It is difficult to run tensorflow in a lambda:
- Size constraints from the lambda environment (<50MB)
- Native code compilation for the lambda environment
This family of packages will handle that heavy lifting for you.
See the GitHub repository for version information. This package will always contain a matrix of the currently supported AWS Lambda Node environments and the @tensorflow/tfjs-node
latest major versions.
Example matrix:
{
"lambda": ["nodejs10.x", "nodejs12.x", "nodejs14.x"],
"tensorflow": ["1.7.4", "2.8.6", "3.6.1"]
}
Example releases:
nodejs10.x-tf1.7.4.br
nodejs10.x-tf2.8.6.br
nodejs10.x-tf3.6.1.br
nodejs12.x-tf1.7.4.br
nodejs12.x-tf2.8.6.br
nodejs12.x-tf3.6.1.br
nodejs14.x-tf1.7.4.br
nodejs14.x-tf2.8.6.br
nodejs14.x-tf3.6.1.br
Using GitHub actions, these versions are updated daily around midnight. Lock your versions with the --save-exact
flag.
npm install --save --save-exact tfjs-node-lambda tfjs-node-lambda-helpers
npm install --save-dev --save-exact @tensorflow/tfjs-node @tensorflow/tfjs
import loadTf from 'tfjs-node-lambda';
const tf: typeof import('@tensorflow/tfjs') = await loadTf(readStream);
The tfjs-node-lambda-helpers
can help determine environments, generate release urls, download releases, and avoid timeouts. If you would prefer to have full control, see the readStream examples below.
In development, loadTf
will run your locally installed version of @tensorflow/tfjs-node
. If running in an AWS Lambda environment, loadTf
expects a readStream of the release:
import { Readable } from 'stream';
const response = await axios.get(
'https://github.com/jlarmstrongiv/tfjs-node-lambda/releases/download/v2.0.10/nodejs12.x-tf2.8.6.br',
{ responseType: 'arraybuffer' },
);
const readStream = Readable.from(response.data);
Or, if you have saved the file to the tmp directory:
import os from 'os';
import path from 'path';
const readStream = fs.createReadStream(
path.join(os.tmpdir(), 'nodejs12.x-tf2.8.6.br'),
);
const tf = await loadTf(readStream);
Due to having multiple supported versions of @tensorflow/tfjs-node
and an irregular bundle method, you must specify the types:
const tf: typeof import('@tensorflow/tfjs') = await loadTf(readStream);
We welcome contributions!
Inspired by and originally forked from Luc Leray’s tensorflow-lambda.
†tfjs-node-lambda-releases is soft deprecated. Due to file size limitations, 6 releases are too big to be published on npm. Use the assets on GitHub Releases instead.