elektronn2.utils.cnncalculator module

elektronn2.utils.cnncalculator.cnncalculator(filters, poolings, desired_patch_size=None, mfp=False, force_center=False, desired_output=None, ndim=1)

Helper to calculate CNN architectures

This is a function, but it returns an object that has various architecture values as attributes. Useful is also to simply print ‘d’ as in the example.

Parameters:

• filters (list) – Filter shapes (for anisotropic filters the shapes are again a list)
• poolings (list) – Pooling factors
• desired_patch_size (int or list[int]) – Desired patch_size size(s). If None a range of suggestions can be found in the attribute valid_patch_sizes
• mfp (list[bool] or bool) – Whether to apply Max-Fragment-Pooling in this Layer and check compliance with max-fragment-pooling (requires other patch_size sizes than normal pooling)
• force_center (Bool) – Check if output neurons/pixel lie at center of patch_size neurons/pixel (and not in between)
• desired_output (None or int or list[int]) – Alternative to desired_patch_size
• ndim (int) – Dimensionality of CNN

Examples

Calculation for anisotropic “flat” 3d CNN with mfp in the first layers only:

>>> desired_patch_size   = [8, 211, 211]
>>> filters         = [[1,6,6], [4,4,4], [2,2,2], [1,1,1]]
>>> pool            = [[1,2,2], [2,2,2], [2,2,2], [1,1,1]]
>>> mfp             = [True,    True,    False,    False ]
>>> ndim=3
>>> d = cnncalculator(filters, pool, desired_patch_size, mfp=mfp, force_center=True, desired_output=None, ndim=ndim)
patch_size (8) changed to (10) (size too small)
patch_size (211) changed to (210) (size not possible)
patch_size (211) changed to (210) (size not possible)
>>> print(d)
patch_size: [10, 210, 210]
Layer/Fragment sizes:   [(1, 24, 24), (1, 24, 24), (3, 49, 49), (10, 102, 102)]
Unpooled Layer sizes:   [(1, 24, 24), (2, 48, 48), (7, 99, 99), (10, 205, 205)]
Receptive fields:       [(9, 23, 23), (9, 23, 23), (5, 15, 15), (1, 7, 7)]
Strides:                [(4, 8, 8), (4, 8, 8), (2, 4, 4), (1, 2, 2)]
Overlap:                [(5, 15, 15), (5, 15, 15), (3, 11, 11), (0, 5, 5)]
Offset:         [4.5, 11.5, 11.5]
If offset is non-int: output neurons lie centered on patch_size neurons,they have an odd FOV

elektronn2.utils.cnncalculator.get_cloesest_valid_patch_size(filters, poolings, desired_patch_size=100, mfp=False, ndim=1)

elektronn2.utils.cnncalculator.get_valid_patch_sizes(filters, poolings, desired_patch_size=100, mfp=False, ndim=1)