[FRONTEND][MXNET] L2Normalization is not supported in nnvm

[ndcuong91]

Hi all. I got this error when i run model VGG_VOC0712_SSD_300x300 in mxnet through command
nnvm_sym, nnvm_params = nnvm.frontend.from_mxnet(mx_sym, args, auxs)

Error is:
File "/home/prdcv181/PycharmProjects/darknet_nnvm/Phase2/from_mxnet.py", line 86, in
nnvm_sym, nnvm_params = nnvm.frontend.from_mxnet(mx_sym, args, auxs)
File "/usr/local/lib/python2.7/dist-packages/nnvm-0.8.0-py2.7.egg/nnvm/frontend/mxnet.py", line 371, in from_mxnet
sym = _from_mxnet_impl(symbol, {})
File "/usr/local/lib/python2.7/dist-packages/nnvm-0.8.0-py2.7.egg/nnvm/frontend/mxnet.py", line 334, in _from_mxnet_impl
childs = [_from_mxnet_impl(childs[i], graph) for i in range(len(childs.list_outputs()))]
File "/usr/local/lib/python2.7/dist-packages/nnvm-0.8.0-py2.7.egg/nnvm/frontend/mxnet.py", line 334, in _from_mxnet_impl
childs = [_from_mxnet_impl(childs[i], graph) for i in range(len(childs.list_outputs()))]
File "/usr/local/lib/python2.7/dist-packages/nnvm-0.8.0-py2.7.egg/nnvm/frontend/mxnet.py", line 334, in _from_mxnet_impl
childs = [_from_mxnet_impl(childs[i], graph) for i in range(len(childs.list_outputs()))]
File "/usr/local/lib/python2.7/dist-packages/nnvm-0.8.0-py2.7.egg/nnvm/frontend/mxnet.py", line 334, in _from_mxnet_impl
childs = [_from_mxnet_impl(childs[i], graph) for i in range(len(childs.list_outputs()))]
File "/usr/local/lib/python2.7/dist-packages/nnvm-0.8.0-py2.7.egg/nnvm/frontend/mxnet.py", line 334, in _from_mxnet_impl
childs = [_from_mxnet_impl(childs[i], graph) for i in range(len(childs.list_outputs()))]
File "/usr/local/lib/python2.7/dist-packages/nnvm-0.8.0-py2.7.egg/nnvm/frontend/mxnet.py", line 334, in _from_mxnet_impl
childs = [_from_mxnet_impl(childs[i], graph) for i in range(len(childs.list_outputs()))]
File "/usr/local/lib/python2.7/dist-packages/nnvm-0.8.0-py2.7.egg/nnvm/frontend/mxnet.py", line 334, in _from_mxnet_impl
childs = [_from_mxnet_impl(childs[i], graph) for i in range(len(childs.list_outputs()))]
File "/usr/local/lib/python2.7/dist-packages/nnvm-0.8.0-py2.7.egg/nnvm/frontend/mxnet.py", line 334, in _from_mxnet_impl
childs = [_from_mxnet_impl(childs[i], graph) for i in range(len(childs.list_outputs()))]
File "/usr/local/lib/python2.7/dist-packages/nnvm-0.8.0-py2.7.egg/nnvm/frontend/mxnet.py", line 334, in _from_mxnet_impl
childs = [_from_mxnet_impl(childs[i], graph) for i in range(len(childs.list_outputs()))]
File "/usr/local/lib/python2.7/dist-packages/nnvm-0.8.0-py2.7.egg/nnvm/frontend/mxnet.py", line 336, in _from_mxnet_impl
node = _convert_symbol(op_name, childs, attr)
File "/usr/local/lib/python2.7/dist-packages/nnvm-0.8.0-py2.7.egg/nnvm/frontend/mxnet.py", line 295, in _convert_symbol
_raise_not_supported('Operator: ' + op_name)
File "/usr/local/lib/python2.7/dist-packages/nnvm-0.8.0-py2.7.egg/nnvm/frontend/mxnet.py", line 24, in _raise_not_supported
raise NotImplementedError(err)
NotImplementedError: Operator: L2Normalization is not supported in nnvm.

I checked that mxnet in nnvm now only support those layer:
'Activation' : _activations,
'BatchNorm' : _batch_norm,
'BatchNorm_v1' : _batch_norm,
'Cast' : _rename('cast'),
'Concat' : _concat,
'Convolution' : _conv2d,
'Convolution_v1': _conv2d,
'Deconvolution' : _conv2d_transpose,
'Dropout' : _dropout,
'Flatten' : _rename('flatten'),
'FullyConnected': _dense,
'LeakyReLU' : _leaky_relu,
'Pooling' : _pooling,
'Pooling_v1' : _pooling,
'Reshape' : _reshape,
'SliceChannel' : _split,
'split' : _split,
'Softmax' : _rename('softmax'),
'SoftmaxOutput' : _softmax_output,
'concat' : _concat,
'max_axis' : _rename('max'),
'min_axis' : _rename('min'),
'reshape' : _reshape,
'sum_axis' : _rename('sum'),
'UpSampling' : _upsampling,
'clip' : _clip

So, how can i work around with this issue?

[PariksheetPinjari909]

L2Normaization is supported in tvm. I have raised PR #1157 to support it in NNVM. Waiting for review. @tqchen Please review the PR #1157 so we will not face these issues.

[ndcuong91]

@PariksheetPinjari909 i see. But it seem to be an error during test phase. I hope to see this feature as soon as possible

[tqchen]

#1157

[ifeherva]

I just built TVM from source, yet I am getting the same error (L2Normalization is not supported in nnvm). My model is from mxnet. What symbol version is needed for this to work?

[yjcn]

@ifeherva @titikid I meet the same problem. Is the L2Normalization op still not suppoted in nnvm?