As per title.
This will be useful to make Tensor shape propogations and einsum operations immediately understandable.
Moreover, it will better support a future symbolic engine.
The convention that should be followed:
einsum equations are always in lower case.
Comments and docstrings are always in upper case.
Letters and their meanings:
F (f): fold dimension;
B (b): batch dimension (possibly *B or ... if multiple);
I,J,K,O (i,j,k,o): units dimension (K general use; I,O (Ki, Ko): input/output; I,J: paired input);
R (r): rank (units with low rank) dimension;
H (h): layer inputs (arity) dimension, e.g., H=2 for Tucker2 layers;
D (d): variables dimension, e.g., D=784 for MNIST;
C (c): channels dimension, e.g., C=3 for RGB images.
P (p): replica dimension, useful for input layers.
As per title. This will be useful to make Tensor shape propogations and
einsum
operations immediately understandable. Moreover, it will better support a future symbolic engine.The convention that should be followed:
einsum
equations are always in lower case.Letters and their meanings:
*B
or...
if multiple);