[NNVM][FRONTEND][ONNX] Fix the gemm conversion in onnx frontend#1241
Conversation
nishi-t
changed the title
nishi-t
changed the title
| inputs[0] = _sym.transpose(inputs[0], axes=(1, 0)) | ||
| if not transB: | ||
| inputs[1] = _sym.transpose(inputs[1], axes=(1, 0)) | ||
| inputs[0] = _sym.flatten(inputs[0]) |
There was a problem hiding this comment.
Can we verify if flatten is necessary? i.e. check the dimension and only insert the flatten if necessary
There was a problem hiding this comment.
I'm trying it, but I have not already done so well. Because, I'm struggling to get the input data's dimension in the _impl_v1 for checking necessity of flatten. I try to use _infer_shape auxiliary function to get the input[0]'s output_shape. But, it does not work when input[0] is pooling. Could you give me some advices?
There was a problem hiding this comment.
It is strange that you cannot get the output shape when the input[0] is pooling. Maybe can you look into what happens here?
There was a problem hiding this comment.
Thank you for your advice. I analyzed what happened here and found out a few things.
In the stage of nnvm.frontend.from_onnx, maybe infer_shape does not work completly, because the input's shape may not be decided.
When I putted the following code before the flatten and ran it, out_shapes was [[]]
g = _graph.create(inputs[0])
shape_dict = {k: v.shape for k, v in params.items()}
_, out_shapes = graph_util.infer_shape(g, **shape_dict)
In following lines called from above infer_shape, dshape.ndim() became 0 and returned false.
https://github.com/dmlc/tvm/blob/master/nnvm/src/top/nn/pooling.cc#L33
https://github.com/dmlc/tvm/blob/master/nnvm/src/top/nn/convolution.cc#L62
It seems difficult to use infer_shape for my purpose without making any changes. Please let me know what you think.
There was a problem hiding this comment.
infer shape is not available because input shape is not ready. Currently I cannot find a way to detect if flatten is necessary as well.
There was a problem hiding this comment.
OK, maybe an easier way is to add an optimization to remove flatten in nnvm when it is not necessary. I am going to merge this in for now
|
OK, maybe an easier way is to add an optimization to remove flatten in nnvm when it is not necessary. I am going to merge this in for now |
| inputs[0] = _sym.transpose(inputs[0], axes=(1, 0)) | ||
| if not transB: | ||
| inputs[1] = _sym.transpose(inputs[1], axes=(1, 0)) | ||
| inputs[0] = _sym.flatten(inputs[0]) |
There was a problem hiding this comment.
OK, maybe an easier way is to add an optimization to remove flatten in nnvm when it is not necessary. I am going to merge this in for now
3 participants
This PR addressed #1231
I think that the problem caused by passing 4 dimensions no flatten data to
denseoperator. I locally confirmed thatvgg19is compiled successfully. Please review.