Create onnx model with reshape node

Hi All,

Currently, I create a simple ONNX model with reshape node with below code:

shape = [2,4]
shape2 = [2,4]

input_data = np.random.randn(*shape).astype(np.int64)

original_shape = [2, 3, 4]
input_data2 = np.random.random_sample(original_shape).astype(np.int64)

const_node = onnx.helper.make_node('Constant', inputs=[], outputs=['const_shape'],
                               value=onnx.helper.make_tensor(
                                   name='const_tensor',
                                   data_type=onnx.TensorProto.INT64,
                                   dims=input_data.shape,
                                   vals=input_data.flatten()))

node1 = onnx.helper.make_node('Identity', inputs=['add1'], outputs=['identity1'], name='identity_node1')
node2 = onnx.helper.make_node('Abs', inputs=['identity1'], outputs=['add2'], name='abs_node2')
node3 = onnx.helper.make_node('Identity', inputs=['add2'], outputs=['identity2'], name='identity_node2')
node4 = onnx.helper.make_node('Abs', inputs=['identity2'], outputs=['add3'], name='abs_node3')
node5 = onnx.helper.make_node('Identity', inputs=['add3'], outputs=['identity3'], name='identity_node3')
node6 = onnx.helper.make_node('Abs', inputs=['identity3'], outputs=['add4'], name='abs_node4')
node7 = onnx.helper.make_node('Identity', inputs=['add4'], outputs=['identity4'], name='identity_node4')
node8 = onnx.helper.make_node('Abs', inputs=['identity4'], outputs=['add5'], name='abs_node5')
node9 = onnx.helper.make_node('Add', inputs=['add5', 'identity3'], outputs=['add6'], name='add_node1')
node10 = onnx.helper.make_node('Identity', inputs=['add6'], outputs=['identity5'], name='identity_node5')
node11 = onnx.helper.make_node('Abs', inputs=['identity5'], outputs=['add7'], name='abs_node6')
node12 = onnx.helper.make_node('Reshape', inputs=['add7', 'const_shape'], outputs=['reshape'])

graph = onnx.helper.make_graph([const_node, node1, node2, node3, node4, node5, node6, node7, node8, node9, node10, node11, node12], 'test_graph',
                    [onnx.helper.make_tensor_value_info('add1', onnx.TensorProto.FLOAT, shape)],
                    [onnx.helper.make_tensor_value_info('reshape', onnx.TensorProto.FLOAT, shape2)])

model = onnx.helper.make_model(graph, producer_name='ngraph ONNX Importer')

onnx.save(model, 'model.onnx')

model_path = 'model.onnx'
session = onnxruntime.InferenceSession(model_path, None)
input_name1 = session.get_inputs()[0].name

a = np.random.rand(*shape).astype("float32")

print(a)
raw_result = session.run(None, {input_name1: a})

output_data = np.array(raw_result[0])
print(output_data)

the error log as below:

    sess = C.InferenceSession(session_options, self._model_path, True, self._read_config_from_model)
onnxruntime.capi.onnxruntime_pybind11_state.Fail: [ONNXRuntimeError] : 1 : FAIL : Load model from model.onnx failed:Node () Op (Reshape) [ShapeInferenceError] Invalid position of 0

I have refer the sample code https://www.programcreek.com/python/example/122583/onnx.helper.make_node to create onnx model with reshape node, but I still get error.

So could anyone show me the way to create onnx model with reshape node?

[hoaquocphan]

Hi @jcwchen

Thank you for your support. I can create the simple model with reshape node and it can run successfully. I also have 2 questions need support from you:

1/ Please see my below snippet code

original_shape = [2, 3, 4]
shape_of_shape=[3,]
new_shape=[1, 1, 24]
node = onnx.helper.make_node('Reshape', inputs=['data', 'shape'], outputs=['reshaped'])
graph = onnx.helper.make_graph([node], 'test_graph',
                    [onnx.helper.make_tensor_value_info('data', onnx.TensorProto.FLOAT, original_shape), 
                    onnx.helper.make_tensor_value_info('shape', onnx.TensorProto.INT64, shape_of_shape)],
                    [onnx.helper.make_tensor_value_info('reshaped', onnx.TensorProto.FLOAT, new_shape)])
model = onnx.helper.make_model(graph, producer_name='ngraph ONNX Importer')
onnx.save(model, 'model.onnx')
model_path = 'model.onnx'
session = onnxruntime.InferenceSession(model_path, None)
input_name1 = session.get_inputs()[0].name
input_name2 = session.get_inputs()[1].name
data = np.random.random_sample(original_shape).astype(np.float32)
data_shape = np.array([4, 2, 3], dtype=np.int64)
raw_result = session.run(None, {input_name1: data, input_name2: data_shape})
output_data = np.array(raw_result[0])
print("data:")
print(data)
print("output data:")
print(output_data)

You can see that I can create the model with output shape [1, 1, 24] is: new_shape=[1, 1, 24] the model will define output of reshaped node is [1, 1, 24]

But when I feed the model with output shape [4, 2, 3]: data_shape = np.array([4, 2, 3], dtype=np.int64) The model create output with shape [4, 2, 3]

this mean, the output shape define in [onnx.helper.make_tensor_value_info(‘reshaped’, onnx.TensorProto.FLOAT, new_shape)]) doesn’t impact to the output shape of model. So, why we need to define the new_shape=[1, 1, 24] when create the model?

2/ Please get below models from my git repo: https://github.com/hoaquocphan/shared_files.git

my_model.onnx mobilenetv2-1.0.onnx

You can see that the reshape node in my_model.onnx (created by above code snippet) have 2 inputs but the reshape node in mobilenetv2-1.0.onnx model have only one input (the reshape_attr_tensor421 input is not output of any other node)

So How to create the model have reshape node as mobilenetv2-1.0.onnx (we fix the input shape to the model 1x1000, don’t need to feed the input shape 1x1000)?

[hoaquocphan]

Hi @jcwchen

I can create the reshape node with fix shape input in the model.