Error when checking model with external data

Bug Report

Describe the bug

I saved a simple model with save_as_external_data=True, then load the model into memory with load_external_data=False, will get an error when using checker.check_model() for this model.

System information

• OS Platform and Distribution (e.g. Linux Ubuntu 16.04): win10
• ONNX version (e.g. 1.7): 1.9.0
• Python version: 3.8.0

Reproduction instructions

• Describe the code to reproduce the behavior.

import numpy as np
import onnx
from onnx import helper
from onnx import AttributeProto, TensorProto, GraphProto
import json
from google.protobuf.json_format import MessageToJson

#======== 1. create an onnx model
dtype = np.float32
x = np.random.randn(3, 5).astype(dtype)

data_w = np.random.randn(2, 5).astype(dtype)
w = helper.make_tensor(name = 'w', data_type = TensorProto.FLOAT, dims = data_w.shape, vals = data_w.flatten().astype(dtype).tobytes(), raw=True)

Z = helper.make_tensor_value_info('Z', TensorProto.FLOAT, [2, 3])

# Create node (NodeProto)
X = helper.make_node(
    'Constant',
    inputs=[],
    outputs=['X'],
    value=onnx.helper.make_tensor(
        name='const_tensor_x',
        data_type=onnx.TensorProto.FLOAT,
        dims=x.shape,
        vals=x.flatten().astype(float),
    ),
)
node_gemm = helper.make_node(
    'Gemm',                  # name
    ['w', 'X'], # inputs
    ['Z'],                  # outputs
    name = 'gemm1'
)
# Create the graph (GraphProto)
graph_def = helper.make_graph(
    [X, node_gemm],        # nodes
    'test-model',      # name
    [],  # inputs
    [Z],  # outputs
    initializer = [w],
)
model_def = helper.make_model(graph_def, producer_name='onnx-example')

#======== 2. save an onnx model
model_path = '../outs/model.onnx'
# save model with separate data files
onnx.save_model(model_def, model_path
                , save_as_external_data=True
                , all_tensors_to_one_file=False
                , size_threshold=0)

#======== 4. load and try save
# (1) load with data, then check: will success
model_full = onnx.load(model_path)
onnx.checker.check_model(model_full)

# (2) load without data, then check: will fail
model_sep = onnx.load(model_path, load_external_data=False)
onnx.checker.check_model(model_sep)  # onnx.onnx_cpp2py_export.checker.ValidationError: Data of TensorProto ( tensor name: w) should be stored in w, but it doesn't exist or is not accessible.

• Error:

"C:\Program Files\Python38\python.exe" D:/pythonProject/test/issue1_check_model.py
Traceback (most recent call last):
  File "D:/pythonProject/test/issue1_check_model.py", line 61, in <module>
    onnx.checker.check_model(model_sep)  # onnx.onnx_cpp2py_export.checker.ValidationError: Data of TensorProto ( tensor name: w) should be stored in w, but it doesn't exist or is not accessible.
  File "C:\Program Files\Python38\lib\site-packages\onnx\checker.py", line 104, in check_model
    C.check_model(protobuf_string)
onnx.onnx_cpp2py_export.checker.ValidationError: Data of TensorProto ( tensor name: w) should be stored in w, but it doesn't exist or is not accessible.

[jcwchen]

We are targeting the next release for this feature, but typically ONNX is released semi-annually. (So it will be supported by official release in six months) However, once we have implemented this feature and got it merged, you can still use it from building the latest ONNX from source or weekly TestPyPI package. I will post the updates in this thread if any. Thanks!

[yjydfnhc]

If you load the model first and then call checker it does not know the data dir and therefore tries to find the data in the current working dir and complains if it does not find it. If your script and data are not in the same directory, directly use the checker api and providing the model path .i.e check_model_path(“D:/mytest/outs/model.onnx”)

great, according to your advice, it truly does not fail by using checker.check_model(model_path) instead, thanks!