SwinTransformer 學習紀錄

參考 https://github.com/microsoft/Swin-Transformer?tab=readme-ov-file

參考 https://github.com/SwinTransformer/Swin-Transformer-Object-Detection

參考 https://mmdetection.readthedocs.io/en/latest/get_started.html

參考 https://github.com/open-mmlab/mmdetection/blob/master/docs/en/get_started.md

參考 https://github.com/open-mmlab/mmcv

參考 https://github.com/open-mmlab/mim?tab=readme-ov-file

Swin-Transformer 是原始版本，主要針對圖片分類

Swin-Transformer-Object-Detection 主要針對物件偵測(基於 mmdetection)

mmdetection 主要針對物件偵測，其中不只是 Swin-Transformer, 包含各式各樣先進的演算法

mmdetection 可在 mmdet/__init__.py 可以查詢到 mmcv 所需要的版本

Swin-Transformer-Object-Detection 要求使用 mmcv(1.4.0)

但 mmdetection 最後對應到 mmcv(1.4.0) 的版本為(v2.18.1)

所以以下放棄使用 Swin-Transformer-Object-Detection, 直接使用 mmdetection 的最新版本

$ docker run --gpus all -it --rm nvcr.io/nvidia/pytorch:21.05-py3

$ docker run --gpus all -it --name SwinTransformer nvcr.io/nvidia/pytorch:21.05-py3

$ docker start SwinTransformer

$ docker attach SwinTransformer

# <ctrl+p><ctrl+q>

$ docker attach SwinTransformer

$ docker stop SwinTransformer

$ docker rm SwinTransformer

$ git clone https://github.com/microsoft/Swin-Transformer.git

$ git clone https://github.com/SwinTransformer/Swin-Transformer-Object-Detection.git

$ git clone https://github.com/open-mmlab/mmdetection.git

$ docker run --gpus all -it --name SwinTransformer --shm-size=8G \

  -v /mnt/Data/SwinTransformer/Swin-Transformer:/workspace/Swin-Transformer \

  -v /mnt/Data/SwinTransformer/mmdetection:/workspace/mmdetection \

  -v /mnt/Data/SwinTransformer/Swin-Transformer-Object-Detection:/workspace/Swin-Transformer-Object-Detection \

  -v /mnt/QNAP_A/ImageData/ImageNet:/workspace/ImageNet \

  nvcr.io/nvidia/pytorch:21.05-py3

# pip install timm==0.4.12

# pip install opencv-python==4.4.0.46 termcolor==1.1.0 yacs==0.1.8 pyyaml scipy

# cd /workspace/Swin-Transformer/kernels/window_process/

# python setup.py install

# cd /workspace/Swin-Transformer/

# wget https://github.com/SwinTransformer/storage/releases/download/v1.0.0/swin_tiny_patch4_window7_224.pth

# python -m torch.distributed.launch --nproc_per_node 1 --master_port 12345 main.py \

  --eval --cfg configs/swin/swin_tiny_patch4_window7_224.yaml \

  --resume swin_tiny_patch4_window7_224.pth --data-path /workspace/ImageNet \

  --batch_size=64

#### 出現下列錯誤

####  File "/opt/conda/lib/python3.8/site-packages/PIL/_typing.py", line 10, in <module>

####    NumpyArray = npt.NDArray[Any]

####AttributeError: module 'numpy.typing' has no attribute 'NDArray'

#### 解決

# pip install Pillow==9.5.0

#### 出現下列錯誤

####RuntimeError: Found 0 files in subfolders of: /workspace/ImageNet/val

####Supported extensions are: .jpg,.jpeg,.png,.ppm,.bmp,.pgm,.tif,.tiff,.webp

#### 解決

$ cd /mnt/QNAP_A/ImageData/ImageNet/

$ mv val val_a ;mkdir val; mv val_a val

#### 出現下列錯誤

ERROR: Unexpected bus error encountered in worker. This might be caused by insufficient shared memory (shm).

#### 解決 docker run 時須加上 --shm-size=8G

# cd /workspace/mmdetection

# pip install -U openmim

#### 出現下列錯誤

####ERROR: After October 2020 you may experience errors when installing or updating packages. This is because pip will change the way that it resolves dependency conflicts.

####We recommend you use --use-feature=2020-resolver to test your packages with the new resolver before it becomes the default.

#### 不用擔心

# mim list

# mim install mmengine

# mim install "mmcv>=2.0.0"

# pip install -v -e .

    # "-v" means verbose, or more output

    # "-e" means installing a project in editable mode,

    # thus any local modifications made to the code will take effect without reinstallation.

# mim download mmdet --config yolov3_mobilenetv2_8xb24-320-300e_coco --dest .

# mim list

# ls

# python demo/image_demo.py demo/demo.jpg yolov3_mobilenetv2_8xb24-320-300e_coco.py \

  --weights yolov3_mobilenetv2_320_300e_coco_20210719_215349-d18dff72.pth

#### 出現下列錯誤

####    import cv2

####  File "/opt/conda/lib/python3.8/site-packages/cv2/__init__.py", line 5, in <module>

####    from .cv2 import *

####ImportError: libGL.so.1: cannot open shared object file: No such file or directory

#### 出現下列錯誤

####    cv.gapi.wip.GStreamerPipeline = cv.gapi_wip_gst_GStreamerPipeline

####AttributeError: partially initialized module 'cv2' has no attribute 'gapi_wip_gst_GStreamerPipeline' (most likely due to a circular import)

#### 解決

# pip install opencv-python-headless==4.4.0.46

#### 出現下列錯誤

####    assert (mmcv_version >= digit_version(mmcv_minimum_version)

####AssertionError: MMCV==2.2.0 is used but incompatible. Please install mmcv>=2.0.0rc4, <2.2.0.

#### 解決

# mim install "mmcv==2.0.0rc4"

# mim download mmdet --config mask-rcnn_swin-t-p4-w7_fpn_1x_coco --dest .

# python demo/image_demo.py demo/demo.jpg mask-rcnn_swin-t-p4-w7_fpn_1x_coco.py \

  --weights mask_rcnn_swin-t-p4-w7_fpn_1x_coco_20210902_120937-9d6b7cfa.pth