how to convert a keras or tensorflow model with multiple inputs to a onnx one?

[lk1983823]

Describe the issue

how to convert a keras or tensorflow model with multiple inputs to a onnx one?

To reproduce

I have a tensorflow model with inputs like this:
model(past_data1,past_data2, past_data3, future_features1,future_features2, tsidx)
where:
past_data1.shape,past_data2.shape, past_data3.shape, future_features1.shape,future_features2.shape, tsidx.shape are:

(TensorShape([64, 720]),
 TensorShape([8, 720]),
 TensorShape([1, 720]),
 TensorShape([8, 30]),
 TensorShape([1, 30]),
 TensorShape([64]))

Now I try to convert this model to a onnx one to gain some accelerations with the code:

input1 = tf.TensorSpec(shape=(64, 720), dtype=tf.float32, name='x1')
input2 = tf.TensorSpec(shape=(8, 720), dtype=tf.float32, name='x2')
input3 = tf.TensorSpec(shape=(1, 720), dtype=tf.float32, name='x3')
input4 = tf.TensorSpec(shape=(8, 30), dtype=tf.float32, name='x4')
input5 = tf.TensorSpec(shape=(1, 30), dtype=tf.float32, name='x5')
input6 = tf.TensorSpec(shape=(64, ), dtype=tf.float32, name='x6')
model_input_signature = (input1, input2, input3, input4, input5, input6)

output_path = "./modellk.onnx"
onnx_model, _ = tf2onnx.convert.from_keras(model,
    output_path=output_path,
    input_signature=model_input_signature
)

Finally, it gives errors:

TypeError                                 Traceback (most recent call last)
Cell In[46], line 10
      7 model_input_signature = (input1, input2, input3, input4, input5, input6)
      9 output_path = "[./modellk.onnx](https://file+.vscode-resource.vscode-cdn.net/media/lk/disk1/lk_git/2_TimeSeries/2023_TimeSeries/google_ts/tide/modellk.onnx)"
---> 10 onnx_model, _ = tf2onnx.convert.from_keras(model,
     11     output_path=output_path,
     12     input_signature=model_input_signature
     13 )

File [~/anaconda3/envs/dpc/lib/python3.8/site-packages/tf2onnx/convert.py:452](https://file+.vscode-resource.vscode-cdn.net/media/lk/disk1/lk_git/2_TimeSeries/2023_TimeSeries/google_ts/tide/~/anaconda3/envs/dpc/lib/python3.8/site-packages/tf2onnx/convert.py:452), in from_keras(model, input_signature, opset, custom_ops, custom_op_handlers, custom_rewriter, inputs_as_nchw, outputs_as_nchw, extra_opset, shape_override, target, large_model, output_path, optimizers)
    449 except TypeError as e:
    450     # Legacy keras models don't accept the training arg tf provides so we hack around it
    451     if "got an unexpected keyword argument 'training'" not in str(e):
--> 452         raise e
    453     model_call = model.call
    454     def wrap_call(*args, training=False, **kwargs):

File [~/anaconda3/envs/dpc/lib/python3.8/site-packages/tf2onnx/convert.py:448](https://file+.vscode-resource.vscode-cdn.net/media/lk/disk1/lk_git/2_TimeSeries/2023_TimeSeries/google_ts/tide/~/anaconda3/envs/dpc/lib/python3.8/site-packages/tf2onnx/convert.py:448), in from_keras(model, input_signature, opset, custom_ops, custom_op_handlers, custom_rewriter, inputs_as_nchw, outputs_as_nchw, extra_opset, shape_override, target, large_model, output_path, optimizers)
    446 function = _saving_utils.trace_model_call(model, input_signature)
    447 try:
--> 448     concrete_func = function.get_concrete_function()
    449 except TypeError as e:
    450     # Legacy keras models don't accept the training arg tf provides so we hack around it
    451     if "got an unexpected keyword argument 'training'" not in str(e):
...
TypeError: in user code:


    TypeError: tf__call() missing 5 required positional arguments: 'past_data2', 'past_data3', 'future_features1', 'future_features2', and 'tsidx'

So how could I deal with it ? Thanks!

Urgency

No response

Platform

Linux

OS Version

Ubuntu 18.04

ONNX Runtime Installation

Released Package

ONNX Runtime Version or Commit ID

1.12.0

ONNX Runtime API

Python

Architecture

X64

Execution Provider

Default CPU

Execution Provider Library Version

No response

[justinchuby]

Please create an issue in https://github.com/onnx/tensorflow-onnx/issues instead. Be sure to include all the relavent package versions. Thanks!