ONNX optimzer crashes for initializers not listed as graph-input

[thilow]

Initializers are not required being listed in graph inputs (To my understanding this requirement was lifted in context of issue #1449).

Problem: The ONNX optimizer fails for models that have initializers which are not listed in graph inputs.

Below there is a sample script demonstrating the issue. For the first version of the model (model_1) the optimizer passes while for the second it fails.
See further below for script output

#%%
import onnx
from onnx import helper, optimizer
from onnx import AttributeProto, TensorProto, GraphProto

# Creating a simple model that adds 42 to the input-value.
# The initializer is not added to he graph-inputs
def create_simpel_model():
    # Create one input (ValueInfoProto)
    a = helper.make_tensor_value_info('a', TensorProto.FLOAT, [1])

    # Create an initialzer with value 42
    b = helper.make_tensor("b",TensorProto.FLOAT,[1], [42])

    # Create one output (ValueInfoProto)
    Y = helper.make_tensor_value_info('Y', TensorProto.FLOAT, [1])

    # Create a node (NodeProto)
    node_def = helper.make_node(
        'Add', # node name
        ['a', 'b'], # inputs
        ['Y'], # outputs
    )

    # Create the graph (GraphProto)
    graph_def = helper.make_graph(
        [node_def],
        'test-model',
        [a],
        [Y],
        [b]
    )

    # Create the model (ModelProto)
    return helper.make_model(graph_def, producer_name='onnx-example')


print("ONNX version: ", onnx.__version__)
#%%
model_1 = create_simpel_model()

# Postprocessing the model by adding the initializer 'b' to the input.
# This prevents the optimizer from crashing.
new_input = helper.make_tensor_value_info('b', TensorProto.FLOAT,[1])
model_1.graph.input.append(new_input)

#%% check model_1
onnx.checker.check_model(model_1)
print('model_1 is checked and ok!')

#%%
# Optimize the model
passes = [
    "eliminate_deadend",
    "eliminate_identity",
    "eliminate_nop_dropout",
    "eliminate_nop_monotone_argmax",
    "eliminate_nop_pad",
    "eliminate_nop_transpose",
    "extract_constant_to_initializer",
    "fuse_consecutive_concats",
    "fuse_consecutive_log_softmax",
    "fuse_consecutive_reduce_unsqueeze",
    "fuse_consecutive_squeezes",
    "fuse_consecutive_transposes"]

model_1 = optimizer.optimize(model_1, passes)
print("model_1 is optimized")

#%%
# Now we don't add the initialzer to the inputs
model_2 = create_simpel_model()


# check model_2
onnx.checker.check_model(model_2)
print('model_2 is checked and ok too')

print("trying to optimize model_2")
# Next line fails!
model_1 = optimizer.optimize(model_2, passes)
print("model_2 is optimized")

The script prduces output below:

ONNX version:  1.5.0
model_1 is checked and ok!
model_1 is optimized
model_2 is checked and ok too
trying to optimize model_2
Traceback (most recent call last):
  File "c:\Users\thilow\.vscode\extensions\ms-python.python-2019.6.24221\pythonFiles\ptvsd_launcher.py", line 43, in <module>
    main(ptvsdArgs)
  File "c:\Users\thilow\.vscode\extensions\ms-python.python-2019.6.24221\pythonFiles\lib\python\ptvsd\__main__.py", line 434, in main
    run()
  File "c:\Users\thilow\.vscode\extensions\ms-python.python-2019.6.24221\pythonFiles\lib\python\ptvsd\__main__.py", line 312, in run_file
    runpy.run_path(target, run_name='__main__')
  File "e:\src\BrainWaveMunich\smartReply\OnnxPythonTools\optimizer_bug_repro.py", line 85, in <module>
    model_1 = optimizer.optimize(model_2, passes)
  File "E:\anaconda3\lib\site-packages\onnx\optimizer.py", line 55, in optimize
    optimized_model_str = C.optimize(model_str, passes)
IndexError: invalid unordered_map<K, T> key

[houseroad]

Yes, we should fix, since the spec support such cases.

cc: @spandantiwari

[jcwchen]

Please note that ONNX optimizer has been moved to another repo https://github.com/onnx/optimizer since ONNX 1.9. If you still have questions related to the optimizer, please raise an issue there. Thank you!