wassname/keras-js
1
0
Fork 0
mirror of https://github.com/wassname/keras-js.git synced 2026-09-09 11:25:25 +08:00
Code Issues Packages Projects Releases Wiki Activity

tensor reshaping operation optimizations

Browse Source
This commit is contained in:
Leon Chen
2016-10-06 18:37:25 -04:00
parent 4c6ae6222e
commit f5f6109f13
6 changed files with 54 additions and 91 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/layers/convolutional/AtrousConvolution2D.js
+7 -7
View File
@@ -1,8 +1,6 @@
import Tensor from '../../Tensor'
import Convolution2D from './Convolution2D'
import ops from 'ndarray-ops'
import unpack from 'ndarray-unpack'
import flattenDeep from 'lodash/flattenDeep'
/**
* AtrousConvolution2D layer class
@@ -87,18 +85,20 @@ export default class AtrousConvolution2D extends Convolution2D {
const imColsMat = new Tensor([], [nbPatches, patchLen])
let patch = new Tensor([], [patchLen])
let patch = new Tensor([], [nbRow, nbCol, inputChannels])
let patchRaveled = new Tensor([], [patchLen])
let n = 0
for (let i = 0, limit = inputRows - nbRowDilated; i <= limit; i += this.subsample[0]) {
for (let j = 0, limit = inputCols - nbColDilated; j <= limit; j += this.subsample[1]) {
const patchData = flattenDeep(unpack(
ops.assign(
patch.tensor,
x.tensor
.hi(i + nbRowDilated, j + nbColDilated, inputChannels)
.lo(i, j, 0)
.step(this.atrousRate[0], this.atrousRate[1], 1)
))
patch.replaceTensorData(patchData)
ops.assign(imColsMat.tensor.pick(n, null), patch.tensor)
)
patchRaveled.replaceTensorData(patch.tensor.data)
ops.assign(imColsMat.tensor.pick(n, null), patchRaveled.tensor)
n += 1
}
}
src/layers/convolutional/Convolution2D.js
+13 -39
View File
@@ -3,8 +3,6 @@ import Tensor from '../../Tensor'
import Layer from '../../Layer'
import ops from 'ndarray-ops'
import gemm from 'ndarray-gemm'
import unpack from 'ndarray-unpack'
import flattenDeep from 'lodash/flattenDeep'
/**
* Convolution2D layer class
@@ -146,15 +144,14 @@ export default class Convolution2D extends Layer {
const imColsMat = new Tensor([], [nbPatches, patchLen])
let patch = new Tensor([], [patchLen])
let patch = new Tensor([], [nbRow, nbCol, inputChannels])
let patchRaveled = new Tensor([], [patchLen])
let n = 0
for (let i = 0, limit = inputRows - nbRow; i <= limit; i += this.subsample[0]) {
for (let j = 0, limit = inputCols - nbCol; j <= limit; j += this.subsample[1]) {
const patchData = flattenDeep(unpack(
x.tensor.hi(i + nbRow, j + nbCol, inputChannels).lo(i, j, 0)
))
patch.replaceTensorData(patchData)
ops.assign(imColsMat.tensor.pick(n, null), patch.tensor)
ops.assign(patch.tensor, x.tensor.hi(i + nbRow, j + nbCol, inputChannels).lo(i, j, 0))
patchRaveled.replaceTensorData(patch.tensor.data)
ops.assign(imColsMat.tensor.pick(n, null), patchRaveled.tensor)
n += 1
}
}
@@ -174,13 +171,12 @@ export default class Convolution2D extends Layer {
const wRowsMat = new Tensor([], [patchLen, nbFilter])
let patch = new Tensor([], [patchLen])
let patch = new Tensor([], [nbRow, nbCol, inputChannels])
let patchRaveled = new Tensor([], [patchLen])
for (let n = 0; n < nbFilter; n++) {
const patchData = flattenDeep(unpack(
this.weights.W.tensor.pick(null, null, null, n)
))
patch.replaceTensorData(patchData)
ops.assign(wRowsMat.tensor.pick(null, n), patch.tensor)
ops.assign(patch.tensor, this.weights.W.tensor.pick(null, null, null, n))
patchRaveled.replaceTensorData(patch.tensor.data)
ops.assign(wRowsMat.tensor.pick(null, n), patchRaveled.tensor)
}
return wRowsMat
@@ -197,25 +193,12 @@ export default class Convolution2D extends Layer {
x.tensor = x.tensor.transpose(1, 2, 0)
}
let startTime = performance.now()
this._calcOutputShape(x)
let endTime = performance.now()
console.log('_calcOutputShape', endTime - startTime)
startTime = performance.now()
this._padInput(x)
endTime = performance.now()
console.log('_padInput', endTime - startTime)
startTime = performance.now()
const imColsMat = this._im2col(x)
endTime = performance.now()
console.log('imColsMat', endTime - startTime)
startTime = performance.now()
const wRowsMat = this._w2row(x)
endTime = performance.now()
console.log('wRowsMat', endTime - startTime)
startTime = performance.now()
const nbFilter = this.kernelShape[0]
const outputRows = this.outputShape[0]
const outputCols = this.outputShape[1]
@@ -226,10 +209,7 @@ export default class Convolution2D extends Layer {
ops.assigns(matMul.tensor.pick(null, n), this.weights.b.tensor.get(n))
}
}
endTime = performance.now()
console.log('createMatMul', endTime - startTime)
startTime = performance.now()
if (x._useWeblas) {
const bias = this.bias
? this.weights.b.tensor.data
@@ -242,21 +222,15 @@ export default class Convolution2D extends Layer {
} else {
gemm(matMul.tensor, imColsMat.tensor, wRowsMat.tensor, 1, 1)
}
endTime = performance.now()
console.log('gemm', endTime - startTime)
startTime = performance.now()
let output = new Tensor([], this.outputShape)
let outputChannelRaveled = new Tensor([], [outputRows * outputCols])
let outputChannel = new Tensor([], [outputRows, outputCols])
for (let n = 0; n < nbFilter; n++) {
const outputChannelData = flattenDeep(unpack(
matMul.tensor.pick(null, n)
))
outputChannel.replaceTensorData(outputChannelData)
ops.assign(outputChannelRaveled.tensor, matMul.tensor.pick(null, n))
outputChannel.replaceTensorData(outputChannelRaveled.tensor.data)
ops.assign(output.tensor.pick(null, null, n), outputChannel.tensor)
}
endTime = performance.now()
console.log('createOutput', endTime - startTime)
x.tensor = output.tensor
this.activation(x)
src/layers/convolutional/Convolution3D.js
+13 -18
View File
@@ -3,8 +3,6 @@ import Tensor from '../../Tensor'
import Layer from '../../Layer'
import ops from 'ndarray-ops'
import gemm from 'ndarray-gemm'
import unpack from 'ndarray-unpack'
import flattenDeep from 'lodash/flattenDeep'
/**
* Convolution3D layer class
@@ -163,16 +161,15 @@ export default class Convolution3D extends Layer {
const volColsMat = new Tensor([], [nbPatches, patchLen])
let patch = new Tensor([], [patchLen])
let patch = new Tensor([], [kernelDim1, kernelDim2, kernelDim3, inputChannels])
let patchRaveled = new Tensor([], [patchLen])
let n = 0
for (let i = 0, limit = inputDim1 - kernelDim1; i <= limit; i += this.subsample[0]) {
for (let j = 0, limit = inputDim2 - kernelDim2; j <= limit; j += this.subsample[1]) {
for (let k = 0, limit = inputDim3 - kernelDim3; k <= limit; k += this.subsample[2]) {
const patchData = flattenDeep(unpack(
x.tensor.hi(i + kernelDim1, j + kernelDim2, k + kernelDim3, inputChannels).lo(i, j, k, 0)
))
patch.replaceTensorData(patchData)
ops.assign(volColsMat.tensor.pick(n, null), patch.tensor)
ops.assign(patch.tensor, x.tensor.hi(i + kernelDim1, j + kernelDim2, k + kernelDim3, inputChannels).lo(i, j, k, 0))
patchRaveled.replaceTensorData(patch.tensor.data)
ops.assign(volColsMat.tensor.pick(n, null), patchRaveled.tensor)
n += 1
}
}
@@ -193,13 +190,12 @@ export default class Convolution3D extends Layer {
const wRowsMat = new Tensor([], [patchLen, nbFilter])
let patch = new Tensor([], [patchLen])
let patch = new Tensor([], [kernelDim1, kernelDim2, kernelDim3, inputChannels])
let patchRaveled = new Tensor([], [patchLen])
for (let n = 0; n < nbFilter; n++) {
const patchData = flattenDeep(unpack(
this.weights.W.tensor.pick(null, null, null, null, n)
))
patch.replaceTensorData(patchData)
ops.assign(wRowsMat.tensor.pick(null, n), patch.tensor)
ops.assign(patch.tensor, this.weights.W.tensor.pick(null, null, null, null, n))
patchRaveled.replaceTensorData(patch.tensor.data)
ops.assign(wRowsMat.tensor.pick(null, n), patchRaveled.tensor)
}
return wRowsMat
@@ -248,12 +244,11 @@ export default class Convolution3D extends Layer {
}
let output = new Tensor([], this.outputShape)
let outputChannelRaveled = new Tensor([], [outputDim1 * outputDim2 * outputDim3])
let outputChannel = new Tensor([], [outputDim1, outputDim2, outputDim3])
for (let n = 0; n < nbFilter; n++) {
const outputChannelData = flattenDeep(unpack(
matMul.tensor.pick(null, n)
))
outputChannel.replaceTensorData(outputChannelData)
ops.assign(outputChannelRaveled.tensor, matMul.tensor.pick(null, n))
outputChannel.replaceTensorData(outputChannelRaveled.tensor.data)
ops.assign(output.tensor.pick(null, null, null, n), outputChannel.tensor)
}
x.tensor = output.tensor
src/layers/convolutional/Deconvolution2D.js
+13 -16
View File
@@ -3,8 +3,6 @@ import Tensor from '../../Tensor'
import Layer from '../../Layer'
import ops from 'ndarray-ops'
import gemm from 'ndarray-gemm'
import unpack from 'ndarray-unpack'
import flattenDeep from 'lodash/flattenDeep'
/**
* Deconvolution2D layer class
@@ -129,13 +127,12 @@ export default class Deconvolution2D extends Layer {
const [inputRows, inputCols, inputChannels] = x.tensor.shape
const imColsMat = new Tensor([], [inputRows * inputCols, inputChannels])
let channel = new Tensor([], [inputRows * inputCols])
let channelRaveled = new Tensor([], [inputRows * inputCols])
let channel = new Tensor([], [inputRows, inputCols])
for (let c = 0; c < inputChannels; c++) {
const channelData = flattenDeep(unpack(
x.tensor.pick(null, null, c)
))
channel.replaceTensorData(channelData)
ops.assign(imColsMat.tensor.pick(null, c), channel.tensor)
ops.assign(channel.tensor, x.tensor.pick(null, null, c))
channelRaveled.replaceTensorData(channel.tensor.data)
ops.assign(imColsMat.tensor.pick(null, c), channelRaveled.tensor)
}
return imColsMat
}
@@ -150,13 +147,12 @@ export default class Deconvolution2D extends Layer {
const [nbRow, nbCol, inputChannels, nbFilter] = this.weights.W.tensor.shape
const wRowsMat = new Tensor([], [inputChannels, nbRow * nbCol * nbFilter])
let channel = new Tensor([], [nbRow * nbCol * nbFilter])
let channelRaveled = new Tensor([], [nbRow * nbCol * nbFilter])
let channel = new Tensor([], [nbRow, nbCol, nbFilter])
for (let c = 0; c < inputChannels; c++) {
const channelData = flattenDeep(unpack(
this.weights.W.tensor.pick(null, null, c, null)
))
channel.replaceTensorData(channelData)
ops.assign(wRowsMat.tensor.pick(c, null), channel.tensor)
ops.assign(channel.tensor, this.weights.W.tensor.pick(null, null, c, null))
channelRaveled.replaceTensorData(channel.tensor.data)
ops.assign(wRowsMat.tensor.pick(c, null), channelRaveled.tensor)
}
return wRowsMat
}
@@ -211,11 +207,12 @@ export default class Deconvolution2D extends Layer {
const patchShape = [nbRow, nbCol, nbFilter]
let patch = new Tensor([], patchShape)
let patchRaveled = new Tensor([], [nbRow * nbCol * nbFilter])
let index = 0
for (let i = 0; i < inputRows; i++) {
for (let j = 0; j < inputCols; j++) {
const patchData = unpack(matMul.tensor.pick(index, null))
patch.replaceTensorData(patchData)
ops.assign(patchRaveled.tensor, matMul.tensor.pick(index, null))
patch.replaceTensorData(patchRaveled.tensor.data)
const iOutPos = i * this.subsample[0]
const jOutPos = j * this.subsample[1]
ops.addeq(
src/layers/core/Flatten.js
+4 -6
View File
@@ -1,12 +1,9 @@
import Tensor from '../../Tensor'
import Layer from '../../Layer'
import ndarray from 'ndarray'
import unpack from 'ndarray-unpack'
import flattenDeep from 'lodash/flattenDeep'
/**
* Flatten layer class
* Turns tensor into 1-d. Note there is no concept of batch size in these layers (single-batch).
* We use ndarray-unpack first, as ndarray striding/offsets precludes us from simply using x.tensor.data
*/
export default class Flatten extends Layer {
/**
@@ -24,8 +21,9 @@ export default class Flatten extends Layer {
*/
call (x) {
if (x.tensor.shape.length > 1) {
const shape = [x.tensor.shape.reduce((a, b) => a * b, 1)]
x.tensor = ndarray(new x._type(flattenDeep(unpack(x.tensor))), shape)
let raveled = new Tensor([], [x.tensor.shape.reduce((a, b) => a * b, 1)])
raveled.replaceTensorData(x.tensor.data)
x.tensor = raveled.tensor
}
return x
}
src/layers/core/Reshape.js
+4 -5
View File
@@ -1,12 +1,9 @@
import Tensor from '../../Tensor'
import Layer from '../../Layer'
import ndarray from 'ndarray'
import unpack from 'ndarray-unpack'
import flattenDeep from 'lodash/flattenDeep'
/**
* Reshape layer class
* Note there is no concept of batch size in these layers (single-batch).
* We use ndarray-unpack first, as ndarray striding/offsets precludes us from simply using x.tensor.data
*/
export default class Reshape extends Layer {
/**
@@ -32,7 +29,9 @@ export default class Reshape extends Layer {
if (this.targetShape.reduce((a, b) => a * b, 1) !== x.tensor.size) {
throw new Error(`${this.name} [Reshape layer] The total size of new array must be unchanged in reshape layer.`)
}
x.tensor = ndarray(new x._type(flattenDeep(unpack(x.tensor))), this.targetShape)
let reshaped = new Tensor([], this.targetShape)
reshaped.replaceTensorData(x.tensor.data)
x.tensor = reshaped.tensor
return x
}
}