Detectron2: why Inference Demo with Colab trained Models detected 0 instances

[sswhqqw]

When I used command:
python demo/demo.py \
--config-file configs/COCO-InstanceSegmentation/mask_rcnn_R_50_FPN_3x.yaml \
--input /Users/wenzq/Downloads/detectron2/demo/horseface01601.jpg --output outputs/ \
--opts MODEL.WEIGHTS '/Users/wenzq/Downloads/detectron2/train_model/model_final_f10217.pkl' \
MODEL.DEVICE cpu
result in :
detected 3 instances


note:
model_final_f10217.pkl is the model Pre-trained Models

When I used command:
python demo/demo.py \
--config-file configs/COCO-InstanceSegmentation/mask_rcnn_R_50_FPN_3x.yaml \
--input /Users/wenzq/Downloads/detectron2/demo/horseface01601.jpg --output outputs/ \
--opts MODEL.WEIGHTS '/Users/wenzq/Downloads/detectron2/train_model/model_final-3.pth' \
MODEL.DEVICE cpu
detected 0 instances

note:
model_final-3.pth is the model trained on colab which can predict instance nomally on colab.

[davidmurray]

I'm also getting a similar issue.

python demo.py --config-file ../configs/COCO-InstanceSegmentation/mask_rcnn_R_50_FPN_3x.yaml --input C:\Users\dmurray\code\analyses_images\assets\man-horse.jpg --confidence-threshold 0.1 --opts MODEL.DEVICE cpu

The above command yields 0 detections in 2s.

Note: I am running on CPU because I do not have a GPU.

(The image 'man-horse.jpg' is the following: http://images.cocodataset.org/val2017/000000439715.jpg)

Attached is the output when running demo.py:
output.txt

Versions

------------------------  ------------------------------------------------------------------------------------
sys.platform              win32
Python                    3.7.7 (default, Apr 15 2020, 05:09:04) [MSC v.1916 64 bit (AMD64)]
numpy                     1.18.1
detectron2                0.1.2 @C:\Users\dmurray\Anaconda3\envs\analyses_images\lib\site-packages\detectron2
detectron2 compiler       MSVC 192528614
detectron2 CUDA compiler  not available
DETECTRON2_ENV_MODULE     <not set>
PyTorch                   1.5.0 @C:\Users\dmurray\Anaconda3\envs\analyses_images\lib\site-packages\torch
PyTorch debug build       False
CUDA available            False
Pillow                    7.1.2
torchvision               0.6.0 @C:\Users\dmurray\Anaconda3\envs\analyses_images\lib\site-packages\torchvision
fvcore                    0.1.1
cv2                       3.4.2
------------------------  ------------------------------------------------------------------------------------
PyTorch built with:
  - C++ Version: 199711
  - MSVC 191627039
  - Intel(R) Math Kernel Library Version 2020.0.0 Product Build 20191125 for Intel(R) 64 architecture applications
  - Intel(R) MKL-DNN v0.21.1 (Git Hash 7d2fd500bc78936d1d648ca713b901012f470dbc)
  - OpenMP 200203
  - CPU capability usage: AVX2
  - Build settings: BLAS=MKL, BUILD_TYPE=Release, CXX_FLAGS=/DWIN32 /D_WINDOWS  /GR  /w /EHa /bigobj -openmp -DNDEBUG -DUSE_FBGEMM, PERF_WITH_AVX=1, PERF_WITH_AVX2=1, PERF_WITH_AVX512=1, USE_CUDA=0, USE_EXCEPTION_PTR=1, USE_GFLAGS=OFF, USE_GLOG=OFF, USE_MKL=ON, USE_MKLDNN=ON, USE_MPI=OFF, USE_NCCL=OFF, USE_NNPACK=OFF, USE_OPENMP=ON, USE_STATIC_DISPATCH=OFF, 

Thanks.

[davidmurray]

Quick update:

Here is what I get when I run the following command:

python .\social_distancing\inference\inference.py --config-file COCO-Detection\faster_rcnn_R_50_C4_1x.yaml --image-input C:\Users\dmurray\code\analyses_images\assets\man-horse.jpg --confidence-threshold 0.1

Could it be because the image is not being loaded properly?

Note: I understand that detectron2 is not officially supported on Windows, but I have managed to get it to compile, so I am wondering why inference is not working.

Thanks,
David

[davidmurray]

Update: I fixed my problem. I was setting MODEL_WEIGHTS rather than MODEL.WEIGHTS on the config object. I was actually in fact using an untrained model.

[ppwwyyxx]

When loading the model for inference you need the same config, same number of classes, etc.

Also, as the issue template mentions:

If you expect the model to converge / work better, note that we do not give suggestions on how to train a new model. Only in one of the two conditions we will help with it: (1) You're unable to reproduce the results in detectron2 model zoo. (2) It indicates a detectron2 bug.