MXNetError: Invalid Parameter format for shape expect Shape(tuple) but value='<Symbol identity1>', in operator Reshape(name="", shape="<Symbol identity1>")

[tom888888]

I use the example from
https://github.com/onnx/tutorials/blob/master/tutorials/PytorchCaffe2SuperResolution.ipynb
to export super_resolution.onnx
and run in MXNET , this error occurs
I did run programming with GPU on Jupyter notebook

Error Message(Import model with MXNET)

MXNetErrorTraceback (most recent call last)
<ipython-input-2-330c0ff72e5d> in <module>()
      1 import onnx_mxnet
      2 from IPython.core.display import display
----> 3 sym, params = onnx_mxnet.import_model('super_resolution.onnx')

/usr/local/lib/python3.6/dist-packages/onnx_mxnet/__init__.py in import_model(model_file)
     33     # loads model file and returns ONNX protobuf object
     34     model_proto = onnx.load(model_file)
---> 35     sym, params = graph.from_onnx(model_proto.graph)
     36     return sym, params

/usr/local/lib/python3.6/dist-packages/onnx_mxnet/import_onnx.py in from_onnx(self, graph)
    139                 op = self._fix_squeeze(inputs, mx_attr)
    140             else:
--> 141                 op = new_op(name=node_name, *inputs, **mx_attr)
    142 
    143             node_output = self._fix_outputs(op_name, node.output)

/usr/local/lib/python3.6/dist-packages/mxnet/symbol/register.py in reshape(data, shape, reverse, target_shape, keep_highest, name, attr, out, **kwargs)

/usr/local/lib/python3.6/dist-packages/mxnet/_ctypes/symbol.py in _symbol_creator(handle, args, kwargs, keys, vals, name)
    123         c_str_array(keys),
    124         c_str_array([str(v) for v in vals]),
--> 125         ctypes.byref(sym_handle)))
    126 
    127     if args and kwargs:

/usr/local/lib/python3.6/dist-packages/mxnet/base.py in check_call(ret)
    144     """
    145     if ret != 0:
--> 146         raise MXNetError(py_str(_LIB.MXGetLastError()))
    147 
    148 

MXNetError: Invalid Parameter format for shape expect Shape(tuple) but value='<Symbol identity1>', in operator Reshape(name="", shape="<Symbol identity1>")

this is my export model code

import io
import numpy as np

from torch import nn
from torch.autograd import Variable
import torch.utils.model_zoo as model_zoo
import torch.onnx 
import torch.nn as nn
import torch.nn.init as init


class SuperResolutionNet(nn.Module):
    def __init__(self, upscale_factor, inplace=False):
        super(SuperResolutionNet, self).__init__()

        self.relu = nn.ReLU(inplace=inplace)
        self.conv1 = nn.Conv2d(1, 64, (5, 5), (1, 1), (2, 2))
        self.conv2 = nn.Conv2d(64, 64, (3, 3), (1, 1), (1, 1))
        self.conv3 = nn.Conv2d(64, 32, (3, 3), (1, 1), (1, 1))
        self.conv4 = nn.Conv2d(32, upscale_factor ** 2, (3, 3), (1, 1), (1, 1))
        self.pixel_shuffle = nn.PixelShuffle(upscale_factor)

        self._initialize_weights()

    def forward(self, x):
        x = self.relu(self.conv1(x))
        x = self.relu(self.conv2(x))
        x = self.relu(self.conv3(x))
        x = self.pixel_shuffle(self.conv4(x))
        return x

    def _initialize_weights(self):
        init.orthogonal(self.conv1.weight, init.calculate_gain('relu'))
        init.orthogonal(self.conv2.weight, init.calculate_gain('relu'))
        init.orthogonal(self.conv3.weight, init.calculate_gain('relu'))
        init.orthogonal(self.conv4.weight)

torch_model = SuperResolutionNet(upscale_factor=3)

model_url = 'https://s3.amazonaws.com/pytorch/test_data/export/superres_epoch100-44c6958e.pth'
batch_size = 1    # just a random number
torch_model.load_state_dict(model_zoo.load_url(model_url))
torch_model.train(False)
x = Variable(torch.randn(batch_size, 1, 224, 224), requires_grad=True)
torch_out = torch.onnx._export(torch_model , x , "super_resolution.onnx" , export_params=True)

this is my import model code

import onnx_mxnet
from IPython.core.display import display
sym, params = onnx_mxnet.import_model('super_resolution.onnx')

my Python Environment

python 3.6.3
apt-get install protobuf-compiler libprotoc-dev
pip install git+https://github.com/onnx/onnx.git@master
and I try to do
pip install onnx-mxnet
pip install Pillow
and
git clone https://github.com/onnx/onnx-mxnet.git
cd onnx-mxnet
sudo python setup.py install
but both of two methods occurred the same error

Could you please advice how to solve out this error
Thank you

[lupesko]

@spidydev to take a look

[rajanksin]

ONNX changed the definition for Reshape op after in the latest master. Pytorch has already incorporated the change in pytorch/caffe export, hence the problem. @anirudhacharya is already working on a solution as MXNET reshape operator doesnt match with new ONNX definition.
There is a related issue raised on Mxnet github: apache/mxnet#10789

[glingyan]

@spidydev Is it possible that I roll back onnx to a old version to bypass this problem

[glingyan]

this is from onnx ONNX_OPERATOR_SET_SCHEMA(
Reshape,
5,
OpSchema()
.SetDoc(Reshape_ver5_doc)
.Input(0, "data", "An input tensor.", "T")
.Input(1, "shape", "Specified shape for output.", "tensor(int64)") //where is the problem?
.Output(0, "reshaped", "Reshaped data.", "T")

[rajanksin]

@glingyan The issue should be resolved in latest MXNet (please install using : pip install mxnet --pre)

The problem was "shape" input was previously an "attribute" in reshape operator(https://github.com/onnx/onnx/blob/master/docs/Changelog.md#Reshape-1). Also, MXNet always expects the "shape" as an attribute not a input symbol. This scenario has been handle in the latest MXNet now (master).

Rolling back ONNX woudn't help as pytorch converted ONNX model will have "shape" as input, so needs handling in import code to be fixed.
Please try with the latest and let us know if the issue persists.

[glingyan]

@spidydev I already use latest mxnet, the Branches: master, remotes/origin/marcoabreu-patch-1, remotes/origin/master
Follows: utils
Precedes:

minor fixes to example/ssd (#11138)

[glingyan]

this is the error I got
File "/home/lingyan/python3_env/pytorch3/lib/python3.5/site-packages/mxnet-1.3.0-py3.5.egg/mxnet/contrib/onnx/_import/import_onnx.py", line 114, in from_onnx
mxnet_sym = self._convert_operator(node_name, op_name, onnx_attr, inputs)
File "/home/lingyan/python3_env/pytorch3/lib/python3.5/site-packages/mxnet-1.3.0-py3.5.egg/mxnet/contrib/onnx/_import/import_onnx.py", line 66, in _convert_operator
mxnet_sym = new_op(*inputs, **new_attrs)
File "", line 116, in reshape
File "/home/lingyan/python3_env/pytorch3/lib/python3.5/site-packages/mxnet-1.3.0-py3.5.egg/mxnet/_ctypes/symbol.py", line 125, in _symbol_creator
ctypes.byref(sym_handle)))
File "/home/lingyan/python3_env/pytorch3/lib/python3.5/site-packages/mxnet-1.3.0-py3.5.egg/mxnet/base.py", line 210, in check_call
raise MXNetError(py_str(_LIB.MXGetLastError()))
mxnet.base.MXNetError: Invalid Parameter format for shape expect Shape(tuple) but value='', in operator Reshape(name="", shape="")

[glingyan]

this is the onnx out
%311 : Float(512) = onnx::Normaxis=1, keepdims=0, ord=2, scope: eNet
%312 : Dynamic = onnx::Constantvalue= 512 1 1 [ CPULongTensor{3} ], scope: eNet
%313 : Float(512, 1, 1) = onnx::Reshape(%311, %312), scope: eNet

[rajanksin]

sorry forgot to mention, as this repo is deprecated , all our fixes goes in the import module inside mxnet. please use http://mxnet.incubator.apache.org/versions/1.2.0/api/python/contrib/onnx.html for reference and usage.

[glingyan]

@spidydev thanks, I already use latest mxnet

[rajanksin]

@glingyan Are you using the import module we added inside latest MXNet.
i.e. import mxnet.contrib.onnx as onnx_mxnet

[glingyan]

yes, my code is simple

"""Testing super_resolution model conversion"""
from future import absolute_import as _abs
from future import print_function
from collections import namedtuple
import logging
import numpy as np

import mxnet as mx
from mxnet.test_utils import download
import mxnet.contrib.onnx as onnx_mxnet

set up logger

logging.basicConfig()
LOGGER = logging.getLogger()
LOGGER.setLevel(logging.INFO)

def import_onnx():
LOGGER.info("Converting onnx format to mxnet's symbol and params...")
sym, arg_params, aux_params = onnx_mxnet.import_model('enet.onnx')
LOGGER.info("Successfully Converted onnx format to mxnet's symbol and params...")
return sym, arg_params, aux_params

if name == 'main':
import_onnx()

[glingyan]

could you tell me which patch in mxnet is to fix this , I will add print to double check

[rajanksin]

the reshape operator convertor (https://github.com/apache/incubator-mxnet/blob/master/python/mxnet/contrib/onnx/_import/op_translations.py#L346) was expected to fix this, but we need to look what is the issue.

[glingyan]

just checked my code , seems do not have this patch , will update and report to you

[glingyan]

still have problem
inputs = {list} <class 'list'>: [, ]
0 = {Symbol}
handle = {c_void_p} c_void_p(220823152)
name = {str} 'norm0'
1 = {Symbol}
handle = {c_void_p} c_void_p(45194928)
name = {str} 'identity0'
len = {int} 2

reshape_shape = list(proto_obj._params[inputs[1].name].asnumpy())
inputs[1].name is identity0, it should be a number

[rajanksin]

Thanks @glingyan for the details. We are looking into the issue.

[lxy5513]

Had the problem already been fixed ?, anyone can tell me how to solve it, Thank you

[lupesko]

@lxy5513 this repo is deprecated, functionality is now built into mxnet.
Please track the issue in MXNet.