Not getting even near accurate results trying to make a identity function using convolution node

Nobody has claimed this yet.

### Describe the feature request

def idenity(shape):
input_tensor = onnx.helper.make_tensor_value_info('input_tensor',onnx.TensorProto.FLOAT,input_shape)
output = onnx.helper.make_tensor_value_info('output',onnx.TensorProto.FLOAT,['N','C','H','W'])
kernel1_val = (np.eye(input_shape[1]).astype(np.float32)).flatten()
kernel = onnx.helper.make_tensor('kernel',onnx.TensorProto.FLOAT,[input_shape[1],input_shape[1],1,1],kernel1_val)
bias = onnx.helper.make_tensor('bias',onnx.TensorProto.FLOAT,[1],[0.0])
iden_node = onnx.helper.make_node('Conv',inputs=["input_tensor","kernel","bias"],outputs=["output"],kernel_shape = [1,1],pads = [0,0,0,0],strides=[1,1],group=1,name="identity_conv")
graph = onnx.helper.make_graph(nodes=[iden_node],name='identity',inputs=[input_tensor],outputs=[output],initializer=[kernel,bias])
return graph
#this is a identity convolution which I used instead of direct identity node

### Describe scenario use case

I have taken a more than 500+ sample of input tensors with different channels and dimensions and random values when I compare the input and output after each iteration during the start input and output with shape and value tensors are equal after 5 to 10 iterations the output tensors where giving "nan" the shape is same as the input tensor
```[tasklist]
### Tasks
```

No source file, test, runtime version, or complete reproducer is identified. Start by reproducing the reported identity Conv graph with the 500+ random tensor cases and compare values across iterations, then trace where NaNs first appear. Done means the cause is isolated and a regression test or precise reproducible diagnosis is recorded.

Written by the indexing model from the issue text.