Code

Initial implementation of autograd#30

Code

Merged

umar456 merged 20 commits intoarrayfire:masterarrayfire/arrayfire-ml:masterfrom pavanky:autogradpavanky/arrayfire-ml:autogradCopy head branch name to clipboardJul 6, 2017

+1,315-959

Commits

Commits on Jul 2, 2017

• Reorganizing the CMake files

  pavanky

  committed
  5ed7a88

  Copy full SHA for 5ed7a88

• Reorganizing the include files and namespace

  pavanky

  committed
  dfa8fda

  Copy full SHA for dfa8fda

• First attempt at implementing autograd

  pavanky

  committed
  562b860

  Copy full SHA for 562b860

• Store gradients as autograd::Variable instead of af::array

  pavanky

  committed
  3f832a0

  Copy full SHA for 3f832a0

Commits on Jul 5, 2017

• Refactor autograd::Variable, option to disable grad calculations

  Show description for c94ee3d
  pavanky

  committed
  c94ee3d

  Copy full SHA for c94ee3d

• Changing autograd::backward function to Variable::backward method

  pavanky

  committed
  7bb0b6c

  Copy full SHA for 7bb0b6c

• Moving autograd from header only lib to a compiled lib

  pavanky

  committed
  a316af0

  Copy full SHA for a316af0

• Adding negate, reciprocal, subtract and divide

  pavanky

  committed
  d7edafc

  Copy full SHA for d7edafc

• Add scalar support for operators

  pavanky

  committed
  664cf7c

  Copy full SHA for 664cf7c

• Adding exp, sin, cos, tanh, and sigmoid functions

  pavanky

  committed
  45b21da

  Copy full SHA for 45b21da

• Adding expandAs, reduceAs, and transpose

  pavanky

  committed
  3b985b0

  Copy full SHA for 3b985b0

• Adding matmul, matmulTN, and matmulNT functions

  pavanky

  committed
  9b05273

  Copy full SHA for 9b05273

• Add option to explicitly request higher order gradients.

  Show description for 49b8917
  pavanky

  committed
  49b8917

  Copy full SHA for 49b8917

Commits on Jul 6, 2017

• Convert Variable::build and Variable::buildSubGraph to static functions

  pavanky

  committed
  8bf7f1b

  Copy full SHA for 8bf7f1b

• Overhaul of af::nn to use af::autograd

  Show description for 5eda600
  pavanky

  committed
  5eda600

  Copy full SHA for 5eda600

• Fixing bugs in backward pass for activation functions

  pavanky

  committed
  9aefea4

  Copy full SHA for 9aefea4

• Fixing perceptron example to use smaller batch size

  Show description for 6d5751a
  pavanky

  committed
  6d5751a

  Copy full SHA for 6d5751a

• Adding model.eval() and model.train()

  pavanky

  committed
  a01504b

  Copy full SHA for a01504b

• Formatting changes

  pavanky

  committed
  2776aa2

  Copy full SHA for 2776aa2

• Use references while iterating when possible

  pavanky

  committed
  04cd450

  Copy full SHA for 04cd450