AttributeError: module 'onnx' has no attribute 'load'

[ezyang]

When I run test/onnx/test_utility_funs.py locally I get an error:

======================================================================
ERROR: test_strip_doc_string (__main__.TestUtilityFuns)
----------------------------------------------------------------------
Traceback (most recent call last):
  File "test/onnx/test_utility_funs.py", line 226, in test_strip_doc_string
    self.assertTrue(is_model_stripped(io.BytesIO()))
  File "test/onnx/test_utility_funs.py", line 220, in is_model_stripped
    model = onnx.load(io.BytesIO(f.getvalue()))
AttributeError: module 'onnx' has no attribute 'load'

======================================================================
ERROR: test_strip_doc_string (__main__.TestUtilityFuns_opset10)
----------------------------------------------------------------------
Traceback (most recent call last):
  File "test/onnx/test_utility_funs.py", line 226, in test_strip_doc_string
    self.assertTrue(is_model_stripped(io.BytesIO()))
  File "test/onnx/test_utility_funs.py", line 220, in is_model_stripped
    model = onnx.load(io.BytesIO(f.getvalue()))
AttributeError: module 'onnx' has no attribute 'load'

----------------------------------------------------------------------

I can fix it by running pip install onnx. But the error message here is unexpected; if onnx wasn't installed, why did I get an error that it imported successfully and just doesn't have any attributes?

When I print onnx.__file__ I get None. My best guess is that we are accidentally picking up test/onnx folder as a namespace.

cc @houseroad @spandantiwari @lara-hdr @BowenBao @neginraoof

[spandantiwari]

Just add some context (i.e., not exactly the answer to the question) - to the best of my knowledge, onnx is not installed as a dependency of PyTorch. It is only a test dependency.

[nolifenameless]

This error here is caused by the module 'onnx'（before pip install onnx) which can avoid python import inspection,

I try to recur this situation but still got nothing. @ezyang would you like provide info about this situation? that would be very helpful.

IMO, wrap this case "test_strip_doc_string" in some context which scenario should run(Ex. method _skipper in file test/onnx/test_pytorch_common.py ) may solve this problem

[ezyang]

This issue is pretty old so I'm happy to just assume it's been silently fixed since then.

[Manoj-Borkar]

I am still getting the same error with onnx.load. Is it resolved?? If yes, could someone pls point me in the right direction

[saadnajib]

I had the same problem fixed it in a way that, I had my python file name "onnx.py" same name as the library I was importing, changing the name of my python file from "onnx.py" to "converting_to_onnx.py" solved the issue.

[dron-dronych]

for those curious on why this happens is because when you run a py file named "onnx.py" it first adds that directory to your python search path. It then searches for a name to import (import onnx) and the first one it encounters is your directory with YOUR onnx.py. This behavior should be avoided as it adds unexpected behavior to the python import mechanism.

[kingydw]

I had the same problem fixed it in a way that, I had my python file name "onnx.py" same name as the library I was importing, changing the name of my python file from "onnx.py" to "converting_to_onnx.py" solved the issue.

I jumped into this trap too...😂