For the latest update, please go to DeepCriticalLearning

Deep Robustness: ProSelfLC-CVPR 2021 + Example Weighting (DM+IMAE)

ProSelfLC: Progressive Self Label Correction for Training Robust Deep Neural Networks

• Paper link: https://arxiv.org/abs/2005.03788

Mean Absolute Error Does Not Treat Examples Equally and Gradient Magnitude’s Variance Matters

Derivative Manipulation: Example Weighting via Emphasis Density Funtion in the context of DL

• Novelty: moving from loss design to derivative design

For any specific discussion or potential future collaboration, please feel free to contact me.

See Citation Details

Please kindly cite the following papers if you find this repo useful.

@inproceddings{wang2021proselflc,
  title={ {ProSelfLC}: Progressive Self Label Correction
  for Training Robust Deep Neural Networks},
  author={Wang, Xinshao and Hua, Yang and Kodirov, Elyor and Clifton, David A and Robertson, Neil M},
  booktitle={CVPR},
  year={2021}
}
@phdthesis{wang2020example,
  title={Example weighting for deep representation learning},
  author={Wang, Xinshao},
  year={2020},
  school={Queen's University Belfast}
}
@article{wang2019derivative,
  title={Derivative Manipulation for General Example Weighting},
  author={Wang, Xinshao and Kodirov, Elyor and Hua, Yang and Robertson, Neil},
  journal={arXiv preprint arXiv:1905.11233},
  year={2019}
}
@article{wang2019imae,
  title={{IMAE} for Noise-Robust Learning: Mean Absolute Error Does Not Treat Examples Equally and Gradient Magnitude’s Variance Matters},
  author={Wang, Xinshao and Hua, Yang and Kodirov, Elyor and Robertson, Neil M},
  journal={arXiv preprint arXiv:1903.12141},
  year={2019}
}

PyTorch Implementation for ProSelfLC, Derivative Manipulation, Improved MAE

• Easy to install
• Easy to use
• Easy to extend: new losses, new networks, new dataset and loaders
• Easy to run experiments and sink results
• Easy to put sinked results into your technical reports and academic papers.

Install

See Install Guidelines

Set the Pipenv From Scratch

• sudo apt update && sudo apt upgrade
• sudo apt install python3.8
• curl https://bootstrap.pypa.io/get-pip.py -o get-pip.py
• python3.8 get-pip.py
• vim ~/.bashrc -> add export PATH="/home/ubuntu/.local/bin:$PATH" -> source ~/.bashrc
• pip3 install pipenv

Build env for this repo using pipenv

• git clone [email protected]:XinshaoAmosWang/ProSelfLC-CVPR2021.git
• cd ProSelfLC-CVPR2021
• pipenv install -e .

How to use

Run experiments

• cd ProSelfLC-CVPR2021
• pipenv shell
• python tests/convnets/test_trainer_cnn_vision_cifar100_cce_convets.py
• python tests/convnets/test_trainer_cnn_vision_cifar100_lscplc_covnets.py
• python tests/convnets/test_trainer_cnn_vision_cifar100_proselflc_covnets.py
• python tests/derivativemanipulation_imae/test_dm_imae.py

Visualize results

• The results are well sinked and organised, e.g., proselflc_experiments/cifar100_symmetric_noise_rate_0.2/shufflenetv2

How to extend this repo

• Add dataset and dataloader: see examples in src/proselflc/slices/datain
• Add losses: see examples in src/proselflc/slices/losses
• Add networks: see examples in src/proselflc/slices/networks
• Add optimisers: see examples in src/proselflc/optim
• Extend the slicegetter: src/proselflc/slicegetter
• Write run scripts: see examples in tests/convnets/

Examples of sinked experimental configs and resutls

See Sinked Results

• The results are well sinked and organised, e.g., proselflc_experiments/cifar100_symmetric_noise_rate_0.2/shufflenetv2

• 

• 

• accuracy_loss.xlsx

epoch	clean_test	noisy_subset	clean_subset	cleaned_noisy_subset
4	0.1801	0.0091	0.169	0.1617
8	0.3329	0.0089	0.32655	0.3065
12	0.3848	0.0093	0.380025	0.3451
16	0.391	0.0091	0.391875	0.3561
20	0.4225	0.009	0.4119	0.3693
24	0.4214	0.0077	0.416975	0.3728
28	0.4222	0.0108	0.43095	0.3884
32	0.4709	0.0097	0.47	0.4254
36	0.4155	0.0097	0.42955	0.3886
40	0.449	0.0083	0.463125	0.4165
44	0.448	0.0074	0.441125	0.4006
48	0.3856	0.0078	0.383025	0.3496
52	0.4672	0.0083	0.479475	0.4373
56	0.4428	0.0081	0.437075	0.3891
60	0.4164	0.0095	0.422675	0.3815
64	0.4635	0.0079	0.483225	0.4386
68	0.4506	0.0085	0.4654	0.4145
72	0.4428	0.0081	0.459825	0.4105
76	0.4553	0.0086	0.4579	0.4151
80	0.6108	0.0104	0.670775	0.5751
84	0.5989	0.0107	0.67395	0.5636
88	0.6026	0.0098	0.6832	0.5703
92	0.5949	0.013	0.679925	0.5608
96	0.5976	0.0122	0.67985	0.5487
100	0.5838	0.0123	0.67	0.5426
104	0.592	0.0113	0.67465	0.5544
108	0.603	0.0117	0.6723	0.5552
112	0.58	0.0117	0.662175	0.5394
116	0.5767	0.0126	0.660825	0.5421
120	0.5829	0.0121	0.654925	0.5382
124	0.5828	0.0127	0.655875	0.5426
128	0.5825	0.013	0.652575	0.5405
132	0.5641	0.0111	0.625275	0.5303
136	0.5779	0.0112	0.635275	0.5355
140	0.6462	0.0139	0.76175	0.6095
144	0.6464	0.0159	0.76965	0.6175
148	0.6412	0.0169	0.773475	0.616
152	0.6458	0.0161	0.775025	0.6127
156	0.6353	0.019	0.768825	0.6036
160	0.6385	0.0168	0.768125	0.6115
164	0.6338	0.0181	0.763825	0.6122
168	0.6316	0.0164	0.75755	0.6011
172	0.6225	0.0171	0.747675	0.5942
176	0.6312	0.0153	0.749425	0.5989
180	0.6552	0.021	0.7966	0.6264
184	0.653	0.022	0.8036	0.6292
188	0.6544	0.0212	0.807125	0.6256
192	0.6545	0.0209	0.80905	0.6286
196	0.6531	0.0222	0.811075	0.6276
200	0.6572	0.0229	0.8102	0.627

• params.csv

data_name	num_classes	device	num_workers	batch_size	counter	lr	total_epochs	milestones	gamma	loss_name	symmetric_noise_rate	network_name	warmup_epochs	exp_base	transit_time_ratio	summary_writer_dir	train	total_iterations	momentum	weight_decay
cifar100	100	gpu	4	128	iteration	0.1	200	60	0.2	proselflc	0.2	shufflenetv2	16	6	0.3	/home/xinshao/tpami_proselflc_experiments/cifar100_symmetric_noise_rate_0.2/shufflenetv2/050_proselflc_warm16_b6_transit0.3_20210904-172732	True	78200	0.9	0.0005
cifar100	100	gpu	4	128	iteration	0.1	200	120	0.2	proselflc	0.2	shufflenetv2	16	6	0.3	/home/xinshao/tpami_proselflc_experiments/cifar100_symmetric_noise_rate_0.2/shufflenetv2/050_proselflc_warm16_b6_transit0.3_20210904-172732	True	78200	0.9	0.0005
cifar100	100	gpu	4	128	iteration	0.1	200	160	0.2	proselflc	0.2	shufflenetv2	16	6	0.3	/home/xinshao/tpami_proselflc_experiments/cifar100_symmetric_noise_rate_0.2/shufflenetv2/050_proselflc_warm16_b6_transit0.3_20210904-172732	True	78200	0.9	0.0005

Link to Slide, Poster, Final version

Link to reviewers' comments

List of Content

0. Storyline
1. Open ML Research Questions
2. Noticeable Findings
3. Literature Review
4. In Self LC, a core question is not well answered
5. Underlying Principle of ProSelfLC
6. Mathematical Details of ProSelfLC
7. Design Reasons of ProSelfLC
8. Related Interesting Work