HRNet stopped working after NumPy 1.24

[rodrigoberriel]

My actions before raising this issue

• Read/searched the docs
• Searched past issues

Three weeks ago NumPy 1.24.0 was released and they finally expired the deprecation for np.int introduced in Numpy 1.20:

The deprecation for the aliases np.object, np.bool, np.float,
np.complex, np.str, and np.int is expired (introduces NumPy
1.20). Some of these will now give a FutureWarning in addition to
raising an error since they will be mapped to the NumPy scalars in
the future.

Therefore, using np.int now results in an error instead of a warning. It just happens that HRNet uses np.int and the NumPy version is not pinned in their requirements.txt. So, if you try to deploy HRNet today, it will pull NumPy >= 1.24 and the following error will be thrown:

AttributeError: module 'numpy' has no attribute 'int'

23.01.11 15:42:04.675 sor.http.w0.python.logger (E) Exception raised while running init_context {"worker_id": "0"}
Traceback (most recent call last):
  File "/opt/nuclio/_nuclio_wrapper.py", line 432, in <module>
    run_wrapper()
  File "/opt/nuclio/_nuclio_wrapper.py", line 420, in run_wrapper
    loop.run_until_complete(wrapper_instance.initialize())
  File "/usr/lib/python3.8/asyncio/base_events.py", line 616, in run_until_complete
    return future.result()
  File "/opt/nuclio/_nuclio_wrapper.py", line 147, in initialize
    await self._initialize_context()
  File "/opt/nuclio/_nuclio_wrapper.py", line 158, in _initialize_context
    init_context_result = getattr(self._entrypoint_module, 'init_context')(self._context)
  File "/opt/nuclio/main.py", line 14, in init_context
    model = ModelHandler()
  File "/opt/nuclio/model_handler.py", line 45, in __init__
    self.net = utils.load_is_model(checkpoint_path, self.device)
  File "/opt/nuclio/hrnet/isegm/inference/utils.py", line 33, in load_is_model
    return load_single_is_model(state_dict, device, **kwargs)
  File "/opt/nuclio/hrnet/isegm/inference/utils.py", line 37, in load_single_is_model
    model = load_model(state_dict['config'], **kwargs)
  File "/opt/nuclio/hrnet/isegm/utils/serialization.py", line 64, in load_model
    return model_class(**model_args)
  File "/opt/nuclio/hrnet/isegm/utils/serialization.py", line 39, in new_init
    init(self, *args, **kwargs)
  File "/opt/nuclio/hrnet/isegm/model/is_hrnet_model.py", line 15, in __init__
    self.feature_extractor = HighResolutionNet(width=width, ocr_width=ocr_width, small=small,
  File "/opt/nuclio/hrnet/isegm/model/modeling/hrnet_ocr.py", line 208, in __init__
    last_inp_channels = np.int(np.sum(pre_stage_channels))
  File "/usr/local/lib/python3.8/dist-packages/numpy/__init__.py", line 284, in __getattr__
    raise AttributeError("module {!r} has no attribute "
AttributeError: module 'numpy' has no attribute 'int'
23.01.11 15:42:04.928 sor.http.w0.python.logger (W) Failed to read from connection {"err": "EOF"}
23.01.11 15:42:04.931 sor.http.w0.python.logger (E) Unexpected termination of child process {"error": null, "status": "exit status 1"}

Expected Behaviour

HRNet should be deployed with success.

Current Behaviour

The HRNet does not deploy successfully as of 3 weeks ago.

Possible Solution

Add an extra step to function-gpu.yaml, downgrading NumPy: pip install "numpy<1.24".

I tested it in my environment and it worked. I can submit a PR if this approach is okay.

Alternative:

• Patch the requirements.txt before installing it, replacing the numpy line to numpy<1.24.

Steps to Reproduce (for bugs)

1. Deploy HRNet
2. Try to use it
3. Check the logs: docker logs nuclio-nuclio-pth-saic-vul-hrnet

Context

I re-deployed HRNet after #5570, and it didn't work.

Your Environment

• Git hash commit (git log -1): 43ed0a2
• Docker version docker version (e.g. Docker 17.0.05): 20.10.18
• Are you using Docker Swarm or Kubernetes? No.
• Operating System and version (e.g. Linux, Windows, MacOS): Linux.

[nmanovic]

@rodrigoberriel , thanks for the report. Could you please send us a PR?

[rodrigoberriel]

@rodrigoberriel , thanks for the report. Could you please send us a PR?

Done!