Relax frontend Reshape issue

Development unity
#1

I’m trying to load onnx model with relax.from_onnx() with the latest unity branch(c0b8953), got the following error while constructing the Reshape node, dose anyone met it before? and perhaps solution?

    5: tvm::relax::Normalizer::Normalize(tvm::RelayExpr const&)
          at /path/tvm/src/relax/ir/block_builder.cc:468
    4: tvm::relax::Normalizer::VisitExpr(tvm::RelayExpr const&)
          at /path/tvm/src/relax/ir/block_builder.cc:544
    3: tvm::relax::Normalizer::VisitExpr_(tvm::relax::CallNode const*)
          at /path/tvm/src/relax/ir/block_builder.cc:599
    2: tvm::relax::Normalizer::InferStructInfo(tvm::relax::Call const&)
          at /path/tvm/src/relax/ir/block_builder.cc:757
    1: tvm::relax::InferStructInfoReshape(tvm::relax::Call const&, tvm::relax::BlockBuilder const&)
          at /path/tvm/src/relax/op/tensor/manipulate.cc:823
    0: tvm::relax::BlockBuilderImpl::ReportFatal(tvm::Diagnostic const&)
          at /path/tvm/src/relax/ir/block_builder.cc:138
    File "/path/tvm/src/relax/ir/block_builder.cc", line 138
    TVMError: Reshape requires the input new shape to be Shape. However, the given one is relax.TensorStructInfo
    [18:14:48] /path/tvm/src/relax/ir/block_builder.cc:64: Warning: BlockBuilder destroyed with remaining blocks!

in the onnx graph the reshape op’ input is relax.Call, one is relax.add, the other is relax.concat()

1694427809417
1694427809417687×840 34 KB

#2

That might be an issue of the ONNX frontend, it’s not hard to fix it in the file https://github.com/apache/tvm/blob/unity/python/tvm/relax/frontend/onnx/onnx_frontend.py

1 Like