Create a mobile friendly pytorch model for action recognition, object detection and object tracking

the state of the art pytorch model should have the following capabilities:

  1. all of its layers and operations should be compatible with the native mobile devices (Android and iOS)
  2. for Android: pytorch -> ONNX -> TensorFlow -> tflite
  3. for iOS: pytorch -> coremltools -> coreml
  4. model should able to use hardware acceleration on mobile devices (if the mobile supports)

Any help will be very much appreciated :slightly_smiling_face:
Also, I’ve 2 months’ time frame for this. So, I’ll keep posting my progress here itself (just in case, if it can be of help to anyone).

I have created a simple neural net (for image classification on CIFAR dataset, given on pytorch official docs 60 mins. blitz section). Now, after doing the following conversion:
pytorch -> ONNX -> TensorFlow -> tflite

I’m using following code to convert .pb -> .tflite:

converter = tf.compat.v1.lite.TFLiteConverter.from_frozen_graph(
    graph_def_file="../models/my_cifar_net.pb", input_arrays=["my_input"], output_arrays=["my_output"])

converter.optimizations = [tf.lite.Optimize.DEFAULT]
tf_lite_model = converter.convert()

with'../models/my_cifar_net.tflite', 'wb') as f:

All other converted models (like, .pth, .onnx, .pb) are giving correct output on an example data except .tflite, as I’m getting the following error:

And, I’m using following versions of diff. libraries:

  • numpy==1.18.2
  • pandas==1.0.3
  • torch==1.5.1
  • torchvision==0.6.1
  • onnx==1.7.0
  • coremltools==4.0b1
  • onnxmltools==1.7.0
  • onnx-tf==1.6.0
  • tensorflow==2.2.0
  • tensorflow-addons==0.10.0