[PyTorch] Missing op aten::rnn_tanh

[cVladu]

I built a simple RNN network in PyTorch and tried to convert it using relay.frontend.from_pytorch interface

Expected behavior

The network is converted

Actual behavior

NotImplementedError: The following operators are not implemented: ['aten::rnn_tanh'] is raised

Environment

Operating system:
Distributor ID: Ubuntu
Description: Ubuntu 18.04.6 LTS
Release: 18.04
Codename: bionic

TVM version: 0.9.dev0
Steps to build the TVM were followed from: https://tvm.apache.org/docs/install/from_source.html -- no change to config.cmake file

Steps to reproduce

def test_RNN_torch(num_layers: int,
                                    bidirectional: bool,
                                    use_bias: bool,
                                    hidden_size: int,
                                    input_size: int,
                                    seq_len: int,
                                    batch_first: bool,
                                    batch_size: int):
    r''' 
    Args:
        num_layers (int): num_layers to be passed to torch.nn.RNN
        bidirectional (bool): whether to build bidirectional RNN or not
        use_bias (bool): whether to use bias or not
        hidden_size (int): hidden_size of RNN cells
        input_size (int): Input features
        seq_len (int): Timesteps in input data
        batch_first (bool): Whether batch dimension is first or second dimension in input tensor
        batch_size (int): Batch size of input. If 0, unbatched input will be fed to network
    '''

    if batch_first:
        input_shape = (batch_size, seq_len, input_size)
    else:
        input_shape = (seq_len, batch_size, input_size)
    pytorch_net = torch.nn.Sequential(
        torch.nn.RNN(input_size,
                     hidden_size,
                     batch_first=batch_first,
                     num_layers=num_layers,
                     bidirectional=bidirectional,
                     bias=use_bias)
    )

    scripted_model = torch.jit.trace(pytorch_net.eval(),
                                     torch.randn(input_shape))

    mod, params = relay.frontend.from_pytorch(scripted_model,
                                              [('input', input_shape)])
    mod = relay.transform.InferType()(mod)
    print(mod.astext())

if __name__ == "__main__":

    test_RNN_torch(1,
                   False,
                   True,
                   5,
                   5,
                   15,
                   True,
                   32)

full traceback:

Traceback (most recent call last):
  File "lstm_transform.py", line 477, in <module>
    test_RNN_torch(1,
  File "lstm_transform.py", line 469, in test_RNN_torch
    mod, params = relay.frontend.from_pytorch(scripted_model,
  File "/home/catalinvladu/Mambaforge/envs/tvm-build/lib/python3.8/site-packages/tvm/relay/frontend/pytorch.py", line 4030, in from_pytorch
    converter.report_missing_conversion(op_names)
  File "/home/catalinvladu/Mambaforge/envs/tvm-build/lib/python3.8/site-packages/tvm/relay/frontend/pytorch.py", line 3236, in report_missing_conversion
    raise NotImplementedError(msg)
NotImplementedError: The following operators are not implemented: ['aten::rnn_tanh']

If this is not the correct place to raise the issue or if my expected result is wrong, please let me know

[yuanfz98]

Hello cVladu,

It is the correct place to raise the issue. :)

When you call

relay.frontend.from_pytorch

You are converting pytorch model (or jit traced model) to "tvm model". And if tvm doesn't have implemented the corresponding equivalent operator (in your case, aten::rnn_tanh), error occurs.

So you need contributors come to help. I can take a stab.