Brain.js: GPU support both slow and wrong - almost hangs computer, node and chrome
What is wrong?
Not sure what to expect from GPU support. Weird training results.
Where does it happen?
Running Windows 7, Node v12.13.1, Chrome 80.0.3987.132, brain.js several versions - downloaded brain-browser.js from github, cloned repo development branch, master branch, 2.0.0-alpha.2
Geforce GTX Titan X with 441.12
How do we replicate the issue?
In browser
<html>
<head>
<script src="../brain.js/dist/brain-browser.js"></script>
<script>
const net = new brain.NeuralNetworkGPU();
const xorTrainingData = [
{input: [0, 0], output: [0]},
{input: [0, 1], output: [1]},
{input: [1, 0], output: [1]},
{input: [1, 1], output: [0]}
];
const opts = {
log: true,
logPeriod: 100,
iterations: 300
}
console.log(net.train(xorTrainingData, opts));
console.log(net.run([0,0]), Math.round( net.run([0,0]) ));
console.log(net.run([0,1]), Math.round( net.run([0,1]) ));
console.log(net.run([1,0]), Math.round( net.run([1,0]) ));
console.log(net.run([1,1]), Math.round( net.run([1,1]) ));
</script>
</head>
<body></body>
</html>
This pretty much hangs chrome for me, the tab becomes unresponsive and I have to pause javascript execution to have any chance to close the tab without force shutting down chrome. Increasing iterations or changing hidden layers or experimenting with other parameters have basically no effect.
//all four outputs
0.5000461935997009
0.5000461935997009
0.5000461935997009
0.5000461935997009
In node
const brain = require('brain.js');
const start = new Date()
console.log(brain)
const net = new brain.NeuralNetworkGPU({ hiddenLayers: [10] });
const xorTrainingData = [
{input: [0, 0], output: [0]},
{input: [0, 1], output: [1]},
{input: [1, 0], output: [1]},
{input: [1, 1], output: [0]}];
console.log(net.train(xorTrainingData, {
log: true,
logPeriod: 100,
iterations: 20000
}));
console.log( net.run([0,0]), Math.round( net.run([0,0]) ));
console.log( net.run([0,1]), Math.round( net.run([0,1]) ));
console.log( net.run([1,0]), Math.round( net.run([1,0]) ));
console.log( net.run([1,1]), Math.round( net.run([1,1]) ));
const after = new Date()
console.log((after.getTime() - start.getTime()) / 1000)
With GPU
iterations: 4100, training error: 0.006740194745361805
iterations: 4200, training error: 0.005839236546307802
iterations: 4300, training error: 0.005128919146955013
iterations: 4400, training error: 0.004557380452752113
{ error: 0.004557380452752113, iterations: 4400 }
Float32Array [ 0.09155591577291489 ] 0
Float32Array [ 0.09155591577291489 ] 0 <------------- slow AND wrong
Float32Array [ 0.09155591577291489 ] 0
Float32Array [ 0.09155591577291489 ] 0
7.165 (seconds)
Without GPU - (new brain.NeuralNetwork) on row 7
iterations: 4200, training error: 0.006874370443914483
iterations: 4300, training error: 0.006211034862270221
iterations: 4400, training error: 0.005647163558867768
iterations: 4500, training error: 0.005163651528723463
{ error: 0.004997787408525928, iterations: 4538 }
Float32Array [ 0.052155524492263794 ] 0
Float32Array [ 0.9345424175262451 ] 1
Float32Array [ 0.9339274764060974 ] 1
Float32Array [ 0.09189336001873016 ] 0
0.055 (seconds)
How important is this (1-5)?
4
Expected behavior (i.e. solution).
Faster training time in both Node and Browser environment with same training results.
Other Comments
Loving brain.js :heart: - I'm solving a Rubiks Cube using it for fun, very interesting!
ErikLarsson82
All 6 comments
Reproduced, will fix today.
robertleeplummerjr
on 16 Mar 2020
Found the issue, at least for the active branch I'm working on, which is rnn-cleanup. I'll have it released in the next day or two.
robertleeplummerjr
on 16 Mar 2020
Im having a similar issue with both Linux+Radeon480card or macOS/macookPro with Intel Iris Plus Graphics 650, GPU results are much much slower than CPU ones - maybe that's expected behavior with those type of card? Normal CPU feedforward training works like a charm and is great!
panzad3
on 4 May 2020
Fixes have been implemented already, will be available in next beta version.
mubaidr
on 4 May 2020
@mubaidr how can we test that beta version?
goferito
on 17 May 2020
Will be published soon.
Confirmed! Problem still exists in 2.0.Beta.1
node "d:\current\brainjs-test\index.js"
CPU training: 48.933ms
CPU training output { error: 0.004996030448576466, iterations: 5257 }
GPU training: 13341.557ms
GPU training output { error: 0.0047736987471580505, iterations: 6100 }
const brain = require("brain.js");
const trainingData = [
{ input: [0, 1], output: [1] },
{ input: [0, 0], output: [0] },
{ input: [1, 1], output: [0] },
{ input: [1, 0], output: [1] },
];
const net = new brain.NeuralNetwork({
// hiddenLayers: [12, 8],
});
console.time("CPU training");
const output = net.train(trainingData, {
iterations: 50000, // maximum training iterations
log: true, // console.log() progress periodically
logPeriod: 10000, // number of iterations between logging
});
console.timeEnd("CPU training");
console.log(`CPU training output`, output);
console.time("GPU training");
const netGPU = new brain.NeuralNetworkGPU({
// hiddenLayers: [12, 8],
});
const outputGPU = netGPU.train(trainingData, {
iterations: 50000, // maximum training iterations
log: true, // console.log() progress periodically
logPeriod: 10000, // number of iterations between logging
});
console.timeEnd("GPU training");
console.log(`GPU training output`, outputGPU);
@goferito cc
mubaidr
on 17 May 2020
Related issues
VEXLife
路
4Comments
nezzard
路
5Comments
tyt34
路
5Comments
ximik753
路
3Comments
AkashGutha
路
5Comments
Most helpful comment
Will be published soon.Confirmed! Problem still exists in
2.0.Beta.1@goferito cc