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

Fix output sequence issue of lstm with backward direction and both direction #123

Merged
merged 7 commits into from
Jan 14, 2025
Merged

Fix output sequence issue of lstm with backward direction and both direction #123

Show file tree
Hide file tree
Changes from 3 commits
Commits
Show all changes
7 commits
Select commit Hold shift + click to select a range
814c241
update reshape without null
BruceDai Jan 6, 2025
f7e4fff
fix output sequence issue with backward and both directions
BruceDai Jan 7, 2025
613f056
address comments
BruceDai Jan 7, 2025
58cb672
fix sequence issue with returnSequence option
BruceDai Jan 8, 2025
ffceac7
fix lint error
BruceDai Jan 8, 2025
0ae6903
refine implementation of lstm
BruceDai Jan 8, 2025
d16218b
address comments
BruceDai Jan 9, 2025
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
31 changes: 27 additions & 4 deletions src/lstm.js
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -3,8 +3,10 @@
import {concat} from './concat.js';
import {lstmCell} from './lstm_cell.js';
import {reshape, squeeze} from './reshape.js';
import {reverse} from './reverse.js';
import {sizeOfShape, Tensor} from './lib/tensor.js';
import {sigmoid} from './sigmoid.js';
import {split} from './split.js';
import {slice} from './slice.js';
import {tanh} from './tanh.js';
import {validateLstmParams} from './lib/validate-input.js';
Expand Down Expand Up @@ -85,10 +87,11 @@ export function lstm(input, weight, recurrentWeight, steps, hiddenSize,
currentHidden[dir], currentCell[dir], hiddenSize, {bias: currentBias[dir],
recurrentBias: currentRecurrentBias[dir], peepholeWeight: currentPeepholeWeight[dir],
layout: layout, activations: activations});
// Expand [batchSize, hiddenSize] to [numDirections, batchSize, hiddenSize]
const output = reshape(results[0], [1, batchSize, hiddenSize]);
const cell = reshape(results[1], [1, batchSize, hiddenSize]);

const output = reshape(results[0], [1, null, hiddenSize]);
const cell = reshape(results[1], [1, null, hiddenSize]);

// Concat along axis 0 (numDirections dimension)
nextHidden = (nextHidden ? concat([nextHidden, output], 0) : output);
nextCell = (nextCell ? concat([nextCell, cell], 0) : cell);
}
Expand All @@ -97,10 +100,30 @@ export function lstm(input, weight, recurrentWeight, steps, hiddenSize,
cellState = nextCell;

if (returnSequence) {
nextHidden = reshape(nextHidden, [1, numDirections, null, hiddenSize]);
// Expand [numDirections, batchSize, hiddenSize] to
// [steps, numDirections, batchSize, hiddenSize]
nextHidden = reshape(nextHidden, [1, numDirections, batchSize, hiddenSize]);
// Concat output sequence along axis 0 (steps dimension)
sequence = (sequence ? concat([sequence, nextHidden], 0) : nextHidden);
}
}

if (direction === 'backward') {
// Refer to https://www.w3.org/TR/webnn/#api-mlgraphbuilder-lstm, Spec says the
// sequence should contain every output from each time step in the temporal sequence, while
// the loop for steps concatenates sequence in a reversed order when direction is backward,
// so here need reverse output sequence along axis 0 (steps dimension).
sequence = reverse(sequence, {axes: [0]});
} else if (direction === 'both') {
// Split output sequence into forward-sequence and backward-sequence two sequences along axis 1
// (numDirections dimension)
const [sequenceForward, sequenceBackward] = split(sequence, 2, {axis: 1});
// Reverse backward-sequence along axis 0 (steps dimension)
const reversedSequenceBackward = reverse(sequenceBackward, {axes: [0]});
sequence = concat([sequenceForward, reversedSequenceBackward], 1);
} else {
// No need update sequence for 'forward' direction
}

return (sequence ? [hiddenState, cellState, sequence] : [hiddenState, cellState]);
}
74 changes: 67 additions & 7 deletions test/lstm_test.js
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -45,7 +45,6 @@ describe('test lstm', function() {
input, weight, recurrentWeight, steps, hiddenSize,
{bias, recurrentBias, peepholeWeight, initialHiddenState,
initialCellState, returnSequence, activations});
console.log('outputs: ', outputs);
utils.checkShape(outputs[0], [numDirections, batchSize, hiddenSize]);
utils.checkShape(outputs[1], [numDirections, batchSize, hiddenSize]);
utils.checkShape(outputs[2], [steps, numDirections, batchSize, hiddenSize]);
Expand All @@ -65,7 +64,7 @@ describe('test lstm', function() {
}
});

it('lstm steps=2 direction="backward" returnSequence=true' +
it('lstm steps=2 direction="backward" returnSequence=true ' +
'activations=[relu, relu, relu]', function() {
const steps = 2;
const numDirections = 1;
Expand Down Expand Up @@ -106,22 +105,83 @@ describe('test lstm', function() {
input, weight, recurrentWeight, steps, hiddenSize,
{bias, recurrentBias, peepholeWeight, initialHiddenState,
initialCellState, direction, returnSequence, activations});
console.log('outputs: ', outputs);
utils.checkShape(outputs[0], [numDirections, batchSize, hiddenSize]);
utils.checkShape(outputs[1], [numDirections, batchSize, hiddenSize]);
utils.checkShape(outputs[2], [steps, numDirections, batchSize, hiddenSize]);
const expected = [
[10.469, 58.02899999999999, 74.529, 518.9490000000001],
[5.51, 20.009999999999998, 19.11, 75.21000000000001],
[
1,
8,
1,
8,
10.469,
58.02899999999999,
74.529,
518.9490000000001,
1,
8,
1,
8,
],
];
for (let i = 0; i < expected.length; ++i) {
utils.checkValue(outputs[i], expected[i]);
}
});

it('lstm steps=2 direction="both" returnSequence=true', function() {
const steps = 2;
const numDirections = 2;
const batchSize = 2;
const inputSize = 2;
const hiddenSize = 2;
const input = new Tensor([steps, batchSize, inputSize],
new Float32Array([1, 2, 2, 1, 3, 4, 1, 2]));
const weight = new Tensor([numDirections, 4 * hiddenSize, inputSize],
new Float32Array([
1, -1, 2, -2, 1, -1, 2, -2,
1, -1, 2, -2, 1, -1, 2, -2,
1, -1, 2, -2, 1, -1, 2, -2,
1, -1, 2, -2, 1, -1, 2, -2,
]));
const recurrentWeight = new Tensor([numDirections, 4 * hiddenSize, hiddenSize],
new Array(2 * 4 * hiddenSize * hiddenSize).fill(0.1));
const bias = new Tensor([numDirections, 4 * hiddenSize],
new Float32Array([
1, 2, 1, 2, 1, 2, 1, 2, 1, 2, 1, 2, 1, 2, 1, 2,
]));
const recurrentBias = new Tensor([numDirections, 4 * hiddenSize],
new Float32Array([
1, 2, 1, 2, 1, 2, 1, 2, 1, 2, 1, 2, 1, 2, 1, 2,
]));
const returnSequence = true;
const direction = 'both';
const outputs = lstm(
input, weight, recurrentWeight, steps, hiddenSize,
{bias, recurrentBias, direction, returnSequence});
utils.checkShape(outputs[0], [numDirections, batchSize, hiddenSize]);
utils.checkShape(outputs[1], [numDirections, batchSize, hiddenSize]);
utils.checkShape(outputs[2], [steps, numDirections, batchSize, hiddenSize]);
const expected = [
[
0.5764073262004139, 0.8236227651782412,
0.6612355785279247, 0.8442635760318142,
0.5764073262004139, 0.8236227651782412,
0.8635294727880538, 0.9491350760903781,
],
[
1.0171455721466105, 1.6205496282195793,
1.338846378789257, 1.7642604746965693,
1.0171455721466105, 1.6205496282195793,
1.485626937219704, 1.8449554199024933,
],
[
0.36960635293570576, 0.6082834181835157,
0.7037753329989016, 0.7586680430344475,
0.5764073262004139, 0.8236227651782412,
0.8635294727880538, 0.9491350760903781,
0.5764073262004139, 0.8236227651782412,
0.6612355785279247, 0.8442635760318142,
0.36960635293570576, 0.6082834181835157,
0.36960635293570576, 0.6082834181835157,
],
];
for (let i = 0; i < expected.length; ++i) {
Expand Down
Loading