[Training] Could not find an implementation for Cos(7) node

[SunHaoOne]

Describe the issue

Hello, I'm working with two separate models: an encoder and a decoder. Individually, exporting either model works without any issues. However, I encountered a problem when trying to export a unified model that integrates both the encoder and the decoder. Below is the structure of my combined model:

class Model():
    def forward(self, x):
        tmp = encoder(x)
        out = decoder(tmp)
        return out

Traceback (most recent call last):
  File "onnx/test_save_all.py", line 66, in <module>
    onnx_model = ort.InferenceSession("qcnet.onnx",
  File "/home/shy/mydiskBig/miniforge3/envs/onnx/lib/python3.8/site-packages/onnxruntime/capi/onnxruntime_inference_collection.py", line 419, in __init__
    self._create_inference_session(providers, provider_options, disabled_optimizers)
  File "/home/shy/mydiskBig/miniforge3/envs/onnx/lib/python3.8/site-packages/onnxruntime/capi/onnxruntime_inference_collection.py", line 463, in _create_inference_session
    sess.initialize_session(providers, provider_options, disabled_optimizers)
onnxruntime.capi.onnxruntime_pybind11_state.NotImplemented: [ONNXRuntimeError] : 9 : NOT_IMPLEMENTED : Could not find an implementation for Cos(7) node with name '/encoder/Cos'

I have checked the onnx in netron, but it seems ok. The pytorch code is here:

polyline_inputs = torch.zeros([100,4])
orient_pl = polyline_inputs[:, 2].contiguous()
orient_vector_pl = torch.stack([orient_pl.cos(), orient_pl.sin()], dim=-1)

To reproduce

Urgency

No response

ONNX Runtime Installation

Built from Source

ONNX Runtime Version or Commit ID

1.16.3

PyTorch Version

2.1.1+cu118

Execution Provider

Default CPU

Execution Provider Library Version

cu118

[xadupre]

The error means onnxruntime cannot find an implementation for operator Cos with opset 7 which is the most recent opset for this one. Maybe the input type for the operator Cos happens to be outside the supported list. Are running it on Cpu, Cuda, ...? Can you share the instructions you used to convert your model into ONNX?

[SunHaoOne]

Hi, @xadupre
Sorry for the late reply. I have tried some different opset versions, but they seem not to work. I am now running it on the CPU, and here are the instructions. By the way, my model includes two parts: an encoder and a decoder. The problem occurs with the cos operator in the encoder. When I change all the cos to sin, the problem does not occur. The input type they are all fp32

import torch
import torch.onnx


model = Model()
model = model.cpu()
locs, heads, pi, x_pl_resued = model(point_inputs,
                                    polyline_inputs,
                                    edge_pl2pl,
                                    agent_inputs)

input_names = ["point_inputs", "polyline_inputs", "edge_pl2pl", "agent_inputs"]

output_names = ["locs", "heads", "pi", "x_pl"]

onnx_filename = "model.onnx"

dummy_input = (point_inputs, polyline_inputs, edge_pl2pl, agent_inputs)

torch.onnx.export(model, dummy_input, onnx_filename, verbose=False,
                 input_names=input_names, output_names=output_names,
                 export_params=True, opset_version=16,
                 keep_initializers_as_inputs=True, do_constant_folding=True)



import onnx
import onnxruntime as ort



model = onnx.load("model.onnx")


onnx_model = ort.InferenceSession("qcnet.onnx",
                                  providers=['CPUExecutionProvider']) 

import numpy as np
inputs = {
    "point_inputs": point_inputs.numpy().astype(np.float32), 
    "polyline_inputs": polyline_inputs.numpy().astype(np.float32),
    "edge_pl2pl": edge_pl2pl.numpy().astype(np.float32),
    "agent_inputs": agent_inputs.numpy().astype(np.float32)
}

outputs = onnx_model.run(None, inputs)

print(outputs)