In recent years, deep neural network has been proved to be an effective method to solve problems in computer vision, natural language processing and other fields. By constructing an appropriate neural network and training it, the model performance will basically exceed traditional algorithms.
When data is sufficient, the model performance can be significantly improved by increasing the number of parameters through appropriate construction of the network, but this increases the model complexity. Large models are expensive to deploy in actual scenarios.
Redundancy exists in deep neural networks. At present, there are several methods to compress the model to reduce its parameter amount, e.g. pruning, quantization, knowledge distillation, etc. Knowledge distillation refers to the method that helps the training process of a smaller network (student) under the supervision of a larger network (teacher), so as to ensure the small model can obtain relatively large performance improvement and even obtain accuracy similar to large models without increasing parameters.
Knowledge distillation methods can be divided into three different categories: Response based distillation, Feature based distillation, Relation based distillation. Detailed introduction is as follows.
Knowledge distillation (KD) was first proposed by Hinton, who introduced KL divergence to the training loss function in addition to the cross entropy between the output logits and the ground truth labels. The accuracy of models trained with KD exceeds the accuracy of the same models trained only using ground truth loss. It should be noted that a larger teacher model needs to be trained first to guide the training process of student models.
PaddleClas proposed a simple yet effective SSLD algorithm [6], removing the dependence on ground truth labels. Combined with a large number of unlabeled data, the accuracy of pretrained models obtained from distillation on 18 models was improved by 3+%.
The aforementioned standard distillation method uses large models as teacher models to guide students to improve the performance. Later, Deep Mutual Learning (DML) distillation method [7] was proposed, i.e., two models with the same architecture learn from each other. Compared with KD and other knowledge distillation algorithms that rely on large teacher models, DML is independent of large teacher models. Such training process is simpler and more efficient.
Heo et al porposed OverHaul of Feature Distillation [8], in which feature map distance between the teacher and the student is calculated as distillation loss. Features of the student are transformed to match the shape of the teacher's features so that the distance can be computed.
Knowledge distillation methods based on feature map distance can be combined with response-based knowledge distillation algorithms mentioned in 3.1, i.e., the student's outputs and its middle feature maps are supervised simultaneously. For DML, such combination is even simpler, since the student's features can be aligned with the teacher's features without transformation. This method is used in PP-OCRv2, improving the accuracy of OCR models significantly.
Papers in 1.1.1 and 1.1.2 mainly consider the outputs and middle feature maps of student and teacher. These distillation algorithms focus on individual outputs and do not consider relations between individuals.
Park et al proposed RKD [10], a distillation algorithm based on relations. In RKD, mutual relations of data examples is considered, and two loss functions are used, the distance-wise distillation loss and angle-wise distillation loss.
The algorithm proposed in this paper, Relational Knowledge Distillation (RKD), transfers the structured relations between the output results obtained from the teacher model to the student model. Unlike the previous algorithms, which only focus on individual output results, the RKD algorithm uses two loss functions: distance-wise distillation loss and angle-wise distillation loss. In the final distillation loss function, both KD loss and RKD loss are considered. The final accuracy is better than that obtained by KD loss distillation only.
Paper:
Cheng Cui, Ruoyu Guo, Yuning Du, Dongliang He, Fu Li, Zewu Wu, Qiwen Liu, Shilei Wen, Jizhou Huang, Xiaoguang Hu, Dianhai Yu, Errui Ding, Yanjun Ma
arxiv, 2021
SSLD is a simple semi-supervised distillation method proposed by Baidu in 2021. By designing an improved JS divergence as the loss function and combining the data mining strategy based on ImageNet22k dataset, the accuracy of the 18 backbone network models was improved by more than 3% on average.
The SSLD configuration is shown below. In the Arch field, you need to define both the student model and the teacher model. The teacher model has fixed parameters, and the pretrained parameters are loaded. In the Loss field, you need to define DistillationDMLLoss as the training loss.
# model architecture
Arch:
name: "DistillationModel" # model name, here distillation model is used
class_num: &class_num 1000 # number of classes, 1000 for ImageNet1k
pretrained_list: # list of pretrained models, leave blank because specified in sub-models
freeze_params_list: # list of freezed params, networks of correspondent index are fixed if set to True
- True
- False
infer_model_name: "Student" # export Student sub-network when exporting model
models: # list of sub-networks
- Teacher: # teacher model
name: ResNet50_vd # model name
class_num: *class_num # number of classes
pretrained: True # pretrained model path, download official pretrained model if set to True
use_ssld: True # whether SSLD pretrained model is used (higher accuracy)
- Student: # student model
name: PPLCNet_x2_5 # model name
class_num: *class_num # number of classes
pretrained: False # pretrained model path, can be bool or string. Set to False here. Student model does not load the pretrained model by default
# loss function config for traing/eval process
Loss: # loss function
Train: # list of training losses
- DistillationDMLLoss: # distillation DMLLoss. DMLLoss is encapsulated to support loss function of distillation (in dict)
weight: 1.0 # weight of loss
model_name_pairs: # model pair used to compute loss. Here loss function between Student and Teacher is computed
- ["Student", "Teacher"]
Eval: # evaluation loss
- CELoss:
weight: 1.0Paper:
Ying Zhang, Tao Xiang, Timothy M. Hospedales, Huchuan Lu
CVPR, 2018
In the DML paper, the process of distillation does not depend on a teacher model. Two models with the same architecture learn from each other and calculate the KL divergence of each other's logits.
Performance on ImageNet1k is shown below.
| Strategy | Backbone | Config | Top-1 acc | Download Link |
|---|---|---|---|---|
| baseline | PPLCNet_x2_5 | PPLCNet_x2_5.yaml | 74.93% | - |
| DML | PPLCNet_x2_5 | PPLCNet_x2_5_dml.yaml | 76.68%(+1.75%) | - |
- Note: Complete PPLCNet_x2_5 The model have been trained for 360 epochs. For comparison, both baseline and DML have been trained for 100 epochs. Therefore, the accuracy is lower than the model (76.60%) opened on the official website.
The DML configuration is shown below. In the Arch field, you need to define both the student model and the teacher model. Both models need to be updated. In the Loss field, you need to define DistillationDMLLoss (JS-Div between student and teacher) and DistillationGTCELoss (CE loss with ground truth labels) as the training loss.
Arch:
name: "DistillationModel"
class_num: &class_num 1000
pretrained_list:
freeze_params_list: # mutual learning, params of both models are not freezed
- False
- False
models:
- Teacher:
name: PPLCNet_x2_5 # mutual learning, so pretrained models are not loaded for both models
class_num: *class_num
pretrained: False
- Student:
name: PPLCNet_x2_5
class_num: *class_num
pretrained: False
Loss:
Train:
- DistillationGTCELoss: # CE loss with ground truth labels needs to be computed for both models because pretrained models are not loaded
weight: 1.0
model_names: ["Student", "Teacher"]
- DistillationDMLLoss:
weight: 1.0
model_name_pairs:
- ["Student", "Teacher"]
Eval:
- CELoss:
weight: 1.0UDML is a teacher-free knowledge distillation algorithm proposed by PaddleCV group. It is improved based on DML. In addition to the outputs, it also considers the middle layers features in the distillation process, so as to further improve the accuracy of knowledge distillation. For more information about UDML and its application, please refer to: PP-ShiTu paper and PP-OCRv3 paper.
Performance on ImageNet1k is shown below.
| Strategy | Backbone | Config | Top-1 acc | Download Link |
|---|---|---|---|---|
| baseline | PPLCNet_x2_5 | PPLCNet_x2_5.yaml | 74.93% | - |
| UDML | PPLCNet_x2_5 | PPLCNet_x2_5_dml.yaml | 76.74%(+1.81%) | - |
Arch:
name: "DistillationModel"
class_num: &class_num 1000
# if not null, its lengths should be same as models
pretrained_list:
# if not null, its lengths should be same as models
freeze_params_list:
- False
- False
models:
- Teacher:
name: PPLCNet_x2_5
class_num: *class_num
pretrained: False
# return_patterns means that in addition to the output logits, the middle feature maps with the corresponding names will also be returned
return_patterns: ["blocks3", "blocks4", "blocks5", "blocks6"]
- Student:
name: PPLCNet_x2_5
class_num: *class_num
pretrained: False
return_patterns: ["blocks3", "blocks4", "blocks5", "blocks6"]
# loss function config for traing/eval process
Loss:
Train:
- DistillationGTCELoss:
weight: 1.0
key: logits
model_names: ["Student", "Teacher"]
- DistillationDMLLoss:
weight: 1.0
key: logits
model_name_pairs:
- ["Student", "Teacher"]
- DistillationDistanceLoss: # distance loss based on features. Here l2 loss is used to calculate the distance between block5s
weight: 1.0
key: "blocks5"
model_name_pairs:
- ["Student", "Teacher"]
Eval:
- CELoss:
weight: 1.0Note(: return_patterns are specified in the network above. The function of returning middle layer features is based on TheseusLayer.
Paper:
Show, attend and distill: Knowledge distillation via attention-based feature matching
Mingi Ji, Byeongho Heo, Sungrae Park
AAAI, 2018
AFD proposes to use attention based meta network to learn the relative similarity between features in the distillation process, and apply the identified similarity relation to control the distillation intensity of all possible feature maps pairs.
Performance on ImageNet1k is shown below.
| Strategy | Backbone | Config | Top-1 acc | Download Link |
|---|---|---|---|---|
| baseline | ResNet18 | ResNet18.yaml | 70.8% | - |
| AFD | ResNet18 | resnet34_distill_resnet18_afd.yaml | 71.68%(+0.88%) | - |
Note: In order to keep alignment with the training configuration in the paper, the number of training iterations is set to be 100 epochs, so the baseline accuracy is lower than the open source model accuracy in PaddleClas (71.0%).
The AFD configuration is shown below. In the Arch field, you need to define both the student model and the teacher model. The teacher model has fixed parameters. In the Loss field, you need to define DistillationKLDivLoss (KL-Div between student and teacher), AFDLoss (AFD loss between student and teacher) and DistillationGTCELoss (CE loss with ground truth labels) as the training loss.
Arch:
name: "DistillationModel"
pretrained_list:
freeze_params_list:
models:
- Teacher:
name: AttentionModel # contains several serial networks. The following networks take outputs of previous networks as input
pretrained_list:
freeze_params_list:
- True
- False
models:
# basic network of AttentionModel
- ResNet34:
name: ResNet34
pretrained: True
# return_patterns means that in addition to the output logits, the middle feature maps with the corresponding names will also be returned
return_patterns: &t_keys ["blocks[0]", "blocks[1]", "blocks[2]", "blocks[3]",
"blocks[4]", "blocks[5]", "blocks[6]", "blocks[7]",
"blocks[8]", "blocks[9]", "blocks[10]", "blocks[11]",
"blocks[12]", "blocks[13]", "blocks[14]", "blocks[15]"]
# transformation network of AttentionModel. It transforms the sub-networks in the basic network
- LinearTransformTeacher:
name: LinearTransformTeacher
qk_dim: 128
keys: *t_keys
t_shapes: &t_shapes [[64, 56, 56], [64, 56, 56], [64, 56, 56], [128, 28, 28],
[128, 28, 28], [128, 28, 28], [128, 28, 28], [256, 14, 14],
[256, 14, 14], [256, 14, 14], [256, 14, 14], [256, 14, 14],
[256, 14, 14], [512, 7, 7], [512, 7, 7], [512, 7, 7]]
- Student:
name: AttentionModel
pretrained_list:
freeze_params_list:
- False
- False
models:
- ResNet18:
name: ResNet18
pretrained: False
return_patterns: &s_keys ["blocks[0]", "blocks[1]", "blocks[2]", "blocks[3]",
"blocks[4]", "blocks[5]", "blocks[6]", "blocks[7]"]
- LinearTransformStudent:
name: LinearTransformStudent
qk_dim: 128
keys: *s_keys
s_shapes: &s_shapes [[64, 56, 56], [64, 56, 56], [128, 28, 28], [128, 28, 28],
[256, 14, 14], [256, 14, 14], [512, 7, 7], [512, 7, 7]]
t_shapes: *t_shapes
infer_model_name: "Student"
# loss function config for traing/eval process
Loss:
Train:
- DistillationGTCELoss:
weight: 1.0
model_names: ["Student"]
key: logits
- DistillationKLDivLoss: # distillation KL-Div loss, features are extracted according to names in model_name_pairs to calculate loss
weight: 0.9 # weight of loss
model_name_pairs: [["Student", "Teacher"]]
temperature: 4
key: logits
- AFDLoss: # AFD loss
weight: 50.0
model_name_pair: ["Student", "Teacher"]
student_keys: ["bilinear_key", "value"]
teacher_keys: ["query", "value"]
s_shapes: *s_shapes
t_shapes: *t_shapes
Eval:
- CELoss:
weight: 1.0Note(: return_patterns are specified in the network above. The function of returning middle layer features is based on TheseusLayer.
Paper:
Decoupled Knowledge Distillation
Borui Zhao, Quan Cui, Renjie Song, Yiyu Qiu, Jiajun Liang
CVPR, 2022
DKD reformulates the classical KD loss into two parts, i.e., target class knowledge distillation (TCKD) and non-target class knowledge distillation (NCKD). The effect of the two parts is studied separately, and their weights can be adjusted independently, improving the accuracy and flexibility of distillation.
Performance on ImageNet1k is shown below.
| Strategy | Backbone | Config | Top-1 acc | Download Link |
|---|---|---|---|---|
| baseline | ResNet18 | ResNet18.yaml | 70.8% | - |
| DKD | ResNet18 | resnet34_distill_resnet18_dkd.yaml | 72.59%(+1.79%) | - |
The DKD configuration is shown below. In the Arch field, you need to define both the student model and the teacher model. The teacher model has fixed parameters, and the pretrained parameters are loaded. In the Loss field, you need to define DistillationDKDLoss (DKD loss between student and teacher) and DistillationGTCELoss (CE loss with ground truth labels) as the training loss.
Arch:
name: "DistillationModel"
# if not null, its lengths should be same as models
pretrained_list:
# if not null, its lengths should be same as models
freeze_params_list:
- True
- False
models:
- Teacher:
name: ResNet34
pretrained: True
- Student:
name: ResNet18
pretrained: False
infer_model_name: "Student"
# loss function config for traing/eval process
Loss:
Train:
- DistillationGTCELoss:
weight: 1.0
model_names: ["Student"]
- DistillationDKDLoss:
weight: 1.0
model_name_pairs: [["Student", "Teacher"]]
temperature: 1
alpha: 1.0
beta: 1.0
Eval:
- CELoss:
weight: 1.0Paper:
Knowledge Distillation from A Stronger Teacher
Tao Huang, Shan You, Fei Wang, Chen Qian, Chang Xu
2022, under review
When using the KD method for distillation, as the accuracy of the teacher model is improved, the effect of distillation is often difficult to improve simultaneously. This paper proposes the DIST method, which uses the Pearson correlation coefficient to represent the difference between the student model and the teacher model, instead of the default KL-divergence in the distillation process, so as to ensure that the model can learn more accurate correlation information.
Performance on ImageNet1k is shown below.
| Strategy | Backbone | Config | Top-1 acc | Download Link |
|---|---|---|---|---|
| baseline | ResNet18 | ResNet18.yaml | 70.8% | - |
| DIST | ResNet18 | resnet34_distill_resnet18_dist.yaml | 71.99%(+1.19%) | - |
The DIST configuration is shown below. In the Arch field, you need to define both the student model and the teacher model. The teacher model has fixed parameters, and the pretrained parameters are loaded. In the Loss field, you need to define DistillationDISTLoss (DIST loss between student and teacher) and DistillationGTCELoss (CE loss with ground truth labels) as the training loss.
Arch:
name: "DistillationModel"
# if not null, its lengths should be same as models
pretrained_list:
# if not null, its lengths should be same as models
freeze_params_list:
- True
- False
models:
- Teacher:
name: ResNet34
pretrained: True
- Student:
name: ResNet18
pretrained: False
infer_model_name: "Student"
# loss function config for traing/eval process
Loss:
Train:
- DistillationGTCELoss:
weight: 1.0
model_names: ["Student"]
- DistillationDISTLoss:
weight: 2.0
model_name_pairs:
- ["Student", "Teacher"]
Eval:
- CELoss:
weight: 1.0Paper:
Masked Generative Distillation
Zhendong Yang, Zhe Li, Mingqi Shao, Dachuan Shi, Zehuan Yuan, Chun Yuan
ECCV 2022
This method performs distillation on the feature map. In the process of distillation, random masks are applied to the features, and the students are forced to use some features to generate all the features of the teacher model, so as to improve the representation ability of the student model. MGD achieve state-of-the-art performance on the feature distillation task, and has been widely verified to be effective in tasks such as detection and segmentation.
Performance on ImageNet1k is shown below.
| Strategy | Backbone | Config | Top-1 acc | Download Link |
|---|---|---|---|---|
| baseline | ResNet18 | ResNet18.yaml | 70.8% | - |
| MGD | ResNet18 | resnet34_distill_resnet18_mgd.yaml | 71.86%(+1.06%) | - |
The MGD configuration is shown below. In the Arch field, you need to define both the student model and the teacher model. The teacher model has fixed parameters, and the pretrained parameters are loaded. In the Loss field, you need to define DistillationPairLoss (MGD loss between student and teacher) and DistillationGTCELoss (CE loss with ground truth labels) as the training loss.
Arch:
name: "DistillationModel"
class_num: &class_num 1000
# if not null, its lengths should be same as models
pretrained_list:
# if not null, its lengths should be same as models
freeze_params_list:
- True
- False
infer_model_name: "Student"
models:
- Teacher:
name: ResNet34
class_num: *class_num
pretrained: True
return_patterns: &t_stages ["blocks[2]", "blocks[6]", "blocks[12]", "blocks[15]"]
- Student:
name: ResNet18
class_num: *class_num
pretrained: False
return_patterns: &s_stages ["blocks[1]", "blocks[3]", "blocks[5]", "blocks[7]"]
# loss function config for traing/eval process
Loss:
Train:
- DistillationGTCELoss:
weight: 1.0
model_names: ["Student"]
- DistillationPairLoss:
weight: 1.0
model_name_pairs: [["Student", "Teacher"]] # calculate mgdloss for Student and Teacher
name: "loss_mgd"
base_loss_name: MGDLoss # MGD loss, the following are parameters of 'MGD loss'
s_keys: ["blocks[7]"] # feature map used to calculate MGD loss in student model
t_keys: ["blocks[15]"] # feature map used to calculate MGD loss in teacher model
student_channels: 512 # channel num for stduent feature map
teacher_channels: 512 # channel num for teacher feature map
Eval:
- CELoss:
weight: 1.0Paper:
Rethinking Soft Labels For Knowledge Distillation: A Bias-variance Tradeoff Perspective
Helong Zhou, Liangchen Song, Jiajie Chen, Ye Zhou, Guoli Wang, Junsong Yuan, Qian Zhang
ICLR, 2021
Weighted Soft Labels (WSL) loss function assigns weights to the KD Loss of each sample according to the CE Loss ratio of the teacher model and the student model with respect to the ground-truth labels. If the student model predicts a certain sample better than the teacher model, a smaller weight will be assigned to the sample. The method is simple and effective. It enables the weight of each sample to be adaptively adjusted, thereby improving the distillation accuracy.
Performance on ImageNet1k is shown below.
| Strategy | Backbone | Config | Top-1 acc | Download Link |
|---|---|---|---|---|
| baseline | ResNet18 | ResNet18.yaml | 70.8% | - |
| WSL | ResNet18 | resnet34_distill_resnet18_wsl.yaml | 72.23%(+1.43%) | - |
The WSL configuration is shown below. In the Arch field, you need to define both the student model and the teacher model. The teacher model has fixed parameters, and the pretrained parameters are loaded. In the Loss field, you need to define DistillationWSLLoss (WSL loss between student and teacher) and DistillationGTCELoss (CE loss with ground truth labels) as the training loss.
# model architecture
Arch:
name: "DistillationModel"
# if not null, its lengths should be same as models
pretrained_list:
# if not null, its lengths should be same as models
freeze_params_list:
- True
- False
models:
- Teacher:
name: ResNet34
pretrained: True
- Student:
name: ResNet18
pretrained: False
infer_model_name: "Student"
# loss function config for traing/eval process
Loss:
Train:
- DistillationGTCELoss:
weight: 1.0
model_names: ["Student"]
- DistillationWSLLoss:
weight: 2.5
model_name_pairs: [["Student", "Teacher"]]
temperature: 2
Eval:
- CELoss:
weight: 1.0- Installation: Please refer to Paddle Installation Tutorial and PaddleClas Installation Tutorial to configure the running environment.
Please prepare the ImageNet-1k dataset on ImageNet website.
Enter PaddleClas directory.
cd path_to_PaddleClas
Enter dataset/ directory, name the downloaded data ILSVRC2012 and store it here. The ILSVRC2012 directory contains the following data:
├── train
│ ├── n01440764
│ │ ├── n01440764_10026.JPEG
│ │ ├── n01440764_10027.JPEG
├── train_list.txt
...
├── val
│ ├── ILSVRC2012_val_00000001.JPEG
│ ├── ILSVRC2012_val_00000002.JPEG
├── val_list.txt
where train/ and val/ are training set and validation set, respectively. train_list.txt and val_list.txt are label files for training set and validation set, respectively.
If unlabeled data similar to the training set scenario is included, they can also be organized in the same way as the training set labels. Place the file in the same directory as the currently labeled dataset, and mark its label value as 0. Suppose the organized tag file is named train_list_unlabel.txt, you can use the following command to generate a label file for SSLD training.
cat train_list.txt train_list_unlabel.txt > train_list_all.txtNote:
- For more information about the format of
train_list.txtandval_list.txt, you may refer to Format Description of PaddleClas Classification Dataset .
In this section, the process of model training, evaluation and prediction of knowledge distillation algorithm will be introduced using the SSLD knowledge distillation algorithm as an example. The configuration file is PPLCNet_x2_5_ssld.yaml. You can use the following command to complete the model training.
export CUDA_VISIBLE_DEVICES=0,1,2,3
python3 -m paddle.distributed.launch \
--gpus="0,1,2,3" \
tools/train.py \
-c ppcls/configs/ImageNet/Distillation/PPLCNet_x2_5_ssld.yamlAfter training the model, the following command can be used to evaluate the performance of the model:
python3 tools/eval.py \
-c ppcls/configs/ImageNet/Distillation/PPLCNet_x2_5_ssld.yaml \
-o Global.pretrained_model=output/DistillationModel/best_modelwhere -o Global.pretrained_model="output/DistillationModel/best_model" specifies the path of the current optimal weights. If you need to specify other weights, you can simply replace the path.
After training is completed, the trained model can be loaded for prediction. A complete example is provided in tools/infer.py. You can use the model for prediction by executing the following command:
python3 tools/infer.py \
-c ppcls/configs/ImageNet/Distillation/PPLCNet_x2_5_ssld.yaml \
-o Global.pretrained_model=output/DistillationModel/best_modelThe outputs are as follows:
[{'class_ids': [8, 7, 86, 82, 21], 'scores': [0.87908, 0.12091, 0.0, 0.0, 0.0], 'file_name': 'docs/images/inference_deployment/whl_demo.jpg', 'label_names': ['hen', 'cock', 'partridge', 'ruffed grouse, partridge, Bonasa umbellus', 'kite']}]
Note:
-
Here
-o Global.pretrained_model="output/ResNet50/best_model"specifies the path of the current optimal weights. If you need to specify other weights, you can simply replace the path. -
Image
docs/images/inference_deployment/whl_demo.jpgis predicted by default. You can also predict other images by adding a field-o Infer.infer_imgs=xxx.
PaddleInference is a native inference library for PaddlePaddle, which can be used on servers and clouds to provide high-performance inference. PaddleInference can use MKLDNN, CUDNN, and TensorRT to accelerate model inference, thereby achieving better performance compared with inference based directly on the trained model. For more information about PaddleInference, please refer to Paddle Inference Tutorial.
The model needs to be exported before inference. For models obtained from knowledge distillation, -o Global.infer_model_name=Student should be specified when exporting to indicate the model to be exported is the student model. The complete command is shown below.
python3 tools/export_model.py \
-c ppcls/configs/ImageNet/Distillation/PPLCNet_x2_5_ssld.yaml \
-o Global.pretrained_model=./output/DistillationModel/best_model \
-o Arch.infer_model_name=Student3 files will be generated in inference directory: inference.pdiparams, inference.pdiparams.info and inference.pdmodel.
For more information about model inference, please refer to: Python Inference.
[1] Hinton G, Vinyals O, Dean J. Distilling the knowledge in a neural network[J]. arXiv preprint arXiv:1503.02531, 2015.
[2] Bagherinezhad H, Horton M, Rastegari M, et al. Label refinery: Improving imagenet classification through label progression[J]. arXiv preprint arXiv:1805.02641, 2018.
[3] Yalniz I Z, Jégou H, Chen K, et al. Billion-scale semi-supervised learning for image classification[J]. arXiv preprint arXiv:1905.00546, 2019.
[4] Cubuk E D, Zoph B, Mane D, et al. Autoaugment: Learning augmentation strategies from data[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. 2019: 113-123.
[5] Touvron H, Vedaldi A, Douze M, et al. Fixing the train-test resolution discrepancy[C]//Advances in Neural Information Processing Systems. 2019: 8250-8260.
[6] Cui C, Guo R, Du Y, et al. Beyond Self-Supervision: A Simple Yet Effective Network Distillation Alternative to Improve Backbones[J]. arXiv preprint arXiv:2103.05959, 2021.
[7] Zhang Y, Xiang T, Hospedales T M, et al. Deep mutual learning[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2018: 4320-4328.
[8] Heo B, Kim J, Yun S, et al. A comprehensive overhaul of feature distillation[C]//Proceedings of the IEEE/CVF International Conference on Computer Vision. 2019: 1921-1930.
[9] Du Y, Li C, Guo R, et al. PP-OCRv2: Bag of Tricks for Ultra Lightweight OCR System[J]. arXiv preprint arXiv:2109.03144, 2021.
[10] Park W, Kim D, Lu Y, et al. Relational knowledge distillation[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2019: 3967-3976.
[11] Zhao B, Cui Q, Song R, et al. Decoupled Knowledge Distillation[J]. arXiv preprint arXiv:2203.08679, 2022.
[12] Ji M, Heo B, Park S. Show, attend and distill: Knowledge distillation via attention-based feature matching[C]//Proceedings of the AAAI Conference on Artificial Intelligence. 2021, 35(9): 7945-7952.
[13] Huang T, You S, Wang F, et al. Knowledge Distillation from A Stronger Teacher[J]. arXiv preprint arXiv:2205.10536, 2022.