"response.arrayBuffer is not a function" while loadLayersModel (Node.js)

[mrqo]

TensorFlow.js version

1.2.9

Browser version

any

Describe the problem or feature request

I have exported a Keras model written in Python, then converted it with tensorflowjs_converter and uploaded it to server. Then, I get error while loading it. I have tried it with @tensorflow/tfjs-node' as well as @tensorflow/tfjs'.

TypeError: response.arrayBuffer is not a function at f:\Developer\x\node_modules\@tensorflow\tfjs-core\dist\io\weights_loader.js:97:90 at Array.map (<anonymous>) at Object.<anonymous> (f:\Developer\x\node_modules\@tensorflow\tfjs-core\dist\io\weights_loader.js:97:48) at step (f:\Developer\x\node_modules\@tensorflow\tfjs-core\dist\io\weights_loader.js:48:23) at Object.next (f:\Developer\x\node_modules\@tensorflow\tfjs-core\dist\io\weights_loader.js:29:53) at fulfilled (f:\Developer\x\node_modules\@tensorflow\tfjs-core\dist\io\weights_loader.js:20:58) at process._tickCallback (internal/process/next_tick.js:68:7)

Code to reproduce the bug / link to feature request

File is publicly available, bin file also resides there.
const tf = require('@tensorflow/tfjs-node');
model = await tf.loadLayersModel('https://storage.googleapis.com/valid-ship-252510.appspot.com/1568271407928-tfjs_model_1/model.json');

[rthadur]

@kqtus getting a CORS exception while loading , can you please check.

[mrqo]

@rthadur hmm, strange. I tried loadLayersModel and through Postman on both model.json and bin on another PC, and files seem to download without CORS exceptions.

-- Edit 1.
IDK if that'll help, but I logged the response I get - no arrayBuffer() indeed.

-- Edit 2.
I bypassed this problem by downloading all files manually and providing file:// url.

[mrqo]

Any estimations on when it will be fixed?

From what I understand, Node.js doesn't support arrayBuffer(). You'd need to change this:

weights_loader.ts, line 60:
const bufferPromises = responses.map(response => response.arrayBuffer());

to this:
const bufferPromises = responses.map(response => response.buffer());

[rthadur]

@kangyizhang any thoughts on above comments from @kqtus ?

[mrqo]

Ok I think I solved it the proper way, having replaced global fetch.

const nodeFetch = require('node-fetch');
global.fetch = nodeFetch;

It is strange though, because previously I tried passing it as a fetchFunc, but it didn't trigger.

[rthadur]

thanks for the update , closing this issue.

[kangyizhang]

tf.loadLayersModel() should automatically use node-fetch when it's in Node environment. @dsmilkov any thoughts why .arrayBuffer() is not available?

[dsmilkov]

@kqtus We'd like to understand why you had to manually do node-fetch. In Node.js, we do override the global fetch internally so the user doesn't have to. I'm unable to reproduce the bug.

Can you provide a bit more info:

• What version of node you are using?
• We don't seem to have access to the model.json on GCS (Access denied). If you can share the model files publicly, that would be great.
• Can you share the error you are getting by simply calling tf.loadLayersModel(url), without specifying any options like fetchFunc?

Thank you!

[dsmilkov]

Another idea. Can you double-check the node-fetch version you had installed before? It is possible that you had manually installed node-fetch 1.x which lacks body.arraybuffer(), while tfjs depends on node-fetch 2.x.

[mrqo]

@dsmilkov

• npm version: 6.9.0
• I deleted the previous model. I have another one, though:
  https://storage.googleapis.com/valid-ship-252510.appspot.com/1568368482017-keras_dogs_cats_model/model.json
• Here's the error log https://pastebin.com/z64J4KEj
• I didn't have node-fetch installed prior to fixing the problem by manual assignment. Having it installed, without overriding global fetch function still doesn't solve the case.

[kangyizhang]

@kqtus can you provide your node version and which operating system you are using?

[mrqo]

@kangyizhang Node v10.16.3, Windows 10 1903. Here's also package.json: https://pastebin.com/swpTke2q.

One additional thing came to my mind - we are having frontend app and backend in the same project (Express + React). Frontend is being transpiled and backend is run as-is - maybe that causes ambiguity.