I have a some problem with convert AttentionOCR to ONNX. I use this model on TF1.

I use this code for convert ONNX.

import tensorflow as tf
import tf2onnx

saved_model_dir = "convert/aocr-model"
onnx_model_path = "aocr_model.onnx"

with tf.compat.v1.Session(graph=tf.Graph()) as sess:
    tf.compat.v1.saved_model.loader.load(sess, ["serve"], saved_model_dir)

    sess.run(tf.compat.v1.global_variables_initializer())

    output_node_names = ["prediction", "probability"]
    frozen_graph_def = tf.compat.v1.graph_util.convert_variables_to_constants(
        sess,
        sess.graph_def,
        output_node_names
    )


    input_names = ["input_image_as_bytes:0"]
    output_names = ["prediction:0", "probability:0"]

    onnx_model = tf2onnx.convert.from_graph_def(
        frozen_graph_def,
        input_names=input_names,
        output_names=output_names,
        output_path=onnx_model_path
    )

print(f"{onnx_model_path}")

Error:

tensorflow.python.framework.errors_impl.InvalidArgumentError: Node 'cond/ExpandDims' has an _output_shapes attribute inconsistent with the GraphDef for output #0: Shapes must be equal rank, but are 1 and 0

Thanks in advance.

I try use tf2onnx and MO in OpenVINO, but doesnt have succes