Skip to content
This repository was archived by the owner on Apr 23, 2025. It is now read-only.

Matrix multiplication operator was changed, code needs update. #15

Merged
merged 7 commits into from
Jun 26, 2018
Merged

Matrix multiplication operator was changed, code needs update. #15

Changes from all commits
Commits
Show all changes
7 commits
Select commit Hold shift + click to select a range
03ae363
Add simple Autoencoder model. Using new `dynamicallyCall` API to inte…
VolodymyrPavliukevych Jun 4, 2018
7f00f7a
Improve model, apply code review issues. Add images. Improve README.
VolodymyrPavliukevych Jun 9, 2018
13d92d8
Apply code review comments.
VolodymyrPavliukevych Jun 12, 2018
53f9f93
Apply code review comments.
VolodymyrPavliukevych Jun 13, 2018
dec64ea
Update to latest version of API instead deprecated `Tensor.dot`.
VolodymyrPavliukevych Jun 13, 2018
2b024b3
Merge branch 'master' of https://github.com/tensorflow/swift-models
VolodymyrPavliukevych Jun 26, 2018
cc7f543
Fixe `MultiplicationPrecedence` operator issue. Apply new API.
VolodymyrPavliukevych Jun 26, 2018
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
24 changes: 12 additions & 12 deletions Autoencoder/Autoencoder.swift
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -128,13 +128,13 @@ extension Autoencoder {
let inputNormalized = input / 255.0

// Forward pass
let z1 = inputNormalized ⊗ w1
let z1 = inputNormalized • w1
let h1 = tanh(z1)
let z2 = h1 ⊗ w2 + b2
let z2 = h1 • w2 + b2
let h2 = z2
let z3 = h2 ⊗ w3
let z3 = h2 • w3
let h3 = tanh(z3)
let z4 = h3 ⊗ w4
let z4 = h3 • w4
let predictions = sigmoid(z4)
let loss: Float = 0.5 * (predictions - inputNormalized).squared().mean()
return (h2, loss, inputNormalized, predictions)
Expand All @@ -148,28 +148,28 @@ extension Autoencoder {
let batchSize = Tensor<Float>(inputNormalized.shapeTensor[0])

// Forward pass
let z1 = inputNormalized ⊗ w1
let z1 = inputNormalized • w1
let h1 = tanh(z1)
let z2 = h1 ⊗ w2 + b2
let z2 = h1 • w2 + b2
let h2 = z2
let z3 = h2 ⊗ w3
let z3 = h2 • w3
let h3 = tanh(z3)
let z4 = h3 ⊗ w4
let z4 = h3 • w4
let predictions = sigmoid(z4)

// Backward pass
let dz4 = ((predictions - inputNormalized) / batchSize)
let dw4 = h3.transposed(withPermutations: 1, 0) ⊗ dz4
let dw4 = h3.transposed(withPermutations: 1, 0) • dz4

let dz3 = matmul(dz4, w4.transposed(withPermutations: 1, 0)) * (1 - h3.squared())
let dw3 = h2.transposed(withPermutations: 1, 0) ⊗ dz3
let dw3 = h2.transposed(withPermutations: 1, 0) • dz3

let dz2 = matmul(dz3, w3.transposed(withPermutations: 1, 0))
let dw2 = h1.transposed(withPermutations: 1, 0) ⊗ dz2
let dw2 = h1.transposed(withPermutations: 1, 0) • dz2
let db2 = dz2.sum(squeezingAxes: 0)

let dz1 = matmul(dz2, w2.transposed(withPermutations: 1, 0)) * (1 - h1.squared())
let dw1 = inputNormalized.transposed(withPermutations: 1, 0) ⊗ dz1
let dw1 = inputNormalized.transposed(withPermutations: 1, 0) • dz1

let loss: Float = 0.5 * (predictions - inputNormalized).squared().mean()

Expand Down