从 ONNX 运行超分辨率模型时出错

Question

测试super resolution model 的 ONNX 模型，我在运行 this sample program 时出错。

我的 ONNX 版本是 1.5.0，onnxruntime 是 1.4.0。 Onnxruntime was installed using pip。 Pytorch 版本是 1.6.0

错误在ort_session = onnxruntime.InferenceSession('/home/itc/pytorch/sub_pixel_cnn_2016/model/super-resolution-10.onnx')

错误在于加载onnx模型。

Traceback (most recent call last):
  File "test.py", line 73, in <module>
    ort_session = onnxruntime.InferenceSession('/home/itc/pytorch/sub_pixel_cnn_2016/model/super-resolution-10.onnx')
  File "/home/itc/pytorch/lib/python3.7/site-packages/onnxruntime/capi/session.py", line 158, in __init__
    self._load_model(providers or [])
  File "/home/itc/pytorch/lib/python3.7/site-packages/onnxruntime/capi/session.py", line 166, in _load_model
    True)
RuntimeError: /onnxruntime_src/onnxruntime/core/session/inference_session.cc:238 onnxruntime::InferenceSession::InferenceSession(const onnxruntime::SessionOptions&, const onnxruntime::Environment&, const string&) status.IsOK() was false. Given model could not be parsed while creating inference session. Error message: Protobuf parsing failed.
 

我该如何解决这个错误？

Answer 1

super-resolution-10.onnx 对我来说似乎加载正常。 我从https://github.com/onnx/models/blob/master/vision/super_resolution/sub_pixel_cnn_2016/model/super-resolution-10.onnx下载了文件

$ pip install onnxruntime
...
Successfully installed onnxruntime-1.5.1

我也试过pip install onnxruntime==1.4.0 - 也很好用。

然后尝试加载它（有一堆警告，但它加载正常）：

In [1]: import onnxruntime

In [2]: onnxruntime.InferenceSession("super-resolution-10.onnx")
2020-10-12 23:25:23.486256465 [W:onnxruntime:, graph.cc:1030 Graph] Initializer conv1.bias appears in graph inputs and will not be treated as constant value/weight. This may prevent some of the graph optimizations, like const folding. Move it out of graph inputs if there is no need to override it, by either re-generating the model with latest exporter/converter or with the tool onnxruntime/tools/python/remove_initializer_from_input.py.
2020-10-12 23:25:23.486293664 [W:onnxruntime:, graph.cc:1030 Graph] Initializer conv1.weight appears in graph inputs and will not be treated as constant value/weight. This may prevent some of the graph optimizations, like const folding. Move it out of graph inputs if there is no need to override it, by either re-generating the model with latest exporter/converter or with the tool onnxruntime/tools/python/remove_initializer_from_input.py.
2020-10-12 23:25:23.486308563 [W:onnxruntime:, graph.cc:1030 Graph] Initializer conv2.bias appears in graph inputs and will not be treated as constant value/weight. This may prevent some of the graph optimizations, like const folding. Move it out of graph inputs if there is no need to override it, by either re-generating the model with latest exporter/converter or with the tool onnxruntime/tools/python/remove_initializer_from_input.py.
2020-10-12 23:25:23.486322663 [W:onnxruntime:, graph.cc:1030 Graph] Initializer conv2.weight appears in graph inputs and will not be treated as constant value/weight. This may prevent some of the graph optimizations, like const folding. Move it out of graph inputs if there is no need to override it, by either re-generating the model with latest exporter/converter or with the tool onnxruntime/tools/python/remove_initializer_from_input.py.
2020-10-12 23:25:23.486335363 [W:onnxruntime:, graph.cc:1030 Graph] Initializer conv3.bias appears in graph inputs and will not be treated as constant value/weight. This may prevent some of the graph optimizations, like const folding. Move it out of graph inputs if there is no need to override it, by either re-generating the model with latest exporter/converter or with the tool onnxruntime/tools/python/remove_initializer_from_input.py.
2020-10-12 23:25:23.486348462 [W:onnxruntime:, graph.cc:1030 Graph] Initializer conv3.weight appears in graph inputs and will not be treated as constant value/weight. This may prevent some of the graph optimizations, like const folding. Move it out of graph inputs if there is no need to override it, by either re-generating the model with latest exporter/converter or with the tool onnxruntime/tools/python/remove_initializer_from_input.py.
2020-10-12 23:25:23.486361862 [W:onnxruntime:, graph.cc:1030 Graph] Initializer conv4.bias appears in graph inputs and will not be treated as constant value/weight. This may prevent some of the graph optimizations, like const folding. Move it out of graph inputs if there is no need to override it, by either re-generating the model with latest exporter/converter or with the tool onnxruntime/tools/python/remove_initializer_from_input.py.
2020-10-12 23:25:23.486384161 [W:onnxruntime:, graph.cc:1030 Graph] Initializer conv4.weight appears in graph inputs and will not be treated as constant value/weight. This may prevent some of the graph optimizations, like const folding. Move it out of graph inputs if there is no need to override it, by either re-generating the model with latest exporter/converter or with the tool onnxruntime/tools/python/remove_initializer_from_input.py.
Out[2]: <onnxruntime.capi.session.InferenceSession at 0x7f58367236d0>

我认为您的 ONNX 文件可能已损坏，请尝试使用 Netron 加载以验证。

附带说明，PyTorch 版本和 onnx 版本应该与加载无关。