Tensorflow lite INT8 Convert Issue

1. System information

Ubuntu 18.04
pip3 install tensorflow-gpu==2.2.0
pip3 install tensorflow-addons==0.10.0
onnx == 1.7.0

onnx model link (https://github.com/onnx/models/blob/master/vision/classification/resnet/model/resnet50-v1-7.onnx)

Conversion works well, but errors occur when performing.

2. Code

onnx -> pb

onnx_model = onnx.load("./onnx_resnet/resnet50-v1-7.onnx")
tf_model_path = "./onnx_resnet/tf_2.2.0/tflite/saved_model"
tf_rep = prepare(onnx_model)
tf_rep.export_graph(tf_model_path)

pb -> tflite

def representative_data_gen():

a = []
datapath = "/MSCOCO/images/test2017"
file_list = os.listdir(datapath)
pixel_mean = (0.485, 0.456, 0.406)
pixel_std = (0.229, 0.224, 0.225)

for i in range(100):
    file_name = file_list[i]
    img = cv2.imread(os.path.join(datapath, file_list[i]))
    img = cv2.resize(img, (224, 224))        
    img = (img - pixel_mean) / pixel_std
    img = img.astype(np.float32)
    a.append(img)
a = np.array(a)
img = tf.data.Dataset.from_tensor_slices(a).batch(1)
for i in img.take(1):
    yield [i]

tf_model_path = "./onnx_resnet/tf_2.2.0/tflite/saved_model"
tflite_model_path = "./tensorflow_resnet/tf.2.2.0/tflite/resnet_int8.tflite"

converter = tf.lite.TFLiteConverter.from_saved_model(tf_model_path)
converter.target_spec.supported_ops = [tf.lite.OpsSet.TFLITE_BUILTINS_INT8]
converter.representative_dataset=representative_data_gen
tflite_model = converter.convert()

with open(tflite_model_path, 'wb') as f:
f.write(tflite_model)

3. Failure after conversion

TFlite Test Code

interpreter = tf.lite.Interpreter(model_path="./tensorflow_resnet/tf.2.2.0/tflite/resnet_int8.tflite")
interpreter.allocate_tensors()

Error Log

RuntimeError: tensorflow/lite/kernels/pad.cc:106 op_context.input->type != op_context.constant_values->type (9 != 1)Node number 7 (PADV2) failed to prepare.

Reproduce the conversion using representative_data_gen and the TFLiteConverter.from_saved_model flow, then load the generated model with tf.lite.Interpreter. Start with the reported PADV2 failure in tensorflow/lite/kernels/pad.cc and compare the input and constant-value types involved. Done means the converted INT8 model allocates successfully without the reported PADV2 error.