Example of how the IncrementalPCA MemoryError failure can occur.

check_arr_fail.py

import numpy as np
from sklearn.utils import check_array

ut = np.memmap('D:\\my_array2.mmap', dtype=np.float16, mode='w+', shape=(140000,3504))
print repr(ut), ut.shape
res = check_array(ut)
print 'Memory check passed',repr(res),res.shape #This will print the same contents and shape as above

res = check_array(ut, dtype=np.float) # This is the actual line used in IncrementalPCA
                                      # https://github.com/scikit-learn/scikit-learn/blob/0.16.X/sklearn/decomposition/incremental_pca.py#L165
print 'Never reaches here' # You'll get a memory error - this will never print

Amazing spot!!! Thanks! It seems for the time being as I might use np.float64 on a machine with a larger RAM.