Incubator-mxnet: fail to mxnet to onnx

[alicera]

cause;
I convert mxnet model to onnx ,it is ok.
But when I use
onnx2trt *.onnx -o *.trt
the model is fail

env :
pip install mxnet-cu102
pip install onnx==1.3.0 -i https://pypi.tuna.tsinghua.edu.cn/simple

model ;
https://www.dropbox.com/s/akxeqp99jvsd6z7/model-MobileFaceNet-arcface-ms1m-refine-v1.zip?dl=0
ref : https://github.com/deepinsight/insightface

mxnet to onnx code;

`
import mxnet as mx
import numpy as np
from mxnet.contrib import onnx as onnx_mxnet
sym = './model/model-symbol.json'
params = './model/model-0000.params'
input_shape = (1,3,112,112)
onnx_file = './mxnet_exported_mobileface.onnx'
converted_model_path = onnx_mxnet.export_model(sym, params, [input_shape], np.float32, onnx_file)
print("finish ")

`

onnx2trt *.onnx -o *.trt
Error;
Parsing model While parsing node number 5 [Conv -> "conv_2_dw_conv2d"]: ERROR: /opt/onnx-tensorrt/builtin_op_importers.cpp:463 In function importConv: [8] Assertion failed: nbSpatialDims == kernel_weights.shape.nbDims - 2

[TMVector]

Handy dandy colab notebook reproduces a similar error in ONNX Runtime, implying it's an error in the model or in the exporter.
https://colab.research.google.com/drive/163tkcoYT09Ap31vYotWlhSOshlgWd5V-?usp=sharing

[szha]

@alicera thanks for reporting this and @TMVector thanks for sharing the notebook for reproducing the issue. Would you mind trying out mxnet nightly builds for 1.7 and 2.0 and see if the issue still exists?

depending on the result, @ciyongch may need to include the fix for the upcoming 1.7 release.

[szha]

You can install 1.x nightly builds with

pip install --pre --upgrade 'mxnet<2' -f https://dist.mxnet.io/python

or 2.x nightly builds with

pip install --pre --upgrade 'mxnet' -f https://dist.mxnet.io/python

[TMVector]

Just attaching the model zip to get a stable URL: model-MobileFaceNet-arcface-ms1m-refine-v1.zip

[TMVector]

Thanks for the reply @szha. I got the same error with 1.x and 2.x nightly using the commands you provided (notebook is updated):

RuntimeException: [ONNXRuntimeError] : 6 : RUNTIME_EXCEPTION : Non-zero status code returned while running PRelu node. Name:'conv_1_relu' Status Message: /onnxruntime_src/onnxruntime/core/providers/cpu/math/element_wise_ops.h:351 void onnxruntime::BroadcastIterator::Init(int64_t, int64_t) axis == 1 || axis == largest was false. Attempting to broadcast an axis by a dimension other than 1. 56 by 64

[ciyongch]

As confirmed by @TMVector both 1.x and 2.x had such issue. I'm wondering if there's anyone would help to fix this? Thanks!

[szha]

@sandeep-krishnamurthy is this something your team could help out?

[leezu]

Have you tried onnx 1.5.0? @alicera

[ciyongch]

Hi @sandeep-krishnamurthy @szha @leezu , may I know if this is a block issue for 1.7.0. thanks?

[sandeep-krishnamurthy]

I tested with MXNet 1.5.0 / 1.6.0 and unfortunately, the issue exists there as well.
This is a bug that we need to be fixed. But, I think this is not a blocker for v1.7.0 and can be listed as known issue. Suggestions @szha @leezu ?

[cloudhan]

You can monkey patch the conversion script with

from mxnet.contrib import onnx as onnx_mxnet
from mxnet.contrib.onnx.mx2onnx.export_onnx import MXNetGraph as mx_op
from mxnet.contrib.onnx.mx2onnx._op_translations import get_inputs, get_boolean_attribute_value, parse_helper, convert_string_to_list

# FUCKSHIT: 1d shape of prelu gamma is not unidirectional broadcastable!!!
@mx_op.register("LeakyReLU")
def convert_leakyrelu(node, **kwargs):
    # write your own conversion code here. hint: unsqueeze the gamma

[leezu]

@cloudhan would you like to open a PR with a fix?