Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

🔨 refactor(layer_2d): expose indices in VQ2D #99

Merged
merged 3 commits into from
Jun 10, 2023
Merged

🔨 refactor(layer_2d): expose indices in VQ2D #99

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
03d0090
Expose indices in VQ2D.
jean-francoisreboud Jun 10, 2023
a4f50f6
Update changelog.
jean-francoisreboud Jun 10, 2023
8e5b21f
Public K.
jean-francoisreboud Jun 10, 2023
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
1 change: 1 addition & 0 deletions CHANGELOG.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -4,6 +4,7 @@ All notable changes to this project will be documented in this file.

## [unreleased]

🔨 **layer_2d:** expose indices in VQ2D ([#99](https://github.com/owkin/GrAIdient/pull/99))\
🪜 **layer_2d:** loosen range contraint in ColorJitterHSV ([#98](https://github.com/owkin/GrAIdient/pull/98))\
🪜 **layer_2d:** SimilarityError2D & dirty losses ([#97](https://github.com/owkin/GrAIdient/pull/97))\
🔨 **core:** LayerWeightInit ([#96](https://github.com/owkin/GrAIdient/pull/96))\
Expand Down
68 changes: 34 additions & 34 deletions Sources/GrAIdient/Layer2D/VQ2D.swift
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -11,7 +11,7 @@ import Foundation
public class VQ2D: Layer2D, LayerWeightInit
{
/// The number of vector approximations.
let _K: Int
public let K: Int

/// Coefficient for commitment.
public var beta: Double
Expand All @@ -20,7 +20,7 @@ public class VQ2D: Layer2D, LayerWeightInit
/// Indices of maximal elements.
/// Shape ~ (batch, height, width).
///
var _indices: MetalBuffer<Int32>! = nil
public var indices: MetalBuffer<Int32>! = nil

///
/// Grid of weights.
Expand Down Expand Up @@ -59,7 +59,7 @@ public class VQ2D: Layer2D, LayerWeightInit
}

var weightsTmp = [Float]()
for k in 0..<_K {
for k in 0..<K {
for depth in 0..<nbChannels
{
weightsTmp.append(Float(_wArrays.w(k, depth)))
Expand Down Expand Up @@ -98,7 +98,7 @@ public class VQ2D: Layer2D, LayerWeightInit
public var connectivityIO: (Int, Int)
{
get {
return (nbChannels, _K)
return (nbChannels, K)
}
}

Expand All @@ -123,7 +123,7 @@ public class VQ2D: Layer2D, LayerWeightInit
beta: Double,
params: GrAI.Model.Params)
{
_K = K
self.K = K
self.beta = beta
super.init(
layerPrev: layerPrev,
Expand All @@ -146,7 +146,7 @@ public class VQ2D: Layer2D, LayerWeightInit
{
let values = try decoder.container(keyedBy: Keys.self)

_K = try values.decode(Int.self, forKey: .K)
K = try values.decode(Int.self, forKey: .K)
beta = try values.decode(Double.self, forKey: .beta)

try super.init(from: decoder)
Expand All @@ -170,7 +170,7 @@ public class VQ2D: Layer2D, LayerWeightInit
{
var container = encoder.container(keyedBy: Keys.self)

try container.encode(_K, forKey: .K)
try container.encode(K, forKey: .K)
try container.encode(beta, forKey: .beta)

let weightsList: [Float]
Expand Down Expand Up @@ -210,7 +210,7 @@ public class VQ2D: Layer2D, LayerWeightInit
params.context.curID = id

let layer = VQ2D(
layerPrev: layerPrev, K: _K, beta: beta, params: params
layerPrev: layerPrev, K: K, beta: beta, params: params
)
if inPlace
{
Expand Down Expand Up @@ -242,7 +242,7 @@ public class VQ2D: Layer2D, LayerWeightInit
{
super.resetKernelCPU()
_wArrays?.reset()
_indices = nil
indices = nil
}

///
Expand All @@ -256,7 +256,7 @@ public class VQ2D: Layer2D, LayerWeightInit
{
super.resetKernelGPU()

_indices = nil
indices = nil
_wDeltaWeights = nil
_wBuffers?.reset()
}
Expand All @@ -273,9 +273,9 @@ public class VQ2D: Layer2D, LayerWeightInit
_weightsList = generateWeightsList()
}

_wArrays = WeightGrids(width: nbChannels, height: _K)
_wArrays = WeightGrids(width: nbChannels, height: K)

for k in 0..<_K {
for k in 0..<K {
for depth in 0..<nbChannels
{
let offset = depth + nbChannels * k
Expand All @@ -297,12 +297,12 @@ public class VQ2D: Layer2D, LayerWeightInit
}

_wBuffers = WeightBuffers(
nbElems: _K * nbChannels,
nbElems: K * nbChannels,
deviceID: deviceID
)

let weightsPtr = _wBuffers.w_p!.shared.buffer
for elem in 0..<_K * nbChannels
for elem in 0..<K * nbChannels
{
weightsPtr[elem] = _weightsList[elem]
}
Expand All @@ -321,9 +321,9 @@ public class VQ2D: Layer2D, LayerWeightInit
{
try super.checkStateCPU(batchSize: batchSize)

if _indices == nil
if indices == nil
{
_indices = MetalSharedBuffer<Int32>(
indices = MetalSharedBuffer<Int32>(
batchSize * height * width,
deviceID: deviceID
)
Expand All @@ -344,13 +344,13 @@ public class VQ2D: Layer2D, LayerWeightInit
GrAI.Gradient.sample && _wDeltaWeights == nil
{
_wDeltaWeights = MetalPrivateBuffer<Float>(
batchSize * _K * nbChannels, deviceID: deviceID
batchSize * K * nbChannels, deviceID: deviceID
)
}

if _indices == nil
if indices == nil
{
_indices = MetalPrivateBuffer<Int32>(
indices = MetalPrivateBuffer<Int32>(
batchSize * height * width,
deviceID: deviceID
)
Expand All @@ -369,7 +369,7 @@ public class VQ2D: Layer2D, LayerWeightInit
try checkStateCPU(batchSize: batchSize)

let neuronsPrev = layerPrev.neurons
let indicesPtr = (_indices as! MetalSharedBuffer<Int32>).buffer
let indicesPtr = (indices as! MetalSharedBuffer<Int32>).buffer

for elem in 0..<batchSize {
for i in 0..<height {
Expand All @@ -378,7 +378,7 @@ public class VQ2D: Layer2D, LayerWeightInit
var minIndex = -1
var minValue: Double? = nil

for k in 0..<_K
for k in 0..<K
{
var value: Double = 0.0
for depth in 0..<nbChannels
Expand Down Expand Up @@ -425,7 +425,7 @@ public class VQ2D: Layer2D, LayerWeightInit
let pNbChannels: [UInt32] = [UInt32(nbChannels)]
let pNbBatch: [UInt32] = [UInt32(batchSize)]
let pDimensions: [UInt32] = [UInt32(width), UInt32(height)]
let pK: [UInt32] = [UInt32(_K)]
let pK: [UInt32] = [UInt32(K)]

let command = MetalKernel.get.createCommand(
"vq2DForward", deviceID: deviceID
Expand All @@ -437,7 +437,7 @@ public class VQ2D: Layer2D, LayerWeightInit
command.setBytes(pK, atIndex: 4)
command.setBytes(pNbBatch, atIndex: 5)
command.setBuffer(outs.metal, atIndex: 6)
command.setBuffer(_indices.metal, atIndex: 7)
command.setBuffer(indices.metal, atIndex: 7)

command.dispatchThreads(
width: height * width,
Expand All @@ -459,7 +459,7 @@ public class VQ2D: Layer2D, LayerWeightInit
if let layerPrev = self.layerPrev as? Layer2D, mustComputeBackward
{
let neuronsPrev = layerPrev.neurons
let indicesPtr = (_indices as! MetalSharedBuffer<Int32>).buffer
let indicesPtr = (indices as! MetalSharedBuffer<Int32>).buffer

for elem in 0..<batchSize {
for i in 0..<height {
Expand Down Expand Up @@ -498,11 +498,11 @@ public class VQ2D: Layer2D, LayerWeightInit
{
let neuronsPrev = layerPrev.neurons
let coeff = batchSize * height * width
let indicesPtr = (_indices as! MetalSharedBuffer<Int32>).buffer
let indicesPtr = (indices as! MetalSharedBuffer<Int32>).buffer

if !accumulateDeltaWeights
{
for k in 0..<_K {
for k in 0..<K {
for depth in 0..<nbChannels
{
_wArrays.g(k, depth, 0.0)
Expand Down Expand Up @@ -550,7 +550,7 @@ public class VQ2D: Layer2D, LayerWeightInit
let pNbChannels: [UInt32] = [UInt32(nbChannels)]
let pNbBatch: [UInt32] = [UInt32(batchSize)]
let pDimensions: [UInt32] = [UInt32(width), UInt32(height)]
let pK: [UInt32] = [UInt32(_K)]
let pK: [UInt32] = [UInt32(K)]
let pBeta: [Float] = [Float(beta)]
let pDirty: [UInt32] = layerPrev.dirty ? [1] : [0]

Expand All @@ -560,7 +560,7 @@ public class VQ2D: Layer2D, LayerWeightInit
command.setBuffer(layerPrev.outs.metal, atIndex: 0)
command.setBuffer(delta.metal, atIndex: 1)
command.setBuffer(_wBuffers.w.metal, atIndex: 2)
command.setBuffer(_indices.metal, atIndex: 3)
command.setBuffer(indices.metal, atIndex: 3)
command.setBytes(pNbChannels, atIndex: 4)
command.setBytes(pDimensions, atIndex: 5)
command.setBytes(pK, atIndex: 6)
Expand All @@ -586,7 +586,7 @@ public class VQ2D: Layer2D, LayerWeightInit
let pNbChannels: [UInt32] = [UInt32(nbChannels)]
let pNbBatch: [UInt32] = [UInt32(batchSize)]
let pDimensions: [UInt32] = [UInt32(width), UInt32(height)]
let pK: [UInt32] = [UInt32(_K)]
let pK: [UInt32] = [UInt32(K)]
let pAccumulate: [UInt32] = accumulateDeltaWeights ? [1] : [0]

var command: MetalCommand
Expand Down Expand Up @@ -615,14 +615,14 @@ public class VQ2D: Layer2D, LayerWeightInit
)
command.setBuffer(layerPrev.outs.metal, atIndex: 0)
command.setBuffer(_wBuffers.w.metal, atIndex: 1)
command.setBuffer(_indices.metal, atIndex: 2)
command.setBuffer(indices.metal, atIndex: 2)
command.setBytes(pNbChannels, atIndex: 3)
command.setBytes(pDimensions, atIndex: 4)
command.setBytes(pK, atIndex: 5)
command.setBytes(pNbBatch, atIndex: 6)
command.setBuffer(_wBuffers.g.metal, atIndex: 7)

command.dispatchThreads(width: nbChannels, height: _K)
command.dispatchThreads(width: nbChannels, height: K)
command.enqueue()
}
else
Expand All @@ -647,7 +647,7 @@ public class VQ2D: Layer2D, LayerWeightInit
)
command.setBuffer(layerPrev.outs.metal, atIndex: 0)
command.setBuffer(_wBuffers.w.metal, atIndex: 1)
command.setBuffer(_indices.metal, atIndex: 2)
command.setBuffer(indices.metal, atIndex: 2)
command.setBytes(pNbChannels, atIndex: 3)
command.setBytes(pDimensions, atIndex: 4)
command.setBytes(pK, atIndex: 5)
Expand All @@ -656,7 +656,7 @@ public class VQ2D: Layer2D, LayerWeightInit

command.dispatchThreads(
width: nbChannels,
height: batchSize * _K
height: batchSize * K
)
command.enqueue()

Expand All @@ -673,7 +673,7 @@ public class VQ2D: Layer2D, LayerWeightInit
command.setBytes(pAccumulate, atIndex: 4)
command.setBuffer(_wBuffers.g.metal, atIndex: 5)

command.dispatchThreads(width: nbChannels, height: _K)
command.dispatchThreads(width: nbChannels, height: K)
command.enqueue()
}
}
Expand Down