I want you to help me! ! The MXNet-Gluon network got an error.

[ys-yamashita]

Platform: Ubuntu 16.04.4 LTS
Python version: Python 3.6.4(Anaconda)
Source framework with version : MXNet 1.0.0 (GPU)
Destination framework with version: IR (I will convert it to Tensorflow code after that)
Pre-trained model path: https://github.com/ys-yamashita/mxnet_gluon
Running scripts: MMdnn/mmdnn/conversion/_script/convertToIR.py

---

• I will try to convert the model created by MXNet (Gluon) to IR, resulting in an error.

Traceback (most recent call last):
File "/root/anaconda3/envs/vision_pose/lib/python3.6/runpy.py", line 193, in _run_module_as_main
"main", mod_spec)
File "/root/anaconda3/envs/vision_pose/lib/python3.6/runpy.py", line 85, in _run_code
exec(code, run_globals)
File "/root/anaconda3/envs/vision_pose/lib/python3.6/site-packages/mmdnn/conversion/_script/convertToIR.py", line 185, in
_main()
File "/root/anaconda3/envs/vision_pose/lib/python3.6/site-packages/mmdnn/conversion/_script/convertToIR.py", line 180, in _main
ret = _convert(args)
File "/root/anaconda3/envs/vision_pose/lib/python3.6/site-packages/mmdnn/conversion/_script/convertToIR.py", line 98, in _convert
parser.run(args.dstPath)
File "/root/anaconda3/envs/vision_pose/lib/python3.6/site-packages/mmdnn/conversion/common/DataStructure/parser.py", line 22, in run
self.gen_IR()
File "/root/anaconda3/envs/vision_pose/lib/python3.6/site-packages/mmdnn/conversion/mxnet/mxnet_parser.py", line 264, in gen_IR
func(current_node)
File "/root/anaconda3/envs/vision_pose/lib/python3.6/site-packages/mmdnn/conversion/mxnet/mxnet_parser.py", line 499, in rename_Convolution
weight = self.weight_data.get(source_node.name + "_weight").asnumpy()
AttributeError: 'NoneType' object has no attribute 'asnumpy'

*The executed command

$ python3 -m mmdnn.conversion._script.convertToIR -f mxnet -n resnet18-stand-json_500-symbol.json -w resnet18-stand-json_500-0000.params -d mxnet_resnext50 --inputShape 3 64 64

The python code that created the model uses Gluon HybridBlocks.
I think that it is probably caused by the fact that the symbol name has the suffix "_fwd" as its origin.

It is reference page of Gluon HybridBlocks
https://gluon.mxnet.io/chapter07_distributed-learning/hybridize.html

Can I convert it to IR?
Help me.

[namizzz]

Hi @ys-yamashita ,
yes, you can change all the source_node.name + "_something" to source_node.name.replace("fwd", "something"). I think it can work for the Gluon format MXNet.

[ys-yamashita]

Thank you for your reply.
All "_fwd" of resnet18-stand-json_500-symbol.json have been replaced with "_something".
However, it will result in an error in the same place.
Is the approach wrong?

Does "source_node.name.replace" need to be inserted somewhere in convertToIR.py?
Is it necessary for Gluon of MXNet?
please tell me.

[namizzz]

Hi @ys-yamashita

eg. change weight = self.weight_data.get(source_node.name + "_weight").asnumpy() to weight = self.weight_data.get(source_node.name.replace("fwd", "weight")).asnumpy()

I have tried and got a (32, 3, 3, 3) weight matrix, it's right.

[ys-yamashita]

Thank you for your advice.

Follow "weight"
"bias" of the rename_Convolution() function
"gamma", " beta", " running_mean", " running_var" of the rename_BatchNorm() function
Also edited.
However, it got an error.

File "/root/anaconda3/envs/vision_pose/lib/python3.6/runpy.py", line 193, in _run_module_as_main
"main", mod_spec)
File "/root/anaconda3/envs/vision_pose/lib/python3.6/runpy.py", line 85, in _run_code
exec(code, run_globals)
File "/root/anaconda3/envs/vision_pose/lib/python3.6/site-packages/mmdnn/conversion/_script/convertToIR.py", line 185, in
_main()
File "/root/anaconda3/envs/vision_pose/lib/python3.6/site-packages/mmdnn/conversion/_script/convertToIR.py", line 180, in _main
ret = _convert(args)
File "/root/anaconda3/envs/vision_pose/lib/python3.6/site-packages/mmdnn/conversion/_script/convertToIR.py", line 98, in _convert
parser.run(args.dstPath)
File "/root/anaconda3/envs/vision_pose/lib/python3.6/site-packages/mmdnn/conversion/common/DataStructure/parser.py", line 22, in run
self.gen_IR()
File "/root/anaconda3/envs/vision_pose/lib/python3.6/site-packages/mmdnn/conversion/mxnet/mxnet_parser.py", line 267, in gen_IR
self.rename_UNKNOWN(current_node)
File "/root/anaconda3/envs/vision_pose/lib/python3.6/site-packages/mmdnn/conversion/mxnet/mxnet_parser.py", line 379, in rename_UNKNOWN
raise NotImplementedError()
NotImplementedError

An error occurred in the if statement

if source_node.type == "null" and source_node.name != 'label':
else:

source_node.type is confirmed by print () "relu"
source_node.name is "resnet0_residual0_relu0"

"relu" is used within hybrid_foward(MXNet-Gluon)

def hybrid_forward(self, F, x):
    out = F.relu(self.bn1(self.conv1(x)))
    out = self.bn2(self.conv2(out))
    if not self.same_shape:
        x = self.conv3(x)
    return F.relu(out + x)

I do not know what to do. Help me

[namizzz]

Yes, you've got this!

You also need to change some code in mxnet_parser:

1. Add rename_relu and rename_Reshape

def rename_relu(self, source_node):
        self._convert_identity_operation(source_node, new_op="Relu")

def rename_Reshape(self, source_node):
        self.rename_Flatten(source_node)

2. follow the former way to change other weights and bias

[ys-yamashita]

Thank you very much. I was able to create an IR. Appreciation

[mifan0208]

Hello, I have another problem. How did you solve it
Traceback (most recent call last):
File "/usr/local/lib/python3.6/runpy.py", line 193, in _run_module_as_main
"main", mod_spec)
File "/usr/local/lib/python3.6/runpy.py", line 85, in _run_code
exec(code, run_globals)
File "/usr/local/lib/python3.6/site-packages/mmdnn/conversion/_script/convertToIR.py", line 197, in
_main()
File "/usr/local/lib/python3.6/site-packages/mmdnn/conversion/_script/convertToIR.py", line 192, in _main
ret = _convert(args)
File "/usr/local/lib/python3.6/site-packages/mmdnn/conversion/_script/convertToIR.py", line 80, in _convert
parser = MXNetParser(model)
File "/usr/local/lib/python3.6/site-packages/mmdnn/conversion/mxnet/mxnet_parser.py", line 243, in init
self.model, self.weight_data = MXNetParser._load_model(input_arg[1], input_arg[2])
File "/usr/local/lib/python3.6/site-packages/mmdnn/conversion/mxnet/mxnet_parser.py", line 176, in _load_model
sym, arg_params, aux_params = mx.model.load_checkpoint(weights, int(epoch))
File "/usr/local/lib/python3.6/site-packages/mxnet/model.py", line 420, in load_checkpoint
symbol = sym.load('%s-symbol.json' % prefix)
File "/usr/local/lib/python3.6/site-packages/mxnet/symbol/symbol.py", line 2494, in load
check_call(_LIB.MXSymbolCreateFromFile(c_str(fname), ctypes.byref(handle)))
File "/usr/local/lib/python3.6/site-packages/mxnet/base.py", line 146, in check_call
raise MXNetError(py_str(_LIB.MXGetLastError()))
mxnet.base.MXNetError: Cannot find argument 'layout', Possible Arguments:

global_pool : boolean, optional, default=0
Ignore kernel size, do global pooling based on current input feature map.
cudnn_off : boolean, optional, default=0
Turn off cudnn pooling and use MXNet pooling operator.
kernel : Shape(tuple), required
pooling kernel size: (y, x) or (d, y, x)
pool_type : {'avg', 'max', 'sum'}, required
Pooling type to be applied.
pooling_convention : {'full', 'valid'},optional, default='valid'
Pooling convention to be applied.
stride : Shape(tuple), optional, default=[]
stride: for pooling (y, x) or (d, y, x)
pad : Shape(tuple), optional, default=[]
pad for pooling: (y, x) or (d, y, x)
, in operator Pooling(name="pool0_fwd", stride="(1, 1)", pooling_convention="full", pool_type="avg", pad="(0, 0)", layout="NCHW", kernel="(1, 1)", global_pool="True")

[simomagi]

Yes, you've got this!

You also need to change some code in mxnet_parser:

1. Add rename_relu and rename_Reshape

def rename_relu(self, source_node):
        self._convert_identity_operation(source_node, new_op="Relu")

def rename_Reshape(self, source_node):
        self.rename_Flatten(source_node)

1. follow the former way to change other weights and bias

Yes, you've got this!

You also need to change some code in mxnet_parser:

1. Add rename_relu and rename_Reshape

def rename_relu(self, source_node):
        self._convert_identity_operation(source_node, new_op="Relu")

def rename_Reshape(self, source_node):
        self.rename_Flatten(source_node)

1. follow the former way to change other weights and bias

I have the same issue; I have not understood what is the second step. After adding methods rename_relu, rename_Reshape, what should i do?
My issue is little different. The NotImplementedError is raised but the print statement in the function RENAME_UNKNOWN() says:
"Warning: MXNet Parser has not supported operator clip with name mobilenetv20_features_relu60_relu6"

[XiaoXYe]

Hi @simomagi , MMdnn has not supported the operator 'clip' in mxnet model, we will add this operator in the future.