David Fan

Retinanet conversion is not supported right now.

The example of converting keras mask-rcnn is [here](https://github.com/onnx/keras-onnx/blob/master/applications/nightly_build/test_mask_rcnn.py) and running in our nightly build every day. The model has a custom conversion example: `set_converter('CropAndResize', convert_tf_crop_and_resize)`.

To load your custom weights, just change `model.load_weights` [here](https://github.com/onnx/keras-onnx/blob/master/applications/mask_rcnn/mask_rcnn.py#L64). `test_utils` is in our source code, please search inside our repo.

Is there any model that need this tf.map_fn support? The goal is not to convert every tensorflow op, but the ones that is needed in real models.

Can you try the model in model zoo [here](https://github.com/onnx/models/tree/master/vision/object_detection_segmentation/yolov3) to see if it can work?

@XuDongHecs can you post this issue to [this](https://github.com/onnx/onnx-tensorrt) repo? seems more related there.

We have transformers conversion in our nightly build [here](https://github.com/onnx/keras-onnx/blob/master/applications/nightly_build/test_transformers.py), would your model be one of them? Need tensorflow 2.2.0+. Which keras2onnx you use? try the code from source.

Looks like your input data dimension does not match the model input dimension. Since the conversion is successful, you can see our test examples to write simple test for the...

Please take a look at the onnx model to see if the Resize op has 4 or 2 inputs. If it does not, then it is the keras2onnx issue, then...

We use ConvBatchNormOptimizer in onnxconverter-common to fuse conv+batchnorm, see the line [here](https://github.com/microsoft/onnxconverter-common/blob/master/onnxconverter_common/optimizer.py#L858)