model works in nodejs and fails in browser

model works in nodejs and fails in browser

1. saved_model with tfjs_node in nodejs:

  const model = tf.node.loadSavedModel(modelPath);
  const result = await model.predict(image);

pass no issues

1. graph_model with tfjs_node in nodejs:

  const model = tf.loadGraphModel(modelPath);
  const result = await model.predict(image);

fail: error due to dynamic ops

This execution contains the node 'filtered_detections/map/while/Exit_4', which has the dynamic op 'Exit'

why? it’s the same model, just converted from saved_model to graph_model using tensorflowjs_convert in a simplest possible way

ok, lets try again with executeAsync

  const model = tf.loadGraphModel(modelPath);
  const result = await model.executeAsync(image);

pass now it works

1. graph_model with tfjs in browser using webgl backend

  const model = tf.loadGraphModel(modelPath);
  const result = await model.executeAsync(image);

fail no matter what, i cannot get this to work in browser (same model and code that works in nodejs)

this is a typical error:

Error: Size(442368) must match the product of shape 65504,4
  id	@	util.js:281
  oS	@	Reshape.js:26
  kernelFunc	@	engine.js:431
  (anonymous)	@	engine.js:483
  scopedRun	@	engine.js:324
  runKernelFunc	@	engine.js:481
  reshape_	@	reshape.js:58
  reshape__op	@	operation.js:44
  executeOp$g	@	transformation_executor.js:34
  (anonymous)	@	operation_executor.js:78
  (anonymous)	@	engine.js:314
  scopedRun	@	engine.js:324
  tidy	@	engine.js:313
  tidy	@	globals.js:173
  (anonymous)	@	operation_executor.js:78
  executeOp$h	@	operation_executor.js:90
  processStack	@	graph_executor.js:390
  executeWithControlFlow	@	graph_executor.js:350
  _executeAsync	@	graph_executor.js:285
  executeAsync	@	graph_executor.js:257
  executeAsync	@	graph_model.js:304

environment: tfjs 2.6.0 on ubuntu 20.10

[pyu10055]

@vladmandic TensorFlow.js does not dispose GL texture by default, the reason we want to reuse the textures for consecutive runs. If the model is very large and generate a huge amount of intermediate textures, this would cause GL OOM. We do have a flag to enable aggressive texture disposal https://github.com/tensorflow/tfjs/blob/038522b8f228962c1cc0d179a8e445df8ea419a7/tfjs-backend-webgl/src/flags_webgl.ts#L178 You can set the WEBGL_DELETE_TEXTURE_THRESHOLD in bytes to trigger texture disposal if the total texture size exceed this limit.

[google-ml-butler[bot]]

Are you satisfied with the resolution of your issue? Yes No