua when size='small'
almson opened this issue · comments
Aleksandr Dubinsky commented
The YUV conversion is done only on the first 10000 elements of trainData.data
, but the global normalization is done across all elements (including those which are still RGB and which have a different mean/variance). Fix is to slice the training/test data when doing global normalization, and again when "verifying statistics."
As an aside, why do the comments state "luminance is locally normalized and color globally," when in fact local and global normalization is applied to all channels?