v0.11.5

[fcakyon]

What's Changed

• fixes a bug that prevents cuda device selection by @fcakyon in fix device selection, increase version #756

Full Changelog: 0.11.4...0.11.5

---

This discussion was created from the release v0.11.5.

[tureckova]

Hi,
I also run into an issue with device selection with mmdet model.
I am calling the function predict from my script and so far I did not set up the model_device parameter and everything works. But now I must set the model_device parameter to "cpu" or "cuda". If I skip the model_device parameter a seg fault occurs during epoch loading:

load checkpoint from local path: epoch_24.pth
Segmentation fault (core dumped)

I have the following environment packages versions:
sahi==11.5, mmcv==1.7.0, mmdet==2.26.0, torch==1.13.0+cu117.

The model_device parameter might be obligatory or at least print some warning if not used, since the seg fault is hard to track down and I spent quite some time discovering the cause.

[fcakyon]

@tureckova thanks for reporting. Can you tell me what is your pipeline using sahi so that I can recreate your issue?

Are you using predict function from sahi.predict or get_sliced_prediction function from sahi.predict?

[fcakyon]

@tureckova when you dont provide any device, sahi successfully selects cuda:0. You don't have to set it as cpu or cuda. I have just tested it with sahi==0.11.5, mmcv==1.7.0, mmdet==2.26.0, torch==1.12.1+cu113.

Maybe mmdet doesn't support torch 1.13 and cuda 11.7. Have you tried it with lower pytorch and cuda versions? (sahi is tested for cuda 11.3 and pytorch 1.12)

[tureckova]

Well, it is weird. It might be connected to cuda version since I encountered the problem after the upgrade.
If I activate debug mode, it works:

But if I just run the script seg fault occurs.

I tracked down your code (viz the printscreens), and it correctly gives device cuda:0, but in mmdet inference.py on line model.to(device) it is killed. I don't have time to dig deeper, but mmdet default train and test scripts works well for me. So I thought the bug is in sahi. I am pretty confused about where the bug came from.

[tureckova]

Ou, and it works on CPU well (but slowly).

[fcakyon]

The reason might be related to your pytorch and cuda versions

[yzfzzz]

This doesn't seem like a version issue, because my torch=1.11, python=3.8, mmcv=1.6.0, it also goes wrong

[yzfzzz]

when cu=116, do you have a test?

[fcakyon]

mmcv most probably dont support cu11.6. we currently only test for cu11.3, in near future: will be starting to test in cu11.4.

[yzfzzz]

Even though I use the same version as you, this problem still occurs and I think it may be related to cuda

This is my test code

from sahi.predict import predict
from sahi import AutoDetectionModel

import faulthandler


faulthandler.enable()

# get batch predict result
result = predict(
    model_type='mmdet', # one of 'yolov5', 'mmdet', 'detectron2'
    model_path= '../mmdetection/tools/work_dirs/cascade_rcnn_val_r50_fpn_120x_coco_panda/best_bbox_mAP_epoch_115.pth', # path to model weight file
    model_config_path='../mmdetection/configs/panda/cascade_rcnn_r50_fpn_1x_coco_round1_panda.py', # for detectron2 and mmdet models
    model_confidence_threshold=0.5,
    model_device='cuda:0', # or 'cuda:0'
    source="./IMG_14_01.jpg", # image or folder path
    no_standard_prediction=True,
    no_sliced_prediction=False,
    slice_height=3000,
    slice_width=3000,
    overlap_height_ratio=0.5,
    overlap_width_ratio=0.5,
    export_pickle=False,
    export_crop=False,
)

[fcakyon]

Then it might be related with your ide 🤔 I am not having any issues debugging with cuda 11.3 on vs code ide.

[kimlia545]

I had the same problem.
torch '1.9.0+cu111' mmdet '2.25.1' mmcv '1.6.0'
In get_sliced_prediction device is "cuda" work, but in predict, when using "cuda", "Segmentation fault (core dumped)" error occurs and it works with "cpu".

[fcakyon]

What is your sahi version?

[kimlia545]

@fcakyon
sahi '0.11.13'

[fcakyon]

Please try with latest sahi