PoseNet models don't accept batch sizes greater than 1.

[oveddan]

I'm writing some custom code that is to feed a batch of images into the PoseNet models for inference; in my case its a series of frames from long videos. I would eventually open a PR that supports this functionality.

However, when feeding a batch size of greater than one to both the MobileNet and ResNet models, you get an error:

Error: The shape of dict['sub_2'] provided in model.execute(dict) must be [1,-1,-1,3], but was [5,537,953,3]

A way to simply reproduce the error source is:

  const url =
      'https://storage.googleapis.com/tfjs-models/savedmodel/posenet/resnet50/quant1/model-stride16.json'
  const graphModel = await tf.loadGraphModel(url);

  const inputTensor = tf.zeros([5, 100, 100, 3]);

  const result = graphModel.predict(inputTensor);

The above code results in an error:

UnhandledPromiseRejectionWarning: Error: The shape of dict['sub_2'] provided in model.execute(dict) must be [1,-1,-1,3], but was [5,100,100,3]

The same thing occurs for other model urls.

CC @tylerzhu-github

[rthadur]

CC @tylerzhu-github gentle ping to check this error.

[tylerzhu-github]

Thanks for the reminder!
Yes that will require exporting a model that support batch size > 1.
Will handle that in the next release, timeline would be end of next quarter.

[oveddan]

@tylerzhu-github thanks! Does this require re-training the model or just a modification to how the model is exported?

[rthadur]

Will be closing this issue as we will be using latest body segmentation model for same purpose.
Please check latest body segmentation model for latest bodypix model demo https://github.com/tensorflow/tfjs-models/tree/master/body-segmentation

[google-ml-butler]

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