TF1-Keras ZeroPadding2D + Conv2D Bias Term Not Fused.

[leimao]

Describe the bug

ZeroPadding2D + Conv2D with bias term cannot be exported as single Conv2D ONNX operator. It is exported as Conv2D + Add instead.

Urgency

We wish it could be fixed asap.

System information

• OS Platform and Distribution (e.g., Linux Ubuntu 16.04): 20.04.2 LTS
• Tensorflow Version: 1.15.5
• Python version: 3.8.10

To Reproduce

In NVIDIA TensorFlow-1 Docker Container:

docker run -it --rm --gpus all -v $(pwd):/mnt nvcr.io/nvidia/tensorflow:21.08-tf1-py3

Install TF2ONNX (tried both the master branch and the latest release 1.9.2)

pip install git+https://github.com/onnx/tensorflow-onnx

Run the following script to generate ONNX file for Conv2DTranspose.

import onnx
import tf2onnx
import tensorflow as tf

inputs = tf.keras.Input(shape=(64, 256, 256))
middles = tf.keras.layers.ZeroPadding2D(
    padding=(0, 4),
    data_format="channels_first",
    name="padding")(inputs)
outputs = tf.keras.layers.Conv2D(
    filters=1,
    kernel_size=(9, 9),
    strides=(1, 1),
    use_bias=True,
    data_format="channels_first",
    name="conv2d",
)(middles)
model = tf.keras.Model(inputs=inputs, outputs=outputs)

onnx_model, _ = tf2onnx.convert.from_keras(
    model=model,
    opset=13,
    large_model=False,
)

onnx.save_model(onnx_model, "conv2d.onnx")

Screenshots

Additional context

N/A

[TomWildenhain-Microsoft]

As before, hopefully someone else will be able to look at this soon, but feel free to do a PR if you want to.