Tools to convert ONNX files (NCHW) to TensorFlow format (NHWC)

• Improvement of tflite's I/O OP name rewriting process and I/O order control process

  • Writing input/output information to tflite as signature_defs.
  • Rewrite binaries so that input/output names match ONNX using concrete_function.

  • -coion / --copy_onnx_input_output_names_to_tflite output tflite with options so that the signature is embedded in the tflite by default.

    • onnx

    • tflite

      https://github.com/PINTO0309/onnx2tf/releases/download/1.1.28/version-RFB-640.onnx

      onnx2tf -i version-RFB-640.onnx -coion
      

      import numpy as np
      import tensorflow as tf
      
      interpreter = tf.lite.Interpreter(
        model_path="saved_model/version-RFB-640_float32.tflite"
      )
      tf_lite_model = interpreter.get_signature_runner()
      test_data = tf.convert_to_tensor(
        value=np.ones(
            shape=[1,480,640,3],
            dtype=np.float32
        )
      )
      tt_lite_output = tf_lite_model(
        input=test_data,
      )
      print("[TFLite] Model Predictions:", tt_lite_output)
      

      [TFLite] Model Predictions:
      {'boxes': array([[[ 1.3372814e-04,  8.5677393e-04,  1.3687836e-02,  2.6055152e-02],
            [-2.9060002e-03, -1.7406903e-03,  1.7976686e-02,  4.0702272e-02],
            [-5.9925802e-03, -8.2277283e-03,  2.4591457e-02,  5.8303714e-02],
            ...,
            [ 8.6583459e-01,  8.3152449e-01,  1.0436372e+00,  1.0488608e+00],
            [ 8.2547259e-01,  7.7981573e-01,  1.0946229e+00,  1.1140716e+00],
            [ 7.7743149e-01,  7.2188163e-01,  1.1553301e+00,  1.1767489e+00]]],
          dtype=float32),
      'scores': array([[[0.9121739 , 0.08782604],
            [0.9121672 , 0.0878328 ],
            [0.9121579 , 0.08784208],
            ...,
            [0.9487316 , 0.05126836],
            [0.96191823, 0.03808172],
            [0.96088   , 0.03912005]]], dtype=float32)}
      

  • Ref: https://www.tensorflow.org/lite/guide/signatures
• [TODO] Improvement of tflite's I/O OP name rewriting process and I/O order control process #228
• ONNX model input & output names are not preserved in TensorFlow & TensorFlow Lite models #178

• -coion / --copy_onnx_input_output_names_to_tflite output tflite with options so that the signature is embedded in the tflite by default. by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/229

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.7.17...1.7.18