TFLite, Model with Conv2DTranspose fails to convert, fully quantization, int8

[wwwind]

System information

• OS Platform and Distribution (e.g., Linux Ubuntu 16.04): Linux
• TensorFlow installed from (source or binary): tf-nightly
• TensorFlow version (or github SHA if from source): tf-nightly

Command used to run the converter or code if you’re using the Python API
If possible, please share a link to Colab/Jupyter/any notebook.

This issue is very similar to the issue, but
the problematic layer is Conv2DTranspose, so it is different model here.
I tested models with other layers and all are fine, except this one and the issue logged above, separately.

https://colab.research.google.com/drive/1g8wjs5D3N9blNpWYMIQ8R_AipZASUKH8?usp=sharing

import numpy as np
import tensorflow as tf

input_size = [5, 5, 2]
kernel_size = [3, 3, 6]
stride = [2, 2]

input_0 = tf.keras.layers.Input(shape=input_size)
layer_0 = tf.keras.layers.Conv2DTranspose(
            filters=kernel_size[-1],
            kernel_size=kernel_size[0:2],
            strides=stride,
            activation=None,
            use_bias=False,
            name = "transpose_conv"
        )(input_0)
model = tf.keras.models.Model(inputs=[input_0], outputs=[layer_0])
model.summary()

keras_layer = [
  layer for layer in model.layers if layer.name == "transpose_conv"
][0]
keras_layer.set_weights(
            [
                np.random.rand(
                    kernel_size[0],
                    kernel_size[1],
                    kernel_size[2],
                    input_size[2],
                ).astype(np.float32)
            ]
        )

num_calib = 1000
def _get_calib_data_func():
  def representative_data_gen():
    for _ in range(num_calib):
      yield [
        np.random.rand(
          1, input_size[0], input_size[1], input_size[2],
        ).astype(np.float32)
      ]

  return representative_data_gen

converter = tf.lite.TFLiteConverter.from_keras_model(model)
converter.representative_dataset = _get_calib_data_func()

converter.target_spec.supported_ops = [tf.lite.OpsSet.TFLITE_BUILTINS_INT8]
tflite_model_INT8 = converter.convert()

The output from the converter invocation

RuntimeError                              Traceback (most recent call last)

<ipython-input-18-717edec90ae0> in <module>()
      3 
      4 converter.target_spec.supported_ops = [tf.lite.OpsSet.TFLITE_BUILTINS_INT8]
----> 5 tflite_model_INT8 = converter.convert()

3 frames

/usr/local/lib/python3.6/dist-packages/tensorflow/lite/python/optimize/calibrator.py in calibrate_and_quantize(self, dataset_gen, input_type, output_type, allow_float, resize_input)
     91     return self._calibrator.QuantizeModel(
     92         np.dtype(input_type.as_numpy_dtype()).num,
---> 93         np.dtype(output_type.as_numpy_dtype()).num, allow_float)
     94 
     95   def calibrate_and_quantize_single(self,

RuntimeError: Max and min for dynamic tensors should be recorded during calibration: Failed for tensor functional_5/transpose_conv/Shape
Empty min/max for tensor functional_5/transpose_conv/Shape

Also, please include a link to the saved model or GraphDef

# Put link here or attach to the issue.

Failure details
If the conversion is successful, but the generated model is wrong,
state what is wrong:

• Producing wrong results and/or decrease in accuracy
• Producing correct results, but the model is slower than expected (model generated from old converter)

RNN conversion support
If converting TF RNN to TFLite fused RNN ops, please prefix [RNN] in the title.

Any other info / logs

Include any logs or source code that would be helpful to diagnose the problem. If including tracebacks, please include the full traceback. Large logs and files should be attached.

[Saduf2019]

I am able to replicate this issue, please find the gist here . Thanks!

[alxhoff]

I am also experiencing a similar problem with a similar model.

[MeghnaNatraj]

The issue seems to be resolved with the latest tf-nightly ('2.3.0-dev20200610') and TF 2.2.0. Feel free to re-open it if you still face this issue.

[google-ml-butler]

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