Check failed: param.pad_width.ndim() == dshape.ndim() (4 vs. 3)

1

File “/usr/local/lib/python2.7/dist-packages/nnvm-0.8.0-py2.7.egg/nnvm/compiler/graph_util.py”, line 31, in infer_shape
graph = graph.apply(“InferShape”)
File “/usr/local/lib/python2.7/dist-packages/nnvm-0.8.0-py2.7.egg/nnvm/graph.py”, line 234, in apply
check_call(_LIB.NNGraphApplyPasses(self.handle, npass, cpass, ctypes.byref(ghandle)))
File “/usr/local/lib/python2.7/dist-packages/nnvm-0.8.0-py2.7.egg/nnvm/_base.py”, line 75, in check_call
raise NNVMError(py_str(_LIB.NNGetLastError()))
nnvm._base.NNVMError: Error in operator pad27: [16:21:43] /home/zhoukun/FrameWork/nnvm/src/top/nn/nn.cc:576: Check failed: param.pad_width.ndim() == dshape.ndim() (4 vs. 3)

I use tensorflow as backend, keras as frontend, I want to use the nnvm as my compiler to compile the mobilenet model.
this is part of my code .
target = 'llvm’
shape_dict = {‘data’: (1, 3, 224, 224)}

with nnvm.compiler.build_config(opt_level=2):
graph, lib, params = nnvm.compiler.build(sym, target, shape_dict, params=params)

the error occurs in the 4th code

2

I think some where in process we are loosing dimensions to input data.

We use pad operator in convolution and pooling . Ref. https://github.com/dmlc/nnvm/blob/0e00ca3b81d1317bb47dfc75a9ef3ec4684293f9/python/nnvm/frontend/keras.py#L159

Print the shape of insym in all calls to _sym.pad in frontend to find which operator it is.

3

On my environment, the following PR fixes the problem.

1 Like
4

Yes, the above PR should work which make dim back to 4.

5

where did you reshape the data? could you explain this in detail

6

reshape may be an operation before (convolution / pad).

The reshape operation through keras front end was not considering the batch which results in output dim length 3 instead of 4. Hence the pad operator report and error on dimension mismatch.

7

where can I fix this problem

8

I tried the fix reshape solution, and find that the result obtained by NNVM and keras are different, and not right. I used a cat picture and mobileNet.

NNVM top-1 id: 0, class name: tench, Tinca tinca

Keras top-1 id: 688, class name: oscilloscope, scope, cathode-ray oscilloscope, CRO

9

This part require debugging !

You mean on keras the model detected a cat as “oscilloscope, scope, cathode-ray oscilloscope, CRO” ?

I think we first need to check on keras first.

10

And also I found if I use the fix reshape solution, it seem that the nnvm consumes longer time than the pure keras

11

I’ve sent many fixes recently for Keras and I think all of them are necessary for MobileNet.

Can you try the below branch?
https://github.com/kazum/nnvm/tree/keras-mobilenet

It includes all of the patches and test for MobileNet. The test works fine on my environment.

12

OK, thaks, I will try this branch.

13

porblem solved, the fix reshape works

1 Like