Downstream KUMC test compile issue

[zkysss11235]

I followed the guide in ReadMe and compile the STFT using a RTX4090. It successfully compiled but, when I run the finetuning, it outputs the following error:

Traceback (most recent call last):
File "tools/train_net.py", line 16, in
from stft_core.engine.inference import inference
File "/home/kyz/repos/Endo-FM/STFT/stft_core/engine/inference.py", line 9, in
from stft_core.data.datasets.evaluation import evaluate
File "/home/kyz/repos/Endo-FM/STFT/stft_core/data/datasets/evaluation/init.py", line 3, in
from .vid import vid_evaluation
File "/home/kyz/repos/Endo-FM/STFT/stft_core/data/datasets/evaluation/vid/init.py", line 7, in
from .vid_eval import do_vid_evaluation
File "/home/kyz/repos/Endo-FM/STFT/stft_core/data/datasets/evaluation/vid/vid_eval.py", line 11, in
from stft_core.structures.boxlist_ops import boxlist_iou
File "/home/kyz/repos/Endo-FM/STFT/stft_core/structures/boxlist_ops.py", line 6, in
from stft_core.layers import nms as _box_nms
File "/home/kyz/repos/Endo-FM/STFT/stft_core/layers/init.py", line 10, in
from .nms import nms
File "/home/kyz/repos/Endo-FM/STFT/stft_core/layers/nms.py", line 3, in
from stft_core import _C
ImportError: /home/kyz/repos/Endo-FM/STFT/stft_core/_C.cpython-37m-x86_64-linux-gnu.so: undefined symbol: PyCMethod_New
Killing subprocess 870160
Traceback (most recent call last):
File "/data/kyz/anaconda3/envs/endofm/lib/python3.7/runpy.py", line 193, in _run_module_as_main
"main", mod_spec)
File "/data/kyz/anaconda3/envs/endofm/lib/python3.7/runpy.py", line 85, in _run_code
exec(code, run_globals)
File "/data/kyz/anaconda3/envs/endofm/lib/python3.7/site-packages/torch/distributed/launch.py", line 340, in
main()
File "/data/kyz/anaconda3/envs/endofm/lib/python3.7/site-packages/torch/distributed/launch.py", line 326, in main
sigkill_handler(signal.SIGTERM, None) # not coming back
File "/data/kyz/anaconda3/envs/endofm/lib/python3.7/site-packages/torch/distributed/launch.py", line 301, in sigkill_handler
raise subprocess.CalledProcessError(returncode=last_return_code, cmd=cmd)
subprocess.CalledProcessError: Command '['/data/kyz/anaconda3/envs/endofm/bin/python', '-u', 'tools/train_net.py', '--local_rank=0', '--master_port=15153', '--config-file', 'configs/STFT/kumc_R_50_STFT.yaml', 'OUTPUT_DIR', '/home/kyz/annotations+scripts+outputs/outputs/Endo-FM/downstream']' returned non-zero exit status 1.

Do I need to change anything in the compile setting?

[Kyfafyd]

Hi @zkysss11235 thanks for your interest!
I have not met such problem before. It looks like the compilation is not totally successful. May you refer to this pybind/pybind11#3115 and try to build an environment with another python version?

[10086ddd]

Hi @Kyfafyd thanks for your excellent work
I also encountered a problem when I used my own medical dataset containing 1500 endoscopic images to train STFT, but the result I got was overfitting. Does this mean that training it requires a larger dataset?

[Kyfafyd]

@10086ddd thanks for your interest!
May I learn how you split the training and test sets?

[10086ddd]

@Kyfafyd thanks for your reply
Actually, I manually divided it according to the format of the Kumc dataset.

[Kyfafyd]

@10086ddd what about the performance on the training dataset?

[10086ddd]

@Kyfafyd similar to the results above, it is also overfitting

[Kyfafyd]

@10086ddd If the performance on the test set is 100%, this is not overfitting (I think). This indicates the high effectiveness and robustness of the model.

[10086ddd]

@Kyfafyd thanks for your answer, but why didn't it show such performance on the Kumc dataset?

[Kyfafyd]

@10086ddd KUMC dataset is more difficult:)