Clarify error message if onnx_graphsurgeon is not installed?
#353
Comments
|
Hi @mattpopovich , The reason is that we want to make |
|
Got it, I'll make a PR in a bit to add some assertion information! |
|
Hi @mattpopovich , As you have found before, there are some similar issues elsewhere. I just released version 0.6.2, and I also introduce this poor implementation on this new releases for rapid development. The main reason for these things is that I want to make the core functionality of yolort as light as possible. Since there are many similar problems in yolort, I think it's time to resolve this problem in the next release. TorchAudio implement a def is_module_available(*modules: str) -> bool:
r"""Returns if a top-level module with :attr:`name` exists *without**
importing it. This is generally safer than try-catch block around a
`import X`. It avoids third party libraries breaking assumptions of some of
our tests, e.g., setting multiprocessing start method when imported
(see librosa/#747, torchvision/#544).
"""
return all(importlib.util.find_spec(m) is not None for m in modules)And then they implement a decorator from functools import wraps
def is_tensorrt_available():
return is_module_available("tensorrt")
def requires_tensorrt():
if is_tensorrt_available():
def decorator(func):
return func
else:
def decorator(func):
@wraps(func)
def wrapped(*args, **kwargs):
raise RuntimeError(f"{func.__module__}.{func.__name__} requires TensorRT")
return wrapped
return decoratorMore context |
zhiqwang
added
help wanted
|
Would you prefer to write specific I'm leaning towards just using I feel like they made specific |
I agree with you here, and this approach is also really a bit complicated and not necessary for us. Seems the dependencies for Maybe the def _module_available(module_path: str) -> bool:
"""
Check if a path is available in your environment.
>>> _module_available('os')
True
>>> _module_available('bla.bla')
False
"""
try:
return find_spec(module_path) is not None
except AttributeError:
# Python 3.6
return False
except ModuleNotFoundError:
# Python 3.7+
return False
except ValueError:
# Sometimes __spec__ can be None and gives a ValueError
return TrueAnd then they use a def example_requires(module_paths: Union[str, List[str]]):
return requires(module_paths)(lambda: None)() |
|
Check out this commit that I made. It follows the same thing that pytorch/audio did and I think is pretty straightforward. root@user:/home/user/git/yolov5-rt-stack/# python3 function_using_YOLOTRTGraphSurgeon.py
We're using TensorRT: 8.2.3.0 on cuda device: 0.
NOTE! Installing ujson may make loading annotations faster.
[...]
Traceback (most recent call last):
File "function_using_YOLOTRTGraphSurgeon.py", line 63, in <module>
export_tensorrt_engine(
File "/home/user/git/yolov5-rt-stack/yolort/runtime/trt_helper.py", line 69, in export_tensorrt_engine
yolo_gs = YOLOTRTGraphSurgeon(checkpoint_path, version=version, input_sample=input_sample)
File "/home/user/git/yolov5-rt-stack/yolort/utils/module_utils.py", line 39, in wrapped
raise RuntimeError(f"{func.__module__}.{func.__name__} requires {req}")
RuntimeError: yolort.relay.trt_graphsurgeon.YOLOTRTGraphSurgeon requires module: onnx_graphsurgeon
root@user:/home/user/git/yolov5-rt-stack/# pip install onnx_graphsurgeon
[successful]
root@user:/home/user/git/yolov5-rt-stack/# python3 function_using_YOLOTRTGraphSurgeon.py
[no Traceback]If you're okay with this change, I'll do the same for TensorRT, onnxruntime, and lightning. Then I'll make a PR to close this issue. Additionally, let me know if you are okay with the placement of |
Thanks, I think it's very pretty!
And there is already a file aimed to resolve the version and dependencies problem: https://github.com/zhiqwang/yolov5-rt-stack/blob/main/yolort/utils/dependency.py . I guess it will be better if we could move the |
Sounds good. I'll make the PR tomorrow! |
🐛 Describe the bug
When following the TensorRT deployment instructions, there is a point where we try to use
YOLOTRTGraphSurgeon, but to use that, we need toimport onnx_graphsurgeon as gs. Otherwise, this line will fail becausegs = None.I'm not sure why you catch the
ImportErrorifonnx_graphsurgeonis not found asYOLOTRTGraphSurgeonis the only thing that is in that file and it seems like we needonnx_graphsurgeon, but just wanted to bring this to your attention.Low priority 😄
Versions
The text was updated successfully, but these errors were encountered: