can't convert a simple LeNet from Pytorch even after removing pooling layers.

[alirezag]

I have a very simple very standard LeNet in pytorch and I can't convert it to webdnn. The pooling layer give an error with dilation attribute. even after removing them and leaving just convolution I get a similar error but this time from attribute transA.

File "", line 1, in
runfile('/Users/alirezagoudarzi/github/Aya-s-Brain/WebDNN_Test/pytorch/train_model.py', wdir='/Users/alirezagoudarzi/github/Aya-s-Brain/WebDNN_Test/pytorch')

File "/Users/alirezagoudarzi/anaconda/envs/py36/lib/python3.6/site-packages/spyder/utils/site/sitecustomize.py", line 705, in runfile
execfile(filename, namespace)

File "/Users/alirezagoudarzi/anaconda/envs/py36/lib/python3.6/site-packages/spyder/utils/site/sitecustomize.py", line 102, in execfile
exec(compile(f.read(), filename, 'exec'), namespace)

File "/Users/alirezagoudarzi/github/Aya-s-Brain/WebDNN_Test/pytorch/train_model.py", line 142, in
graph = PyTorchConverter().convert(net, dummy_input)

File "/Users/alirezagoudarzi/anaconda/envs/py36/lib/python3.6/site-packages/webdnn-1.2.3-py3.6.egg/webdnn/frontend/pytorch/converter.py", line 92, in convert
graph = ONNXConverter().convert(onnx.load(proto_path))

File "/Users/alirezagoudarzi/anaconda/envs/py36/lib/python3.6/site-packages/webdnn-1.2.3-py3.6.egg/webdnn/frontend/onnx/converter.py", line 94, in convert
self._convert_operator(onnx_op)

File "/Users/alirezagoudarzi/anaconda/envs/py36/lib/python3.6/site-packages/webdnn-1.2.3-py3.6.egg/webdnn/frontend/onnx/converter.py", line 114, in _convert_operator
super(ONNXConverter, self)._convert_operator(proto)

File "/Users/alirezagoudarzi/anaconda/envs/py36/lib/python3.6/site-packages/webdnn-1.2.3-py3.6.egg/webdnn/frontend/converter.py", line 117, in _convert_operator
self._handler_map[self.class.name][operator_key](self, operator)

File "/Users/alirezagoudarzi/anaconda/envs/py36/lib/python3.6/site-packages/webdnn-1.2.3-py3.6.egg/webdnn/frontend/onnx/defs/math.py", line 370, in _convert_gemm
y, = Tensordot(None, axes=(A.order.axes[0 if attrs["transA"].i else 1], B.order.axes[1 if attrs["transB"].i else 0]))(A, B)

KeyError: 'transA'

[milhidaka]

It seems to be some unexpected condition.
Could you upload model definition and training code?

[alirezag]

Thanks @milhidaka Here is my model definition and training code. I've commented out the webdnn imports and code from this on the top and the bottom.

train_model.zip

[milhidaka]

Thank you for uploading the code. I will investigate it.

[milhidaka]

Some attributes may be omitted and have default value in their specification.
I tried to fix the error by the following patch, but failed.

diff --git a/src/graph_transpiler/webdnn/frontend/onnx/defs/math.py b/src/graph_transpiler/webdnn/frontend/onnx/defs/math.py
index a02eaee9..9f8945ed 100644
--- a/src/graph_transpiler/webdnn/frontend/onnx/defs/math.py
+++ b/src/graph_transpiler/webdnn/frontend/onnx/defs/math.py
@@ -367,7 +367,9 @@ def _convert_gemm(converter: ONNXConverter, onnx_op: INodeProto):
     beta = attrs["beta"].f
     broadcast = attrs.get("broadcast", 0)
 
-    y, = Tensordot(None, axes=(A.order.axes[0 if attrs["transA"].i else 1], B.order.axes[1 if attrs["transB"].i else 0]))(A, B)
+    transA = attrs["transA"].i if "transA" in attrs else 0
+    transB = attrs["transB"].i if "transB" in attrs else 0
+    y, = Tensordot(None, axes=(A.order.axes[0 if transA else 1], B.order.axes[1 if transB else 0]))(A, B)
 
     if broadcast:
         check_broadcast_constraints(y, C)

Along with the patch, another is needed in my environment:

diff --git a/src/graph_transpiler/webdnn/frontend/onnx/defs/nn.py b/src/graph_transpiler/webdnn/frontend/onnx/defs/nn.py
index 0f6d55a4..a2c75999 100644
--- a/src/graph_transpiler/webdnn/frontend/onnx/defs/nn.py
+++ b/src/graph_transpiler/webdnn/frontend/onnx/defs/nn.py
@@ -50,9 +50,9 @@ def _convert_max_pool(converter: ONNXConverter, onnx_op: INodeProto):
 
     attrs = attribute_dict(onnx_op)
     ksize = list(attrs["kernel_shape"].ints)
-    dilations = list(attrs["dilations"].ints)
-    if any(d != 1 for d in dilations):
-        raise NotImplementedError("[ONNXConverter] MaxPool is supported only when dilations are 1.")
+    # dilations = list(attrs["dilations"].ints)
+    # if any(d != 1 for d in dilations):
+    #     raise NotImplementedError("[ONNXConverter] MaxPool is supported only when dilations are 1.")
 
     stride = list(attrs["strides"].ints)

The error:

Traceback (most recent call last):
  File "myconvert.py", line 17, in <module>
    graph = PyTorchConverter().convert(net, dummy_input)
  File "/Users/hidaka/miljs/mil-web-dnn/webdnn/src/graph_transpiler/webdnn/frontend/pytorch/converter.py", line 92, in convert
    graph = ONNXConverter().convert(onnx.load(proto_path))
  File "/Users/hidaka/miljs/mil-web-dnn/webdnn/src/graph_transpiler/webdnn/frontend/onnx/converter.py", line 94, in convert
    self._convert_operator(onnx_op)
  File "/Users/hidaka/miljs/mil-web-dnn/webdnn/src/graph_transpiler/webdnn/frontend/onnx/converter.py", line 114, in _convert_operator
    super(ONNXConverter, self)._convert_operator(proto)
  File "/Users/hidaka/miljs/mil-web-dnn/webdnn/src/graph_transpiler/webdnn/frontend/converter.py", line 117, in _convert_operator
    self._handler_map[self.__class__.__name__][operator_key](self, operator)
  File "/Users/hidaka/miljs/mil-web-dnn/webdnn/src/graph_transpiler/webdnn/frontend/onnx/defs/math.py", line 372, in _convert_gemm
    y, = Tensordot(None, axes=(A.order.axes[0 if transA else 1], B.order.axes[1 if transB else 0]))(A, B)
  File "/Users/hidaka/miljs/mil-web-dnn/webdnn/src/graph_transpiler/webdnn/graph/operators/tensordot.py", line 119, in __call__
    """
AssertionError: 
[Tensordot] Reduction size of "A" and "B" must be same:
    (A) = <Variable Variable9 shape=(1, 360), order=[?31, ?32]>
    (B) = <ConstantVariable ConstantVariable4 shape=(10, 250), order=[?18, ?19]>
    (axes) = ((<Axis ?32>,), (<Axis ?19>,))
    (reduction size of A) = 360
    (reduction size of B) = 250

There is another bug which produces incorrect shape variable. I will investigate further.

[milhidaka]

Fixed pooling, convolution error in #813 . Please try master branch. (git clone && python setup.py develop)

[alirezag]

Thanks a bunch. I'll check it out.