Tfjs: mnist-node example: backend.select is not a function

[caisq]

To get help from the community, check out our Google group.

TensorFlow.js version

0.12.4

The example here is not working properly:
https://github.com/tensorflow/tfjs-examples/tree/master/mnist-node

When I add yarn followed by node index.js, I get the following error:

(node:199628) UnhandledPromiseRejectionWarning: TypeError: backend.select is not a function
at environment_1.ENV.engine.runKernel.$condition (/usr/local/google/home/cais/tfjs/tfjs-examples/mnist-node/node_modules/@tensorflow/tfjs-core/dist/ops/logical_ops.js:88:83)
at /usr/local/google/home/cais/tfjs/tfjs-examples/mnist-node/node_modules/@tensorflow/tfjs-core/dist/engine.js:115:26
at Engine.scopedRun (/usr/local/google/home/cais/tfjs/tfjs-examples/mnist-node/node_modules/@tensorflow/tfjs-core/dist/engine.js:95:23)
at Engine.runKernel (/usr/local/google/home/cais/tfjs/tfjs-examples/mnist-node/node_modules/@tensorflow/tfjs-core/dist/engine.js:113:14)
at where_ (/usr/local/google/home/cais/tfjs/tfjs-examples/mnist-node/node_modules/@tensorflow/tfjs-core/dist/ops/logical_ops.js:88:37)
at Object.where (/usr/local/google/home/cais/tfjs/tfjs-examples/mnist-node/node_modules/@tensorflow/tfjs-core/dist/ops/operation.js:23:29)
at Tensor.where (/usr/local/google/home/cais/tfjs/tfjs-examples/mnist-node/node_modules/@tensorflow/tfjs-core/dist/tensor.js:554:26)
at Object.$x (/usr/local/google/home/cais/tfjs/tfjs-examples/mnist-node/node_modules/@tensorflow/tfjs-core/dist/ops/unary_ops.js:113:41)
at Object.backpropagateGradients (/usr/local/google/home/cais/tfjs/tfjs-examples/mnist-node/node_modules/@tensorflow/tfjs-core/dist/tape.js:81:47)
at /usr/local/google/home/cais/tfjs/tfjs-examples/mnist-node/node_modules/@tensorflow/tfjs-core/dist/engine.js:288:20

[caisq]

FYI @nkreeger

[tommedema]

I get the same:

yarn add @tensorflow/tfjs @tensorflow/tfjs-core @tensorflow/tfjs-node

With the following function:

// tslint:disable: no-import-side-effect

import * as tf from '@tensorflow/tfjs'
import '@tensorflow/tfjs-node'

export const predictBinaryCrossEntropy = async (
  xs: number[][], // tslint:disable-line
  ys: number[] // tslint:disable-line
): Promise<void> => {
  const txs = tf.tensor(xs)
  const tys = tf.tensor(ys)

  const xUnits = xs.length * xs[0].length

  const model = tf.sequential()
  model.add(
    tf.layers.dense({
      units: xUnits,
      inputShape: [2],
      activation: 'tanh'
    })
  )
  model.add(tf.layers.dense({ units: 1, activation: 'sigmoid' }))
  model.compile({ optimizer: 'sgd', loss: 'binaryCrossentropy' })

  await model.fit(txs, tys, {
    batchSize: 1,
    epochs: 5000
  })

  const results = model.predict(txs)

  console.log(results)
}

Results in:
TypeError: backend.select is not a function

When run like so:

import { predictBinaryCrossEntropy } from '../src'

describe('predict', () => {
  it('should predict with real xs and binary ys', async () => {
    const xs = [[0, 0], [0, 1], [1, 0], [1, 1]]
    const ys = [0, 1, 1, 0]

    await predictBinaryCrossEntropy(xs, ys)
  })
})

FWIW, also getting a 2018-07-30 17:46:05.489358: I tensorflow/core/platform/cpu_feature_guard.cc:140] Your CPU supports instructions that this TensorFlow binary was not compiled to use: SSE4.2 AVX AVX2 FMA

[caisq]

Also cc @kangyizhang

[kangyizhang]

This should because of some ops updates in tfjs-core KernelBackend. I'm updating NodeJSKernelBakcend to fix it.

[nsthorat]

Just to diagnose this a little more, this is because of the loose versioning here:
https://github.com/tensorflow/tfjs-node/blob/master/package.json#L42

Every time we add a kernel backend and update, you'll break. I wonder if this should be an exact version. cc @dsmilkov

[doomedramen]

What is the temporary fix for this? Use specific tfjs and tfjs-node versions?

[kangyizhang]

It's fixed by tfjs-node 0.1.10