important functions
tensors, shapes, ndims, add, matmul, argmax, reduce_sum, linalg, @, T, transpose, adjoint, band_part, cholesky decomposition, cross product, det, inv, trace, singular value decompostion, einsum, expand dims, squeeze, concat, stack, pad, gather, gather_nd, ragged tensor, boolean mask, from_the_row_length, from_the_row_limits, variable, with cpu/gpu
Development lifecycle:
Task → Data → Model → Error measurement → Training and optimization → Performance metric → Val and test → Fine tuning
Visuals
Hyperparameters: Input size, padding, kernel size, stride
features, parameters, receptive fields,
input + kernel → output
small kernel size, fine features are extracted, large kernel size, bigger patterns are observed
kernels help reduces the parameters to be learned, padding when kernel beyond the activation map, stride how many jumps in the sliding window, padding ensures corner features are picked up.
Binary classification problems:
Error: BinaryCrossEntropy Loss