[PyTorch] ImplementationError: The following operators are not implemented: ['aten::embedding_bag']

[crazydemo]

trafficstars

when I try to relay dlrm (pytorch model), I encounter the error: ImplementationError: The following operators are not implemented: ['aten::embedding_bag']

pytorch=1.7.0, tvm is the latest

[GoodNight-bye]

hello, have you finished importing the DLRM model，i have the same task

[crazydemo]

no i haven't

发自我的iPhone

------------------ Original ------------------ From: GoodNight-bye @.> Date: Mon,Jun 21,2021 4:26 PM To: apache/tvm @.> Cc: Ivy Zhang @.>, Author @.> Subject: Re: [apache/tvm] ImplementationError: The following operators are not implemented: ['aten::embedding_bag'] (#7986)

[GoodNight-bye]

no i haven't 发自我的iPhone … ------------------ Original ------------------ From: GoodNight-bye @.> Date: Mon,Jun 21,2021 4:26 PM To: apache/tvm @.> Cc: Ivy Zhang @.>, Author @.> Subject: Re: [apache/tvm] ImplementationError: The following operators are not implemented: ['aten::embedding_bag'] (#7986)

你还是原先的这个问题吗

[crazydemo]

是的

发自我的iPhone

------------------ Original ------------------ From: GoodNight-bye @.> Date: Mon,Jun 21,2021 4:30 PM To: apache/tvm @.> Cc: Ivy Zhang @.>, Author @.> Subject: Re: [apache/tvm] ImplementationError: The following operators are not implemented: ['aten::embedding_bag'] (#7986)

[GoodNight-bye]

我这边是把网络拆分了，embedding的部分不放在NPU上执行，但是后面有个index算子的问题

[GoodNight-bye]

embedding_bag is supported in relay/frontend/onnx.py, you can refer to this file.

[GoodNight-bye]

I have done the job, DLRM is already running on our NPU. If you have any question, you can contact me.

[TaoCheng-CS]

I have done the job, DLRM is already running on our NPU. If you have any question, you can contact me. 您好， 我目前也遇到这个算子不支持的问题， 能给个联系方式吗？感谢！

[GoodNight-bye]

I have done the job, DLRM is already running on our NPU. If you have any question, you can contact me. 您好， 我目前也遇到这个算子不支持的问题， 能给个联系方式吗？感谢！

aten::embedding_bag relay层直接参考onnx的实现，算子侧的话需要实现对应的take算子

[zgplvyou]

embedding_bag is supported in relay/frontend/onnx.py, you can refer to this file.

Hi, I`m also relay from onnx dlrm model, in this file(onnx.py), the implementation of embedding_bag use a reshape op( indices = _op.reshape(indices, indices_shape)), however, if the indices_shape is [3, -1], the size of indices is [13, ], the indices can not be reshape to the size[3, ?],how to do it!