search

jleuschn / learned_ct_reco_comparison_paper

Code and supplementing material for the Article "Quantitative comparison of deep learning-based image reconstruction methods for low-dose and sparse-angle CT applications"
14 stars 2 forks source link

Problem of ODL #2

Closed Guaishou74851 closed 3 years ago

Guaishou74851 commented 3 years ago

I run lodopab_eval.py on Colab and get:

# with dival==0.6.0
Traceback (most recent call last):
  File "lodopab_eval.py", line 16, in <module>
    from dival.reconstructors.learnedpd_reconstructor import LearnedPDReconstructor
  File "/opt/conda/lib/python3.6/site-packages/dival/reconstructors/learnedpd_reconstructor.py", line 4, in <module>
    from odl.contrib.torch import OperatorModule
ImportError: cannot import name 'OperatorModule'
# with dival==0.5.7
Traceback (most recent call last):
  File "lodopab_eval.py", line 16, in <module>
    from dival.reconstructors.learnedpd_reconstructor import LearnedPDReconstructor
  File "/usr/local/lib/python3.7/dist-packages/dival/reconstructors/learnedpd_reconstructor.py", line 4, in <module>
    from odl.contrib.torch import OperatorModule
ImportError: cannot import name 'OperatorModule' from 'odl.contrib.torch' (/usr/local/lib/python3.7/dist-packages/odl/contrib/torch/__init__.py)
jleuschn commented 3 years ago

Hi, this is most likely due to an older version of ODL, such as the stable 0.7.0. A more recent development version is required, which can be installed with

pip install https://github.com/odlgroup/odl/archive/master.zip --upgrade

FYI about the change in ODL: the related PR is #1516, where amongst other things (which i think also include an important fix to gradient accumulation) OperatorAsModule has been renamed to OperatorModule.

See also supp_material/environment_details/environments.md.

Guaishou74851 commented 3 years ago

Thanks a lot!

Guaishou74851 commented 3 years ago

Hi, I upgraded ODL to the latest version (1.0.0) by using:

pip install https://github.com/odlgroup/odl/archive/master.zip --upgrade

, and I downloaded and unzipped the complete LoDoPaB-CT dataset automatically with dival==0.6.1 Python module, but when I run lodopab_eval.py again, I got:

Traceback (most recent call last):
  File "lodopab_eval.py", line 70, in <module>
    dataset = get_standard_dataset('lodopab', impl=IMPL)
  File "/opt/conda/lib/python3.6/site-packages/dival/datasets/standard.py", line 174, in get_standard_dataset
    dataset = LoDoPaBDataset(**lodopab_kwargs)
  File "/opt/conda/lib/python3.6/site-packages/dival/datasets/lodopab_dataset.py", line 259, in __init__
    self.ray_trafo = self.get_ray_trafo(impl=impl)
  File "/opt/conda/lib/python3.6/site-packages/dival/datasets/lodopab_dataset.py", line 348, in get_ray_trafo
    return odl.tomo.RayTransform(self.space[1], self.geometry, **kwargs)
  File "/opt/conda/lib/python3.6/site-packages/odl/tomo/operators/ray_trafo.py", line 193, in __init__
    impl_type, self.__cached_impl = self._initialize_impl(impl)
  File "/opt/conda/lib/python3.6/site-packages/odl/tomo/operators/ray_trafo.py", line 238, in _initialize_impl
    'implementations.'.format(impl)
ValueError: The 'astra_cuda' `impl` is not found. This `impl` is either not supported, it may be misspelled, or external packages required are not available. Consult `RAY_TRAFO_IMPLS` to find the run-time available implementations.

After some efforts, I am still really confused and do not know how to perform a complete test.

jleuschn commented 3 years ago

Hi,

i think you can find the relevant info in the last paragraph in this docs section, let me shamelessly paste from a previous dival issue:

the error is about the backend implementation for the radon transform. The default is 'astra_cuda', which requires both the astra-toolbox being installed and a cuda-enabled GPU. For testing, you can pass the option impl='skimage' to get_standard_dataset. But for real use, this will probably be too slow.If you are using anaconda, the latest development version of astra can be installed with

conda install astra-toolbox -c astra-toolbox/label/dev

Then the backends 'astra_cpu' and 'astra_cuda' should become available ('astra_cuda' only if cuda is available).

Hope this helps and please let me know if there are any further troubles to get the code running.

Guaishou74851 commented 3 years ago

Hi, thank you very much for your helps. The correct configuration of astra-toolbox really solved my previous problem. My another question is, where can I find the hyperparameter setting (.json) files for the low-dose CT tasks (e.g. lodopab_fbp_hyper_params.json and lodopab_fbpunet_hyper_params.json)? When I want to test the recovery performance of FBP operator and run lodopab_eval.py again (with designating the method parameter as "fbp"), I got:

Traceback (most recent call last):
  File "lodopab_eval.py", line 94, in <module>
    reconstructor.load_params(os.path.join(RESULTS_PATH, name))
  File "/opt/conda/lib/python3.6/site-packages/dival/reconstructors/reconstructor.py", line 351, in load_params
    self.load_hyper_params(hp_path)
  File "/opt/conda/lib/python3.6/site-packages/dival/reconstructors/reconstructor.py", line 245, in load_hyper_params
    with open(path, 'r') as f:
FileNotFoundError: [Errno 2] No such file or directory: '../learned_ct_reco_comparison_paper_results/lodopab_fbp_hyper_params.json'

In fact, I traversed this GitHub repository and the related links but only found the trained networks for sparse-angle CT tasks on Apple-CT dataset and the reconstrution results of two tasks (LoDoPaB-CT Challenge Reconstructions and Apple CT Test Reconstructions, respectively). I don't know if I didn't carefully check all the related directories/files/docs/links and missed something, or actually the (hyperparameter configurations)/(trained networks) are in LoDoPaB-CT Challenge Reconstructions?

jleuschn commented 3 years ago

Hi,

The parameters for the dival reconstructors on the LoDoPaB-CT dataset can be found here (i.e. as a supplementary to dival rather than as a supplementary of this comparison paper repo). Probably this is not very easy to find, i'll add a link in the readme as well as on the supplementary zenodo record with the Apples-CT parameters.

Best regards and thanks for the feedback

Guaishou74851 commented 3 years ago

Hi,

With your helps, I can already perform complete low-dose CT tests on LoDoPaB-CT testset and get the results, but when I run the following command for testing FBPUNet: 

python lodopab_eval.py --method=fbpunet --part=test

, I got:

Traceback (most recent call last):
  File "lodopab_eval.py", line 95, in <module>
    reconstructor.load_params(os.path.join(RESULTS_PATH, name))
  File "/opt/conda/lib/python3.6/site-packages/dival/reconstructors/reconstructor.py", line 435, in load_params
    self.load_learned_params(path)
  File "/opt/conda/lib/python3.6/site-packages/dival/reconstructors/standard_learned_reconstructor.py", line 477, in load_learned_params
    state_dict = torch.load(path, map_location=map_location)
  File "/opt/conda/lib/python3.6/site-packages/torch/serialization.py", line 595, in load
    return _legacy_load(opened_file, map_location, pickle_module, **pickle_load_args)
  File "/opt/conda/lib/python3.6/site-packages/torch/serialization.py", line 764, in _legacy_load
    magic_number = pickle_module.load(f, **pickle_load_args)
_pickle.UnpicklingError: invalid load key, 'v'.

I put these files in a new created directory and test FBP operator successfully, so I think the problem may come from the reading process of the trained network (.pt) files by the Pickle Python module. I checked its version and my environment, but still don't know how to fix it. I guess that there are some mismatches about the format of serialized (.pt) file? By the way, is the reconstruction results of the LoDoPaB-CT testset (not the special challenge set) open access?

Best regards and thanks for your patience

jleuschn commented 3 years ago

Hi,

this is probably because the supp.dival git repo uses git-lfs. If git-lfs is not installed, the files are just small pointer text files. You could either manually download via the Github interface, or install git-lfs and checkout again. Note that the file lodopab_iradonmap.pt is quite large (approx. 0.5 GiB), so you might want to exclude it somehow.

Credits for considering it might be git-lfs related go to this SO answer.

We haven't published reconstructions on the LoDoPaB-CT test set yet. If you have troubles reproducing some results, just let me know.

Kind regards

jleuschn commented 3 years ago

In case you had trouble accessing the files due to my quota limit, i just upgraded the plan for the current month. You should be able to access the weights now.

Regards, Johannes