Skip to content

README.md

Latest commit

 

History

History
222 lines (179 loc) · 14.3 KB

README.md

File metadata and controls

222 lines (179 loc) · 14.3 KB

PIIP for Detection

This folder contains code for applying PIIP on object detection and instance segmentation, developed on top of MMDetection v2.25.3.

The released model weights are provided in the parent folder.

Installation

  • Clone this repo:

    git clone https://github.com/OpenGVLab/PIIP
cd PIIP/

  • Create a conda virtual environment and activate it:

    conda create -n piip python=3.9 -y
conda activate piip

  • Install PyTorch>=1.11<2.0 and torchvision>=0.13.0 with CUDA>=10.2:

    For example, to install torch==1.12.0 with CUDA==11.3:

    pip install torch==1.12.0+cu113 torchvision==0.13.0+cu113 torchaudio==0.12.0 --extra-index-url https://download.pytorch.org/whl/cu113

  • Install flash-attn==0.2.8 :

    If you want to fully replicate my results, please install v0.2.8, otherwise install the latest version.

    This is because different versions of flash attention yield slight differences in results.

    git clone https://github.com/Dao-AILab/flash-attention.git
cd flash-attention
git checkout v0.2.8
pip install ninja
python setup.py install # I use gcc-7.3 to compile this package
# for FusedMLP
cd csrc/fused_dense_lib
pip install .

  • Install other requirements:

    conda install -c conda-forge termcolor yacs pyyaml scipy pip -y
pip install opencv-python
pip install timm==0.6.11
pip install yapf==0.40.1
pip install addict
pip install deepspeed==0.8.0 # please install this old version
pip install pydantic==1.10.2 # later versions may have compatibility issues

  • Install our customized mmcv-full==1.7.0:

    cd mmcv/
export MMCV_WITH_OPS=1
python setup.py develop
cd ../

  • Install our customized mmdetection:

    cd mmdetection/
python setup.py develop

  • Compile the deformable attention:

    cd mmdetection/ops
sh compile.sh

  • Selectively download pretrained ViT weights from DeiT, DeiT III, MAE, InternVL, BEiTv2, DINOv2, AugReg and Uni-Perceiver:

    mkdir mmdetection/pretrained
cd mmdetection/pretrained
# Default: DeiT III and DeiT
wget https://dl.fbaipublicfiles.com/deit/deit_tiny_patch16_224-a1311bcf.pth
wget https://dl.fbaipublicfiles.com/deit/deit_3_small_224_21k.pth
wget https://dl.fbaipublicfiles.com/deit/deit_3_base_224_21k.pth
wget https://dl.fbaipublicfiles.com/deit/deit_3_large_224_21k.pth
# Optional - for InternViT-6B experiments
wget https://dl.fbaipublicfiles.com/mae/pretrain/mae_pretrain_vit_huge.pth
wget https://huggingface.co/OpenGVLab/InternVL/resolve/main/intern_vit_6b_224px.pth
# Optional - for different pretraining methods
# - MAE
wget https://dl.fbaipublicfiles.com/mae/pretrain/mae_pretrain_vit_base.pth
wget https://dl.fbaipublicfiles.com/mae/pretrain/mae_pretrain_vit_large.pth
# - BEiTv2
wget https://github.com/addf400/files/releases/download/BEiT-v2/beitv2_base_patch16_224_pt1k_ft21k.pth
wget https://github.com/addf400/files/releases/download/BEiT-v2/beitv2_large_patch16_224_pt1k_ft21k.pth
# - DINOv2
wget https://dl.fbaipublicfiles.com/dinov2/dinov2_vits14/dinov2_vits14_pretrain.pth
wget https://dl.fbaipublicfiles.com/dinov2/dinov2_vitb14/dinov2_vitb14_pretrain.pth
wget https://dl.fbaipublicfiles.com/dinov2/dinov2_vitl14/dinov2_vitl14_pretrain.pth
# - AugReg
wget https://huggingface.co/OpenGVLab/PIIP/resolve/main/detection/Ti_16-i21k-300ep-lr_0.001-aug_none-wd_0.03-do_0.0-sd_0.0--imagenet2012-steps_20k-lr_0.03-res_384.pth?download=true
wget https://huggingface.co/OpenGVLab/PIIP/resolve/main/detection/S_16-i21k-300ep-lr_0.001-aug_light1-wd_0.03-do_0.0-sd_0.0--imagenet2012-steps_20k-lr_0.03-res_384.pth?download=true
wget https://huggingface.co/OpenGVLab/PIIP/resolve/main/detection/B_16-i21k-300ep-lr_0.001-aug_medium1-wd_0.1-do_0.0-sd_0.0--imagenet2012-steps_20k-lr_0.01-res_384.pth?download=true
wget https://huggingface.co/OpenGVLab/PIIP/resolve/main/detection/L_16-i21k-300ep-lr_0.001-aug_medium1-wd_0.1-do_0.1-sd_0.1--imagenet2012-steps_20k-lr_0.01-res_384.pth?download=true
# - Uni-Perceiver
wget https://github.com/czczup/ViT-Adapter/releases/download/v0.3.1/uni-perceiver-base-L12-H768-224size-torch-pretrained_converted.pth
wget https://github.com/czczup/ViT-Adapter/releases/download/v0.3.1/uni-perceiver-large-L24-H1024-224size-pretrained_converted.pth

  • Prepare the COCO dataset according to the MMDetection guidelines.

Note: the core model code is under mmdet/models/backbones/.

Training

To train PIIP-TSB Mask R-CNN on COCO train2017 on a single node with 8 gpus for 12 epochs run:

sh tools/dist_train.sh configs/piip/3branch/mask_rcnn_beit_tsb_1120_896_448_fpn_1x_coco_bs16.py 8
# or manage jobs with slurm
GPUS=8 sh tools/slurm_train.sh <partition> <job-name> configs/piip/3branch/mask_rcnn_beit_tsb_1120_896_448_fpn_1x_coco_bs16.py

Note: You can modify the deepspeed parameter in the configuration file to decide whether to use deepspeed. If you want to resume the deepspeed pretrained model for finetuning, you need to set deepspeed_load_module_only=True in the config.

Evaluation

To evaluate PIIP-H6B Mask R-CNN on COCO val2017 on a single node with a single gpu:

# w/ deepspeed
python tools/test.py configs/piip/2branch/mask_rcnn_internvit_h6b_1024_512_fpn_1x_coco_bs16_ms work_dirs/mask_rcnn_internvit_h6b_1024_512_fpn_1x_coco_bs16_ms/iter_87961/global_step87960 --eval bbox segm
# w/ deepspeed and set `deepspeed=False` in the configuration file
python tools/test.py configs/piip/2branch/mask_rcnn_internvit_h6b_1024_512_fpn_1x_coco_bs16_ms work_dirs/mask_rcnn_internvit_h6b_1024_512_fpn_1x_coco_bs16_ms/iter_87961/global_step87960/mp_rank_00_model_states.pt --eval bbox segm
# w/o deepspeed
python tools/test.py configs/piip/2branch/mask_rcnn_internvit_h6b_1024_512_fpn_1x_coco_bs16_ms work_dirs/mask_rcnn_internvit_h6b_1024_512_fpn_1x_coco_bs16_ms/mask_rcnn_internvit_h6b_1024_512_fpn_1x_coco_bs16_ms.pth --eval bbox segm

To evaluate PIIP-H6B Mask R-CNN on COCO val2017 on a single node with 8 gpus:

# w/ deepspeed
sh tools/dist_test.sh configs/piip/2branch/mask_rcnn_internvit_h6b_1024_512_fpn_1x_coco_bs16_ms work_dirs/mask_rcnn_internvit_h6b_1024_512_fpn_1x_coco_bs16_ms/iter_87961/global_step87960 8 --eval bbox segm
# w/ deepspeed and set `deepspeed=False` in the configuration file
sh tools/dist_test.sh configs/piip/2branch/mask_rcnn_internvit_h6b_1024_512_fpn_1x_coco_bs16_ms work_dirs/mask_rcnn_internvit_h6b_1024_512_fpn_1x_coco_bs16_ms/iter_87961/global_step87960/mp_rank_00_model_states.pt 8 --eval bbox segm
# w/o deepspeed
sh tools/dist_test.sh configs/piip/2branch/mask_rcnn_internvit_h6b_1024_512_fpn_1x_coco_bs16_ms work_dirs/mask_rcnn_internvit_h6b_1024_512_fpn_1x_coco_bs16_ms/mask_rcnn_internvit_h6b_1024_512_fpn_1x_coco_bs16_ms.pth 8 --eval bbox segm

This should give

Evaluating bbox...
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.558
 Average Precision  (AP) @[ IoU=0.50      | area=   all | maxDets=1000 ] = 0.773
 Average Precision  (AP) @[ IoU=0.75      | area=   all | maxDets=1000 ] = 0.614
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= small | maxDets=1000 ] = 0.396
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=1000 ] = 0.605
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= large | maxDets=1000 ] = 0.711
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.668
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=300 ] = 0.668
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=1000 ] = 0.668
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= small | maxDets=1000 ] = 0.509
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=1000 ] = 0.715
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= large | maxDets=1000 ] = 0.808

Evaluating segm...
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.490
 Average Precision  (AP) @[ IoU=0.50      | area=   all | maxDets=1000 ] = 0.743
 Average Precision  (AP) @[ IoU=0.75      | area=   all | maxDets=1000 ] = 0.531
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= small | maxDets=1000 ] = 0.285
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=1000 ] = 0.531
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= large | maxDets=1000 ] = 0.685
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.593
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=300 ] = 0.593
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=1000 ] = 0.593
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= small | maxDets=1000 ] = 0.420
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=1000 ] = 0.646
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= large | maxDets=1000 ] = 0.754

FLOPs calculation

We provide a simple script to calculate the number of FLOPs. Change the config_list in ../classification/get_flops.py and run

# use the classification environment
conda activate piip_cls
cd ../classification/
python get_flops.py

Then the FLOPs and number of parameters are recorded in flops.txt.

Faster Training

  • Use FusedMLP and discard Projection and LayerNorm in interactions by setting with_proj=False, norm_layer="none" of backbone and mlp_type="fused_mlp" of each branch in the config file. These changes do not affect performance.
Backbone Detector Pretrain Resolution Schd Box mAP Mask mAP Training Time #FLOPs #Param Download
PIIP-SBL w/o Mask R-CNN AugReg 1568/1120/672 1x 48.3 42.6 23h 1874G 498M log | ckpt | cfg
PIIP-SBL w/ Mask R-CNN AugReg 1568/1120/672 1x 48.1 42.4 21h 1741G 470M log | ckpt | cfg
InternViT-6B Mask R-CNN InternVL 1024 1x 53.8 48.1 3d19h 29323G 5919M log | ckpt | cfg
PIIP-H6B w/o Mask R-CNN MAE (H) + InternVL (6B) 1024/512 1x 55.8 49.0 3d14h 11080G 6872M log | ckpt | cfg
PIIP-H6B w/ Mask R-CNN MAE (H) + InternVL (6B) 1024/512 1x 55.6 49.0 2d14h 10692G 6752M log | ckpt | cfg

  • Use fewer interactions, e.g. set interaction_indexes=[[0,5],[6,11]] for ViT-T/S/B and interaction_indexes=[[0,11],[12,23]] for ViT-L. Figure 5(b) shows that when #Interaction >= 2, the growing trends of performance against computation cost are similar.

  • Training speed of multi-branch models can be affected by many factors, e.g. GPU utilization or optimization techniques. Further optimization might be necessary to enhance training efficiency.

Citation

If you find this work helpful for your research, please consider giving this repo a star ⭐ and citing our paper:

@article{piip,
  title={Parameter-Inverted Image Pyramid Networks},
  author={Zhu, Xizhou and Yang, Xue and Wang, Zhaokai and Li, Hao and Dou, Wenhan and Ge, Junqi and Lu, Lewei and Qiao, Yu and Dai, Jifeng},
  journal={arXiv preprint arXiv:2406.04330},
  year={2024}
}