mdistiller

Repo ini

(1) Perpustakaan Pytorch yang menyediakan algoritma distilasi pengetahuan klasik pada tolok ukur CV utama,

(2) Implementasi resmi makalah CVPR-2022: Distilasi Pengetahuan Dipoupled.

(3) Implementasi resmi kertas ICCV-2023: DOT: Pelatih berorientasi distilasi.

Dot: Pelatih yang berorientasi pada penyulingan

Kerangka

Hasil Benchmark Utama

Di CIFAR-100:

Guru Murid	Resnet32x4 Resnet8x4	VGG13 VGG8	Resnet32x4 Shufflenet-v2
KD	73.33	72.98	74.45
KD+DOT	75.12	73.77	75.55

Di Tiny-Imagenet:

Guru Murid	Resnet18 MobileNet-V2	Resnet18 Shufflenet-v2
KD	58.35	62.26
KD+DOT	64.01	65.75

Di imagenet:

Guru Murid	Resnet34 Resnet18	Resnet50 MobileNet-V1
KD	71.03	70.50
KD+DOT	71.72	73.09

Distilasi pengetahuan yang dipisahkan

Kerangka & kinerja

Hasil Benchmark Utama

Di CIFAR-100:

Guru Murid	Resnet56 Resnet20	Resnet110 Resnet32	Resnet32x4 Resnet8x4	WRN-40-2 WRN-16-2	WRN-40-2 WRN-40-1	VGG13 VGG8
KD	70.66	73.08	73.33	74.92	73.54	72.98
DKD	71.97	74.11	76.32	76.23	74.81	74.68

Guru Murid	Resnet32x4 Shufflenet-v1	WRN-40-2 Shufflenet-v1	VGG13 MobileNet-V2	Resnet50 MobileNet-V2	Resnet32x4 MobileNet-V2
KD	74.07	74.83	67.37	67.35	74.45
DKD	76.45	76.70	69.71	70.35	77.07

Di imagenet:

Guru Murid	Resnet34 Resnet18	Resnet50 MobileNet-V1
KD	71.03	70.50
DKD	71.70	72.05

Mdistiller

Perkenalan

MDistiller mendukung metode distilasi berikut pada CIFAR-100, Imagenet dan MS-COCO:

Metode	Tautan kertas	CIFAR-100	Imagenet	MS-COCO
KD	https://arxiv.org/abs/1503.02531	✓	✓
Fitnet	https://arxiv.org/abs/1412.6550	✓
PADA	https://arxiv.org/abs/1612.03928	✓	✓
Nst	https://arxiv.org/abs/1707.01219	✓
PKT	https://arxiv.org/abs/1803.10837	✓
Kdsvd	https://arxiv.org/abs/1807.06819	✓
Ofd	https://arxiv.org/abs/1904.01866	✓	✓
Rkd	https://arxiv.org/abs/1904.05068	✓
Vid	https://arxiv.org/abs/1904.05835	✓
Sp	https://arxiv.org/abs/1907.09682	✓
CRD	https://arxiv.org/abs/1910.10699	✓	✓
Ulasankd	https://arxiv.org/abs/2104.09044	✓	✓	✓
DKD	https://arxiv.org/abs/2203.08679	✓	✓	✓

Instalasi

Lingkungan:

• Python 3.6
• Pytorch 1.9.0
• Torchvision 0.10.0

Instal paket:

 sudo pip3 install -r requirements.txt
sudo python3 setup.py develop

Memulai

0. Wandb sebagai penebang

• Registerasi: https://wandb.ai/home.
• Jika Anda tidak ingin Wandb sebagai logger Anda, atur CFG.LOG.WANDB sebagai False di mdistiller/engine/cfg.py .

1. Evaluasi

• Anda dapat mengevaluasi kinerja model atau model kami yang dilatih sendiri.

• Model kami berada di https://github.com/megvii-research/mdistiller/releases/tag/checkpoints, silakan unduh pos pemeriksaan ke ./download_ckpts

• Jika menguji model di imagenet, silakan unduh dataset di https://image-net.org/ dan letakkan di ./data/imagenet

   # evaluate teachers
  python3 tools/eval.py -m resnet32x4 # resnet32x4 on cifar100
  python3 tools/eval.py -m ResNet34 -d imagenet # ResNet34 on imagenet
  
  # evaluate students
  python3 tools/eval.p -m resnet8x4 -c download_ckpts/dkd_resnet8x4 # dkd-resnet8x4 on cifar100
  python3 tools/eval.p -m MobileNetV1 -c download_ckpts/imgnet_dkd_mv1 -d imagenet # dkd-mv1 on imagenet
  python3 tools/eval.p -m model_name -c output/your_exp/student_best # your checkpoints

2. Pelatihan tentang CIFAR-100

• Unduh cifar_teachers.tar di https://github.com/megvii-research/mdistiller/releases/tag/checkpoints dan untarnya ke ./download_ckpts melalui tar xvf cifar_teachers.tar .

   # for instance, our DKD method.
  python3 tools/train.py --cfg configs/cifar100/dkd/res32x4_res8x4.yaml
  
  # you can also change settings at command line
  python3 tools/train.py --cfg configs/cifar100/dkd/res32x4_res8x4.yaml SOLVER.BATCH_SIZE 128 SOLVER.LR 0.1

3. Pelatihan tentang Imagenet

• Unduh dataset di https://image-net.org/ dan letakkan di ./data/imagenet

   # for instance, our DKD method.
  python3 tools/train.py --cfg configs/imagenet/r34_r18/dkd.yaml

4. Pelatihan tentang MS-COCO

• Lihat Detection.md

5. Ekstensi: Visualisasi

• Jupyter Notebook: TSNE dan CORRELATION_MATRICES

Metode Distilasi Kustom

1. Buat file python di mdistiller/distillers/ dan tentukan penyuling

 from . _base import Distiller

class MyDistiller ( Distiller ):
    def __init__ ( self , student , teacher , cfg ):
        super ( MyDistiller , self ). __init__ ( student , teacher )
        self . hyper1 = cfg . MyDistiller . hyper1
        ...

    def forward_train ( self , image , target , ** kwargs ):
        # return the output logits and a Dict of losses
        ...
    # rewrite the get_learnable_parameters function if there are more nn modules for distillation.
    # rewrite the get_extra_parameters if you want to obtain the extra cost.
  ...

2. Mendaftarkan penyuling di distiller_dict di mdistiller/distillers/__init__.py

3. Mendaftarkan hyper-parameter yang sesuai di mdistiller/engines/cfg.py

4. Buat file konfigurasi baru dan uji.

Kutipan

Jika repo ini bermanfaat untuk penelitian Anda, harap pertimbangkan mengutip kertas:

 @article { zhao2022dkd ,
  title = { Decoupled Knowledge Distillation } ,
  author = { Zhao, Borui and Cui, Quan and Song, Renjie and Qiu, Yiyu and Liang, Jiajun } ,
  journal = { arXiv preprint arXiv:2203.08679 } ,
  year = { 2022 }
}
@article { zhao2023dot ,
  title = { DOT: A Distillation-Oriented Trainer } ,
  author = { Zhao, Borui and Cui, Quan and Song, Renjie and Liang, Jiajun } ,
  journal = { arXiv preprint arXiv:2307.08436 } ,
  year = { 2023 }
}

Lisensi

MDistiller dirilis di bawah lisensi MIT. Lihat lisensi untuk detailnya.

Pengakuan

• Terima kasih untuk CRD dan ReviewKD. Kami membangun perpustakaan ini berdasarkan basis kode CRD dan basis kode ulasan.

• Terima kasih Yiyu Qiu dan Yi Shi atas kontribusi kode selama magang mereka dalam teknologi Megvii.

• Terima kasih Xin Jin untuk diskusi tentang DKD.