├── .gitignore ├── README.md ├── data └── icbhi_dataset │ ├── filename_differences.txt │ ├── filename_format.txt │ ├── metadata.txt │ ├── official_split.txt │ ├── patient_diagnosis.txt │ └── patient_list_foldwise.txt ├── main.py ├── models └── clap.py ├── requirements.txt ├── scripts ├── eval_bts.sh ├── icbhi_audio-clap_ce.sh └── icbhi_bts_meta_all.sh └── util ├── __init__.py ├── icbhi_dataset.py ├── icbhi_util.py ├── meta_description.py └── misc.py /.gitignore: -------------------------------------------------------------------------------- 1 | # Byte-compiled / optimized / DLL files 2 | __pycache__/ 3 | *.py[cod] 4 | *$py.class 5 | 6 | # C extensions 7 | *.so 8 | 9 | # Distribution / packaging 10 | .Python 11 | build/ 12 | develop-eggs/ 13 | dist/ 14 | downloads/ 15 | eggs/ 16 | .eggs/ 17 | lib/ 18 | lib64/ 19 | parts/ 20 | sdist/ 21 | var/ 22 | wheels/ 23 | share/python-wheels/ 24 | *.egg-info/ 25 | .installed.cfg 26 | *.egg 27 | MANIFEST 28 | 29 | # PyInstaller 30 | # Usually these files are written by a python script from a template 31 | # before PyInstaller builds the exe, so as to inject date/other infos into it. 32 | *.manifest 33 | *.spec 34 | 35 | 36 | # Installer logs 37 | pip-log.txt 38 | pip-delete-this-directory.txt 39 | 40 | # Unit test / coverage reports 41 | htmlcov/ 42 | .tox/ 43 | .nox/ 44 | .coverage 45 | .coverage.* 46 | .cache 47 | nosetests.xml 48 | coverage.xml 49 | *.cover 50 | *.py,cover 51 | .hypothesis/ 52 | .pytest_cache/ 53 | cover/ 54 | 55 | # Translations 56 | *.mo 57 | *.pot 58 | 59 | # Django stuff: 60 | *.log 61 | local_settings.py 62 | db.sqlite3 63 | db.sqlite3-journal 64 | 65 | # Flask stuff: 66 | instance/ 67 | .webassets-cache 68 | 69 | # Scrapy stuff: 70 | .scrapy 71 | 72 | # Sphinx documentation 73 | docs/_build/ 74 | 75 | # PyBuilder 76 | .pybuilder/ 77 | target/ 78 | 79 | # Jupyter Notebook 80 | .ipynb_checkpoints 81 | 82 | # IPython 83 | profile_default/ 84 | ipython_config.py 85 | 86 | # pyenv 87 | # For a library or package, you might want to ignore these files since the code is 88 | # intended to run in multiple environments; otherwise, check them in: 89 | # .python-version 90 | 91 | # pipenv 92 | # According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control. 93 | # However, in case of collaboration, if having platform-specific dependencies or dependencies 94 | # having no cross-platform support, pipenv may install dependencies that don't work, or not 95 | # install all needed dependencies. 96 | #Pipfile.lock 97 | 98 | # poetry 99 | # Similar to Pipfile.lock, it is generally recommended to include poetry.lock in version control. 100 | # This is especially recommended for binary packages to ensure reproducibility, and is more 101 | # commonly ignored for libraries. 102 | # https://python-poetry.org/docs/basic-usage/#commit-your-poetrylock-file-to-version-control 103 | #poetry.lock 104 | 105 | # pdm 106 | # Similar to Pipfile.lock, it is generally recommended to include pdm.lock in version control. 107 | #pdm.lock 108 | # pdm stores project-wide configurations in .pdm.toml, but it is recommended to not include it 109 | # in version control. 110 | # https://pdm.fming.dev/#use-with-ide 111 | .pdm.toml 112 | 113 | # PEP 582; used by e.g. github.com/David-OConnor/pyflow and github.com/pdm-project/pdm 114 | __pypackages__/ 115 | 116 | # Celery stuff 117 | celerybeat-schedule 118 | celerybeat.pid 119 | 120 | # SageMath parsed files 121 | *.sage.py 122 | 123 | # Environments 124 | .env 125 | .venv 126 | env/ 127 | venv/ 128 | ENV/ 129 | env.bak/ 130 | venv.bak/ 131 | 132 | # Spyder project settings 133 | .spyderproject 134 | .spyproject 135 | 136 | # Rope project settings 137 | .ropeproject 138 | 139 | # mkdocs documentation 140 | /site 141 | 142 | # mypy 143 | .mypy_cache/ 144 | .dmypy.json 145 | dmypy.json 146 | 147 | # Pyre type checker 148 | .pyre/ 149 | 150 | # pytype static type analyzer 151 | .pytype/ 152 | 153 | # Cython debug symbols 154 | cython_debug/ 155 | 156 | # PyCharm 157 | # JetBrains specific template is maintained in a separate JetBrains.gitignore that can 158 | # be found at https://github.com/github/gitignore/blob/main/Global/JetBrains.gitignore 159 | # and can be added to the global gitignore or merged into this file. For a more nuclear 160 | # option (not recommended) you can uncomment the following to ignore the entire idea folder. 161 | #.idea/ 162 | 163 | # macbook junk 164 | .DS_Store 165 | 166 | # debug outputs 167 | .outputs/ 168 | .temp/ 169 | .local/ 170 | 171 | # Local folders 172 | .vscode/ -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | # BTS: Bridging Text and Sound Modalities for Metadata-Aided Respiratory Sound Classification (INTERSPEECH 2024) 2 | [arXiv](https://arxiv.org/abs/2406.06786) | [Conference](https://www.isca-archive.org/interspeech_2024/kim24f_interspeech.html) | [BibTeX](#bibtex) 3 | 4 |

5 | image 6 |

7 | 8 | 9 | Official Implementation of **BTS: Bridging Text and Sound Modalities for Metadata-Aided Respiratory Sound Classification.**
10 | 11 | **See you in INTERSPEECH 2024!** 12 | 13 | ## Updates 14 | The pretrained BTS checkpoint is available at [Google Drive](https://drive.google.com/file/d/1RziWDkqRTUictS_9zYWbJWj4i6C79RIy/view?usp=sharing). 15 | 16 | The model's performance on Specificity, Sensitivity, and ICBHI Score is **[82.58, 45.11, 63.85**]. 17 | 18 | ## Prerequisites 19 | Please check environments and requirements before you start. If required, we recommend you to either upgrade versions or install them for smooth running. 20 | 21 | ![Ubuntu](https://img.shields.io/badge/Ubuntu-E95420?style=for-the-badge&logo=ubuntu&logoColor=white) 22 | ![Python](https://img.shields.io/badge/python-3670A0?style=for-the-badge&logo=python&logoColor=ffdd54) 23 | ![PyTorch](https://img.shields.io/badge/PyTorch-%23EE4C2C.svg?style=for-the-badge&logo=PyTorch&logoColor=white) 24 | 25 | ### Environments 26 | `Ubuntu xx.xx` 27 | `Python 3.8.xx` 28 | 29 | ## Environmental set-up 30 | 31 | Install the necessary packages with: 32 | 33 | run `requirements.txt` 34 | 35 | ``` 36 | pip install torch torchvision torchaudio 37 | pip install -r requirements.txt 38 | ``` 39 | 40 | For the reproducibility, we used torch=2.0.1+cu117 and torchaudio=2.0.1+cu117, so we highly recommend install as follow: 41 | 42 | ``` 43 | pip install torch==2.0.1 torchvision==0.15.2 torchaudio==2.0.2 44 | ``` 45 | 46 | ## Datasets 47 | Download the ICBHI files and unzip it. 48 | All details is described in the [paper w/ code](https://paperswithcode.com/dataset/icbhi-respiratory-sound-database) 49 | 50 | ``` 51 | wget https://bhichallenge.med.auth.gr/sites/default/files/ICBHI_final_database/ICBHI_final_database.zip 52 | or 53 | wget --no-check-certificate https://bhichallenge.med.auth.gr/sites/default/files/ICBHI_final_database/ICBHI_final_database.zip 54 | ``` 55 | 56 | All `*.wav` and `*.txt` should be saved in data/icbhi_dataset/audio_test_data. (i.e., mkdir `audio_test_data` into `data/icbhi_dataset/` and move `*.wav` and `*.txt` into `data/icbhi_dataset/audio_test_data/`) 57 | 58 | Note that ICBHI dataset consists of a total of 6,898 respiratory cycles, of which 1,864 contain crackles, 886 contain wheezes, and 506 contain both crackles and wheezes, in 920 annotated audio samples from 126 subjects. 59 | 60 | ## Run 61 | 62 | ### Audio-CLAP for Respiratory Sound Classification 63 | ``` 64 | $ ./scripts/icbhi_audio-clap_ce.sh 65 | ``` 66 | 67 | ### BTS for Respiratory Sound Classification 68 | ``` 69 | $ ./scripts/icbhi_bts_meta_all.sh 70 | ``` 71 | 72 | ### Evaluation with BTS for Respiratory Sound Classification 73 | ``` 74 | $ ./scripts/eval_bts.sh 75 | ``` 76 | Note that change `--pretrained_ckpt` with your directory. (e.g. `--pretrained_ckpt /home2/jw/workspace/crisp/save/icbhi_laion/clap-htsat-unfused_ce_bs8_lr5e-5_ep50_seed1_check2/best.pth`) 77 | 78 | 79 | We will provide pretrained checkpoint into the camera-ready version 80 | 81 | ## ICBHI Data 82 | 83 | The database consists of a total of 5.5 hours of recordings containing 6898 respiratory cycles, of which 1864 contain crackles, 886 contain wheezes, and 506 contain both crackles and wheezes, in 920 annotated audio samples from 126 subjects. 84 | 85 | The downloaded data looks like [[kaggle](https://www.kaggle.com/datasets/vbookshelf/respiratory-sound-database), [paper w/ code](https://paperswithcode.com/dataset/icbhi-respiratory-sound-database)]: 86 | 87 | 
 88 | data/icbhi_dataset
 89 | ├── metadata.txt
 90 | │    ├── Patient number
 91 | │    ├── Age
 92 | │    ├── Sex
 93 | │    ├── Adult BMI (kg/m2)
 94 | │    ├── Adult Weight (kg)
 95 | │    └── Child Height (cm)
 96 | │
 97 | ├── official_split.txt
 98 | │    ├── Patient number_Recording index_Chest location_Acqiosotopm mode_Recording equipment
 99 | │    |    ├── Chest location
100 | │    |    |    ├── Trachea (Tc),Anterior left (Al),Anterior right (Ar),Posterior left (Pl)
101 | │    |    |    └── Posterior right (Pr),Lateral left (Ll),Lateral right (Lr)
102 | │    |    |
103 | │    |    ├── Acquisition mode
104 | │    |    |    └── sequential/single channel (sc), simultaneous/multichannel (mc)
105 | │    |    |
106 | │    |    └── Recording equipment 
107 | │    |         ├── AKG C417L Microphone (AKGC417L), 
108 | │    |         ├── 3M Littmann Classic II SE Stethoscope (LittC2SE), 
109 | │    |         ├── 3M Litmmann 3200 Electronic Stethoscope (Litt3200), 
110 | │    |         └── WelchAllyn Meditron Master Elite Electronic Stethoscope (Meditron)
111 | │    |    
112 | │    └── Train/Test   
113 | │
114 | ├── patient_diagnosis.txt
115 | │    ├── Patient number
116 | │    └── Diagnosis
117 | │         ├── COPD: Chronic Obstructive Pulmonary Disease
118 | │         ├── LRTI: Lower Respiratory Tract Infection
119 | │         └── URTI: Upper Respiratory Tract Infection
120 | │
121 | └── patient_list_foldwise.txt
122 |
123 | 124 | ## Result 125 | The proposed BTS achieves a 63.54% Score, which is the new state-of-the-art performance in ICBHI score. 126 |

127 | image 128 |

129 | 130 | 131 | 132 | ## BibTeX 133 | If you find this repo useful for your research, please consider citing our paper: 134 | 135 | ``` 136 | @inproceedings{kim24f_interspeech, 137 | title = {BTS: Bridging Text and Sound Modalities for Metadata-Aided Respiratory Sound Classification}, 138 | author = {June-Woo Kim and Miika Toikkanen and Yera Choi and Seoung-Eun Moon and Ho-Young Jung}, 139 | year = {2024}, 140 | booktitle = {Interspeech 2024}, 141 | pages = {1690--1694}, 142 | doi = {10.21437/Interspeech.2024-492}, 143 | issn = {2958-1796}, 144 | } 145 | ``` 146 | -------------------------------------------------------------------------------- /data/icbhi_dataset/filename_differences.txt: -------------------------------------------------------------------------------- 1 | '101_1b1_Al_sc_AKGC417L' 2 | '101_1b1_Pr_sc_AKGC417L' 3 | '102_1b1_Ar_sc_AKGC417L' 4 | '105_1b1_Tc_sc_LittC2SE' 5 | '108_1b1_Al_sc_LittC2SE' 6 | '111_1b2_Tc_sc_LittC2SE' 7 | '111_1b3_Tc_sc_LittC2SE' 8 | '115_1b1_Ar_sc_LittC2SE' 9 | '116_1b2_Pl_sc_LittC2SE' 10 | '116_1b2_Tc_sc_LittC2SE' 11 | '119_1b1_Ar_sc_AKGC417L' 12 | '121_1b1_Tc_sc_LittC2SE' 13 | '121_1p1_Tc_sc_LittC2SE' 14 | '123_1b1_Al_sc_AKGC417L' 15 | '125_1b1_Tc_sc_LittC2SE' 16 | '126_1b1_Al_sc_AKGC417L' 17 | '127_1b1_Ar_sc_LittC2SE' 18 | '129_1b1_Ar_sc_LittC2SE' 19 | '131_1b1_Al_sc_LittC2SE' 20 | '136_1b1_Ar_sc_AKGC417L' 21 | '137_1b1_Ar_sc_LittC2SE' 22 | '137_1b1_Ll_sc_LittC2SE' 23 | '143_1b1_Al_sc_AKGC417L' 24 | '144_1b1_Al_sc_AKGC417L' 25 | '144_1b1_Tc_sc_AKGC417L' 26 | '148_1b1_Al_sc_LittC2SE' 27 | '149_1b1_Al_sc_LittC2SE' 28 | '149_1b1_Lr_sc_LittC2SE' 29 | '149_1b1_Pl_sc_LittC2SE' 30 | '150_1b2_Al_sc_AKGC417L' 31 | '152_1b1_Al_sc_LittC2SE' 32 | '153_1b1_Al_sc_LittC2SE' 33 | '159_1b1_Al_sc_AKGC417L' 34 | '159_1b1_Ar_sc_AKGC417L' 35 | '159_1b1_Ll_sc_AKGC417L' 36 | '159_1b1_Pr_sc_AKGC417L' 37 | '161_1b1_Al_sc_LittC2SE' 38 | '161_1b1_Pl_sc_LittC2SE' 39 | '164_1b1_Ll_sc_LittC2SE' 40 | '165_1b1_Ar_sc_AKGC417L' 41 | '165_1b1_Pl_sc_AKGC417L' 42 | '165_1b1_Pr_sc_AKGC417L' 43 | '167_1b1_Al_sc_LittC2SE' 44 | '167_1b1_Pr_sc_LittC2SE' 45 | '168_1b1_Al_sc_LittC2SE' 46 | '169_1b1_Lr_sc_AKGC417L' 47 | '169_1b2_Ll_sc_AKGC417L' 48 | '171_1b1_Al_sc_AKGC417L' 49 | '173_1b1_Al_sc_AKGC417L' 50 | '179_1b1_Al_sc_LittC2SE' 51 | '179_1b1_Tc_sc_LittC2SE' 52 | '182_1b1_Tc_sc_LittC2SE' 53 | '183_1b1_Pl_sc_AKGC417L' 54 | '183_1b1_Tc_sc_AKGC417L' 55 | '184_1b1_Ar_sc_LittC2SE' 56 | '187_1b1_Ll_sc_AKGC417L' 57 | '188_1b1_Al_sc_LittC2SE' 58 | '188_1b1_Ar_sc_LittC2SE' 59 | '188_1b1_Pl_sc_LittC2SE' 60 | '188_1b1_Tc_sc_LittC2SE' 61 | '190_1b1_Tc_sc_AKGC417L' 62 | '194_1b1_Lr_sc_AKGC417L' 63 | '194_1b1_Pr_sc_AKGC417L' 64 | '196_1b1_Pr_sc_LittC2SE' 65 | '197_1b1_Al_sc_AKGC417L' 66 | '197_1b1_Tc_sc_AKGC417L' 67 | '201_1b1_Al_sc_LittC2SE' 68 | '201_1b1_Ar_sc_LittC2SE' 69 | '201_1b2_Al_sc_LittC2SE' 70 | '201_1b2_Ar_sc_LittC2SE' 71 | '201_1b3_Al_sc_LittC2SE' 72 | '201_1b3_Ar_sc_LittC2SE' 73 | '202_1b1_Ar_sc_AKGC417L' 74 | '206_1b1_Ar_sc_LittC2SE' 75 | '206_1b1_Lr_sc_LittC2SE' 76 | '206_1b1_Pl_sc_LittC2SE' 77 | '208_1b1_Ll_sc_LittC2SE' 78 | '209_1b1_Tc_sc_LittC2SE' 79 | '210_1b1_Al_sc_LittC2SE' 80 | '210_1b1_Ar_sc_LittC2SE' 81 | '214_1b1_Ar_sc_AKGC417L' 82 | '215_1b2_Ar_sc_LittC2SE' 83 | '215_1b3_Tc_sc_LittC2SE' 84 | '216_1b1_Al_sc_AKGC417L' 85 | '216_1b1_Pl_sc_AKGC417L' 86 | '217_1b1_Tc_sc_LittC2SE' 87 | '224_1b1_Tc_sc_AKGC417L' 88 | '224_1b2_Al_sc_AKGC417L' 89 | '225_1b1_Pl_sc_AKGC417L' 90 | '226_1b1_Al_sc_LittC2SE' 91 | '226_1b1_Ll_sc_LittC2SE' 92 | -------------------------------------------------------------------------------- /data/icbhi_dataset/filename_format.txt: -------------------------------------------------------------------------------- 1 | Elements contained in the filenames: 2 | 3 | Patient number (101,102,...,226) 4 | Recording index 5 | Chest location (Trachea (Tc), {Anterior (A), Posterior (P), Lateral (L)}{left (l), right (r)}) 6 | Acquisition mode (sequential/single channel (sc), simultaneous/multichannel (mc)) 7 | Recording equipment (AKG C417L Microphone, 3M Littmann Classic II SE Stethoscope, 3M Litmmann 3200 Electronic Stethoscope, WelchAllyn Meditron Master Elite Electronic Stethoscope) -------------------------------------------------------------------------------- /data/icbhi_dataset/metadata.txt: -------------------------------------------------------------------------------- 1 | 101 3.0 F NA 19.0 99.0 Pr 2 | 102 0.75 F NA 9.8 73.0 Ar 3 | 103 70.0 F 33.0 NA NA Ar 4 | 104 70.0 F 28.47 NA NA Al 5 | 105 7.0 F NA 32.0 135.0 Tc 6 | 106 73.0 F 21.0 NA NA Pl 7 | 107 75.0 F 33.7 NA NA Ll 8 | 108 3.0 M NA NA NA Al 9 | 109 84.0 F 33.53 NA NA Lr 10 | 110 75.0 M 25.21 NA NA Pr 11 | 111 63.0 M 28.4 NA NA Tc 12 | 112 60.0 M 22.86 NA NA Pr 13 | 113 58.0 M 28.41 NA NA Lr 14 | 114 77.0 M 23.12 NA NA Pr 15 | 115 0.58 M NA 7.14 64.0 Ar 16 | 116 56.0 M 28.58 NA NA Pl 17 | 117 68.0 M 24.4 NA NA Tc 18 | 118 81.0 M 36.76 NA NA Pr 19 | 119 2.0 F NA 15.2 94.0 Ar 20 | 120 78.0 M 35.14 NA NA Ar 21 | 121 13.0 F NA 65.0 170.0 Tc 22 | 122 66.0 M 33.0 NA NA Ar 23 | 123 5.0 M NA 25.0 125.0 Al 24 | 124 65.0 M 29.07 NA NA Al 25 | 125 14.0 M NA 62.0 170.0 Tc 26 | 126 1.0 F NA 10.18 80.0 Al 27 | 127 2.0 M NA 12.6 98.0 Ar 28 | 128 65.0 F 24.3 NA NA Tc 29 | 129 6.0 M NA 23.0 119.0 Ar 30 | 130 85.0 F 17.1 NA NA Pl 31 | 131 3.0 M NA 14.0 97.0 Al 32 | 132 71.0 M 34.0 NA NA Lr 33 | 133 68.0 M 27.4 NA NA Pl 34 | 134 61.0 M 32.0 NA NA Al 35 | 135 70.0 M 21.0 NA NA Ar 36 | 136 5.0 M NA 16.2 110.0 Ar 37 | 137 4.0 M NA 18.0 104.0 Ar 38 | 138 56.0 F 21.6 NA NA Ll 39 | 139 61.0 M 28.68 NA NA Lr 40 | 140 79.0 F 23.0 NA NA Tc 41 | 141 66.0 M 22.4 NA NA Pr 42 | 142 78.0 M 26.1 NA NA Pl 43 | 143 0.25 F NA 8.24 68.0 Al 44 | 144 3.0 M NA 16.7 100.0 Al 45 | 145 69.0 M 23.4 NA NA Lr 46 | 146 67.0 M 28.0 NA NA Pl 47 | 147 77.0 M 25.7 NA NA Ar 48 | 148 4.0 M NA 33.0 110.0 Al 49 | 149 0.67 M NA 9.5 70.0 Lr 50 | 150 0.67 F NA 8.12 74.0 Al 51 | 151 75.0 M 28.4 NA NA Ll 52 | 152 16.0 M NA 70.0 183.0 Al 53 | 153 3.0 M NA 16.7 103.0 Al 54 | 154 65.0 M 28.1 NA NA Ll 55 | 155 69.0 M 26.0 NA NA Al 56 | 156 80.0 M 22.9 NA NA Al 57 | 157 62.0 M 53.5 NA NA Pr 58 | 158 63.0 M 16.5 NA NA Lr 59 | 159 0.83 F NA 11.0 80.0 Al 60 | 160 74.0 M 27.4 NA NA Tc 61 | 161 2.0 F NA 12.0 85.0 Pl 62 | 162 67.0 F 24.9 NA NA Ll 63 | 163 62.0 M 28.3 NA NA Pl 64 | 164 1.0 M NA 13.0 NA Ll 65 | 165 2.0 F NA 12.7 97.0 Pr 66 | 166 71.0 M 25.06 NA NA Ar 67 | 167 1.0 F NA 11.5 86.4 Pr 68 | 168 19.0 F 17.35 NA NA Al 69 | 169 50.0 F 28.81 NA NA Ll 70 | 170 79.0 M 22.6 NA NA Pr 71 | 171 9.0 M NA 32.0 133.0 Al 72 | 172 73.0 M 29.3 NA NA Lr 73 | 173 3.0 M NA 17.3 NA Al 74 | 174 68.0 M 26.4 NA NA Al 75 | 175 63.0 M 28.34 NA NA Pl 76 | 176 65.0 M 30.1 NA NA Pl 77 | 177 56.0 M 22.1 NA NA Pr 78 | 178 58.0 M 30.1 NA NA Tc 79 | 179 10.0 F NA 15.0 104.0 Al 80 | 180 93.0 M 29.03 NA NA Lr 81 | 181 65.0 M 26.4 NA NA Ar 82 | 182 11.0 M NA 33.0 136.0 Tc 83 | 183 14.0 F NA NA NA Tc 84 | 184 2.0 F NA 15.0 100.0 Ar 85 | 185 75.0 M 27.7 NA NA Pl 86 | 186 71.0 M 30.0 NA NA Al 87 | 187 0.5 F NA 8.26 71.0 Ll 88 | 188 3.0 M NA 16.0 100.0 Al 89 | 189 75.0 F 26.2 NA NA Lr 90 | 190 3.0 F NA NA NA Tc 91 | 191 74.0 F 36.0 NA NA Pr 92 | 192 69.0 M 28.0 NA NA Ar 93 | 193 77.0 M 26.3 NA NA Al 94 | 194 2.0 M NA 12.8 86.0 Pr 95 | 195 67.0 M 29.41 NA NA Al 96 | 196 21.0 F 25.5 NA NA Pr 97 | 197 16.0 F NA NA NA Al 98 | 198 71.0 M 18.6 NA NA Pr 99 | 199 71.0 M 20.0 NA NA Ll 100 | 200 72.0 F 27.8 NA NA Ar 101 | 201 73.0 F 28.52 NA NA Al 102 | 202 2.0 M NA 11.84 87.0 Ar 103 | 203 57.0 F 24.0 NA NA Pr 104 | 204 66.0 M 29.76 NA NA Ll 105 | 205 45.0 M 20.1 NA NA Al 106 | 206 3.0 M NA 13.0 92.0 Pl 107 | 207 63.0 F 29.6 NA NA Pl 108 | 208 5.0 F NA 24.1 117.0 Ll 109 | 209 14.0 F NA 80.0 183.0 Tc 110 | 210 1.0 F NA 12.96 76.0 Al 111 | 211 70.0 F 31.1 NA NA Ar 112 | 212 83.0 M 23.0 NA NA Tc 113 | 213 58.0 F 24.7 NA NA Al 114 | 214 5.0 M NA 30.0 118.0 Ar 115 | 215 56.0 F 25.35 NA NA Ar 116 | 216 1.0 M NA 10.25 78.0 Pl 117 | 217 12.0 F NA NA NA Tc 118 | 218 75.0 M 26.29 NA NA Al 119 | 219 81.0 M 26.0 NA NA Ar 120 | 220 66.0 M 35.4 NA NA Al 121 | 221 74.0 F 29.0 NA NA Pl 122 | 222 60.0 M NA NA NA Ar 123 | 223 NA NA NA NA NA Ll 124 | 224 10.0 F NA 32.3 143.0 Tc 125 | 225 0.83 M NA 7.8 74.0 Pl 126 | 226 4.0 M NA 16.7 103.0 Ll 127 | -------------------------------------------------------------------------------- /data/icbhi_dataset/official_split.txt: -------------------------------------------------------------------------------- 1 | 101_1b1_Al_sc_Meditron test 2 | 101_1b1_Pr_sc_Meditron test 3 | 102_1b1_Ar_sc_Meditron test 4 | 103_2b2_Ar_mc_LittC2SE train 5 | 104_1b1_Al_sc_Litt3200 test 6 | 104_1b1_Ar_sc_Litt3200 test 7 | 104_1b1_Ll_sc_Litt3200 test 8 | 104_1b1_Lr_sc_Litt3200 test 9 | 104_1b1_Pl_sc_Litt3200 test 10 | 104_1b1_Pr_sc_Litt3200 test 11 | 105_1b1_Tc_sc_Meditron train 12 | 106_2b1_Pl_mc_LittC2SE train 13 | 106_2b1_Pr_mc_LittC2SE train 14 | 107_2b3_Al_mc_AKGC417L train 15 | 107_2b3_Ar_mc_AKGC417L train 16 | 107_2b3_Ll_mc_AKGC417L train 17 | 107_2b3_Lr_mc_AKGC417L train 18 | 107_2b3_Pl_mc_AKGC417L train 19 | 107_2b3_Pr_mc_AKGC417L train 20 | 107_2b3_Tc_mc_AKGC417L train 21 | 107_2b4_Al_mc_AKGC417L train 22 | 107_2b4_Ar_mc_AKGC417L train 23 | 107_2b4_Ll_mc_AKGC417L train 24 | 107_2b4_Lr_mc_AKGC417L train 25 | 107_2b4_Pl_mc_AKGC417L train 26 | 107_2b4_Pr_mc_AKGC417L train 27 | 107_2b4_Tc_mc_AKGC417L train 28 | 107_2b5_Al_mc_AKGC417L train 29 | 107_2b5_Ar_mc_AKGC417L train 30 | 107_2b5_Ll_mc_AKGC417L train 31 | 107_2b5_Lr_mc_AKGC417L train 32 | 107_2b5_Pl_mc_AKGC417L train 33 | 107_2b5_Pr_mc_AKGC417L train 34 | 107_2b5_Tc_mc_AKGC417L train 35 | 107_3p2_Al_mc_AKGC417L train 36 | 107_3p2_Ar_mc_AKGC417L train 37 | 107_3p2_Ll_mc_AKGC417L train 38 | 107_3p2_Lr_mc_AKGC417L train 39 | 107_3p2_Pl_mc_AKGC417L train 40 | 107_3p2_Pr_mc_AKGC417L train 41 | 107_3p2_Tc_mc_AKGC417L train 42 | 108_1b1_Al_sc_Meditron train 43 | 109_1b1_Al_sc_Litt3200 test 44 | 109_1b1_Ar_sc_Litt3200 test 45 | 109_1b1_Ll_sc_Litt3200 test 46 | 109_1b1_Lr_sc_Litt3200 test 47 | 109_1b1_Pl_sc_Litt3200 test 48 | 109_1b1_Pr_sc_Litt3200 test 49 | 110_1b1_Pr_sc_Meditron train 50 | 110_1p1_Al_sc_Meditron train 51 | 110_1p1_Ll_sc_Meditron train 52 | 110_1p1_Lr_sc_Meditron train 53 | 110_1p1_Pr_sc_Meditron train 54 | 111_1b2_Tc_sc_Meditron train 55 | 111_1b3_Tc_sc_Meditron train 56 | 112_1b1_Ar_sc_Meditron train 57 | 112_1b1_Lr_sc_Meditron train 58 | 112_1p1_Ll_sc_Litt3200 train 59 | 112_1p1_Pl_sc_Litt3200 train 60 | 112_1p1_Pr_sc_Litt3200 train 61 | 113_1b1_Al_sc_Litt3200 test 62 | 113_1b1_Ar_sc_Litt3200 test 63 | 113_1b1_Ll_sc_Litt3200 test 64 | 113_1b1_Lr_sc_Litt3200 test 65 | 113_1b1_Pl_sc_Litt3200 test 66 | 113_1b1_Pr_sc_Litt3200 test 67 | 114_1b4_Al_mc_AKGC417L train 68 | 114_1b4_Ar_mc_AKGC417L train 69 | 114_1b4_Lr_mc_AKGC417L train 70 | 114_1b4_Pl_mc_AKGC417L train 71 | 114_1b4_Pr_mc_AKGC417L train 72 | 115_1b1_Ar_sc_Meditron train 73 | 116_1b2_Pl_sc_Meditron train 74 | 116_1b2_Tc_sc_Meditron train 75 | 117_1b2_Tc_mc_LittC2SE train 76 | 117_1b3_Tc_mc_LittC2SE train 77 | 118_1b1_Al_sc_Litt3200 test 78 | 118_1b1_Ar_sc_Litt3200 test 79 | 118_1b1_Ll_sc_Litt3200 test 80 | 118_1b1_Lr_sc_Litt3200 test 81 | 118_1b1_Pl_sc_Litt3200 test 82 | 118_1b1_Pr_sc_Litt3200 test 83 | 119_1b1_Ar_sc_Meditron test 84 | 120_1b1_Al_sc_Meditron train 85 | 120_1b1_Ar_sc_Meditron train 86 | 120_1b1_Lr_sc_Meditron train 87 | 120_1b1_Pl_sc_Meditron train 88 | 120_1b1_Pr_sc_Meditron train 89 | 121_1b1_Tc_sc_Meditron train 90 | 121_1p1_Tc_sc_Meditron train 91 | 122_2b1_Al_mc_LittC2SE train 92 | 122_2b1_Ar_mc_LittC2SE train 93 | 122_2b1_Tc_mc_LittC2SE train 94 | 122_2b2_Al_mc_LittC2SE train 95 | 122_2b2_Ar_mc_LittC2SE train 96 | 122_2b2_Tc_mc_LittC2SE train 97 | 122_2b3_Al_mc_LittC2SE train 98 | 122_2b3_Ar_mc_LittC2SE train 99 | 122_2b3_Tc_mc_LittC2SE train 100 | 123_1b1_Al_sc_Meditron test 101 | 124_1b1_Al_sc_Litt3200 test 102 | 124_1b1_Ar_sc_Litt3200 test 103 | 124_1b1_Ll_sc_Litt3200 test 104 | 124_1b1_Lr_sc_Litt3200 test 105 | 124_1b1_Pl_sc_Litt3200 test 106 | 124_1b1_Pr_sc_Litt3200 test 107 | 125_1b1_Tc_sc_Meditron train 108 | 126_1b1_Al_sc_Meditron test 109 | 127_1b1_Ar_sc_Meditron train 110 | 128_1b3_Tc_mc_LittC2SE train 111 | 129_1b1_Ar_sc_Meditron train 112 | 130_1p2_Al_mc_AKGC417L train 113 | 130_1p2_Ar_mc_AKGC417L train 114 | 130_1p2_Ll_mc_AKGC417L train 115 | 130_1p2_Lr_mc_AKGC417L train 116 | 130_1p2_Pl_mc_AKGC417L train 117 | 130_1p2_Pr_mc_AKGC417L train 118 | 130_1p2_Tc_mc_AKGC417L train 119 | 130_1p3_Al_mc_AKGC417L train 120 | 130_1p3_Ar_mc_AKGC417L train 121 | 130_1p3_Ll_mc_AKGC417L train 122 | 130_1p3_Lr_mc_AKGC417L train 123 | 130_1p3_Pl_mc_AKGC417L train 124 | 130_1p3_Pr_mc_AKGC417L train 125 | 130_1p3_Tc_mc_AKGC417L train 126 | 130_1p4_Al_mc_AKGC417L train 127 | 130_1p4_Ar_mc_AKGC417L train 128 | 130_1p4_Ll_mc_AKGC417L train 129 | 130_1p4_Lr_mc_AKGC417L train 130 | 130_1p4_Pl_mc_AKGC417L train 131 | 130_1p4_Pr_mc_AKGC417L train 132 | 130_1p4_Tc_mc_AKGC417L train 133 | 130_2b2_Al_mc_AKGC417L train 134 | 130_2b2_Ar_mc_AKGC417L train 135 | 130_2b2_Ll_mc_AKGC417L train 136 | 130_2b2_Lr_mc_AKGC417L train 137 | 130_2b2_Pl_mc_AKGC417L train 138 | 130_2b2_Pr_mc_AKGC417L train 139 | 130_2b2_Tc_mc_AKGC417L train 140 | 130_2b3_Al_mc_AKGC417L train 141 | 130_2b3_Ar_mc_AKGC417L train 142 | 130_2b3_Ll_mc_AKGC417L train 143 | 130_2b3_Lr_mc_AKGC417L train 144 | 130_2b3_Pl_mc_AKGC417L train 145 | 130_2b3_Pr_mc_AKGC417L train 146 | 130_2b3_Tc_mc_AKGC417L train 147 | 130_2b4_Al_mc_AKGC417L train 148 | 130_2b4_Ar_mc_AKGC417L train 149 | 130_2b4_Ll_mc_AKGC417L train 150 | 130_2b4_Lr_mc_AKGC417L train 151 | 130_2b4_Pl_mc_AKGC417L train 152 | 130_2p3_Pl_mc_AKGC417L train 153 | 130_2p5_Al_mc_AKGC417L train 154 | 130_2p5_Ar_mc_AKGC417L train 155 | 130_2p5_Lr_mc_AKGC417L train 156 | 130_2p5_Pl_mc_AKGC417L train 157 | 130_2p5_Pr_mc_AKGC417L train 158 | 130_2p5_Tc_mc_AKGC417L train 159 | 130_3b3_Ll_mc_AKGC417L train 160 | 130_3b4_Al_mc_AKGC417L train 161 | 130_3b4_Ar_mc_AKGC417L train 162 | 130_3b4_Lr_mc_AKGC417L train 163 | 130_3b4_Pl_mc_AKGC417L train 164 | 130_3b4_Pr_mc_AKGC417L train 165 | 130_3p2_Al_mc_AKGC417L train 166 | 130_3p2_Ar_mc_AKGC417L train 167 | 130_3p2_Pl_mc_AKGC417L train 168 | 130_3p2_Pr_mc_AKGC417L train 169 | 130_3p2_Tc_mc_AKGC417L train 170 | 130_3p3_Al_mc_AKGC417L train 171 | 130_3p3_Pl_mc_AKGC417L train 172 | 130_3p3_Pr_mc_AKGC417L train 173 | 130_3p3_Tc_mc_AKGC417L train 174 | 130_3p4_Al_mc_AKGC417L train 175 | 130_3p4_Pl_mc_AKGC417L train 176 | 130_3p4_Pr_mc_AKGC417L train 177 | 130_3p4_Tc_mc_AKGC417L train 178 | 131_1b1_Al_sc_Meditron train 179 | 132_2b1_Lr_mc_LittC2SE train 180 | 132_2b2_Lr_mc_LittC2SE train 181 | 133_2p2_Al_mc_AKGC417L test 182 | 133_2p2_Ar_mc_AKGC417L test 183 | 133_2p2_Pl_mc_AKGC417L test 184 | 133_2p2_Tc_mc_AKGC417L test 185 | 133_2p3_Al_mc_AKGC417L test 186 | 133_2p3_Ar_mc_AKGC417L test 187 | 133_2p3_Pl_mc_AKGC417L test 188 | 133_2p3_Pr_mc_AKGC417L test 189 | 133_2p3_Tc_mc_AKGC417L test 190 | 133_2p4_Al_mc_AKGC417L test 191 | 133_2p4_Ar_mc_AKGC417L test 192 | 133_2p4_Pl_mc_AKGC417L test 193 | 133_2p4_Pr_mc_AKGC417L test 194 | 133_2p4_Tc_mc_AKGC417L test 195 | 133_3p2_Al_mc_AKGC417L test 196 | 133_3p2_Ar_mc_AKGC417L test 197 | 133_3p2_Pl_mc_AKGC417L test 198 | 133_3p2_Pr_mc_AKGC417L test 199 | 133_3p4_Tc_mc_AKGC417L test 200 | 134_2b1_Al_mc_LittC2SE train 201 | 134_2b1_Ar_mc_LittC2SE train 202 | 134_2b2_Al_mc_LittC2SE train 203 | 134_2b2_Ar_mc_LittC2SE train 204 | 134_2b3_Ar_mc_LittC2SE train 205 | 135_2b1_Al_mc_LittC2SE train 206 | 135_2b1_Ar_mc_LittC2SE train 207 | 135_2b1_Pl_mc_LittC2SE train 208 | 135_2b1_Tc_mc_LittC2SE train 209 | 135_2b2_Al_mc_LittC2SE train 210 | 135_2b2_Ar_mc_LittC2SE train 211 | 135_2b2_Pl_mc_LittC2SE train 212 | 135_2b2_Tc_mc_LittC2SE train 213 | 135_2b3_Al_mc_LittC2SE train 214 | 135_2b3_Ar_mc_LittC2SE train 215 | 135_2b3_Pl_mc_LittC2SE train 216 | 135_2b3_Pr_mc_LittC2SE train 217 | 135_2b3_Tc_mc_LittC2SE train 218 | 136_1b1_Ar_sc_Meditron test 219 | 137_1b1_Ar_sc_Meditron train 220 | 137_1b1_Ll_sc_Meditron train 221 | 138_1p2_Al_mc_AKGC417L train 222 | 138_1p2_Ar_mc_AKGC417L train 223 | 138_1p2_Ll_mc_AKGC417L train 224 | 138_1p2_Lr_mc_AKGC417L train 225 | 138_1p2_Pl_mc_AKGC417L train 226 | 138_1p2_Pr_mc_AKGC417L train 227 | 138_1p2_Tc_mc_AKGC417L train 228 | 138_1p3_Al_mc_AKGC417L train 229 | 138_1p3_Ar_mc_AKGC417L train 230 | 138_1p3_Ll_mc_AKGC417L train 231 | 138_1p3_Lr_mc_AKGC417L train 232 | 138_1p3_Pl_mc_AKGC417L train 233 | 138_1p3_Pr_mc_AKGC417L train 234 | 138_1p3_Tc_mc_AKGC417L train 235 | 138_1p4_Ar_mc_AKGC417L train 236 | 138_1p4_Ll_mc_AKGC417L train 237 | 138_1p4_Lr_mc_AKGC417L train 238 | 138_1p4_Pl_mc_AKGC417L train 239 | 138_1p4_Pr_mc_AKGC417L train 240 | 138_1p4_Tc_mc_AKGC417L train 241 | 138_2p2_Al_mc_AKGC417L train 242 | 138_2p2_Ar_mc_AKGC417L train 243 | 138_2p2_Ll_mc_AKGC417L train 244 | 138_2p2_Lr_mc_AKGC417L train 245 | 138_2p2_Pl_mc_AKGC417L train 246 | 138_2p2_Pr_mc_AKGC417L train 247 | 138_2p2_Tc_mc_AKGC417L train 248 | 139_1b1_Al_sc_Litt3200 test 249 | 139_1b1_Ar_sc_Litt3200 test 250 | 139_1b1_Ll_sc_Litt3200 test 251 | 139_1b1_Lr_sc_Litt3200 test 252 | 139_1b1_Pl_sc_Litt3200 test 253 | 139_1b1_Pr_sc_Litt3200 test 254 | 140_2b2_Ll_mc_LittC2SE train 255 | 140_2b2_Tc_mc_LittC2SE train 256 | 140_2b3_Ll_mc_LittC2SE train 257 | 140_2b3_Tc_mc_LittC2SE train 258 | 141_1b1_Pr_mc_LittC2SE train 259 | 141_1b2_Ar_mc_LittC2SE train 260 | 141_1b2_Lr_mc_LittC2SE train 261 | 141_1b2_Pr_mc_LittC2SE train 262 | 141_1b2_Tc_mc_LittC2SE train 263 | 141_1b3_Al_mc_LittC2SE train 264 | 141_1b3_Ar_mc_LittC2SE train 265 | 141_1b3_Pr_mc_LittC2SE train 266 | 142_1b1_Pl_mc_LittC2SE train 267 | 143_1b1_Al_sc_Meditron test 268 | 144_1b1_Al_sc_Meditron test 269 | 144_1b1_Tc_sc_Meditron test 270 | 145_2b2_Al_mc_AKGC417L test 271 | 145_2b2_Ar_mc_AKGC417L test 272 | 145_2b2_Lr_mc_AKGC417L test 273 | 145_2b2_Pr_mc_AKGC417L test 274 | 145_3b2_Ar_mc_AKGC417L test 275 | 145_3b2_Lr_mc_AKGC417L test 276 | 145_3b4_Pl_mc_AKGC417L test 277 | 146_2b2_Pl_mc_AKGC417L test 278 | 146_2b4_Al_mc_AKGC417L test 279 | 146_2b4_Ar_mc_AKGC417L test 280 | 146_2b4_Ll_mc_AKGC417L test 281 | 146_2b4_Lr_mc_AKGC417L test 282 | 146_2b4_Pr_mc_AKGC417L test 283 | 146_8p3_Al_mc_AKGC417L test 284 | 146_8p3_Ar_mc_AKGC417L test 285 | 146_8p3_Lr_mc_AKGC417L test 286 | 146_8p3_Pl_mc_AKGC417L test 287 | 146_8p3_Pr_mc_AKGC417L test 288 | 147_1b2_Tc_mc_AKGC417L test 289 | 147_1b3_Tc_mc_AKGC417L test 290 | 147_1b4_Tc_mc_AKGC417L test 291 | 147_2b2_Al_mc_AKGC417L test 292 | 147_2b2_Ar_mc_AKGC417L test 293 | 147_2b2_Pl_mc_AKGC417L test 294 | 147_2b3_Al_mc_AKGC417L test 295 | 147_2b3_Ar_mc_AKGC417L test 296 | 147_2b3_Ll_mc_AKGC417L test 297 | 147_2b3_Lr_mc_AKGC417L test 298 | 147_2b3_Pl_mc_AKGC417L test 299 | 147_2b4_Al_mc_AKGC417L test 300 | 147_2b4_Ar_mc_AKGC417L test 301 | 147_2b4_Ll_mc_AKGC417L test 302 | 147_2b4_Lr_mc_AKGC417L test 303 | 147_2b4_Pl_mc_AKGC417L test 304 | 148_1b1_Al_sc_Meditron train 305 | 149_1b1_Al_sc_Meditron test 306 | 149_1b1_Lr_sc_Meditron test 307 | 149_1b1_Pl_sc_Meditron test 308 | 150_1b2_Al_sc_Meditron test 309 | 151_2p2_Al_mc_AKGC417L test 310 | 151_2p2_Ar_mc_AKGC417L test 311 | 151_2p2_Ll_mc_AKGC417L test 312 | 151_2p2_Lr_mc_AKGC417L test 313 | 151_2p2_Pl_mc_AKGC417L test 314 | 151_2p2_Pr_mc_AKGC417L test 315 | 151_2p2_Tc_mc_AKGC417L test 316 | 151_2p3_Al_mc_AKGC417L test 317 | 151_2p3_Ar_mc_AKGC417L test 318 | 151_2p3_Ll_mc_AKGC417L test 319 | 151_2p3_Lr_mc_AKGC417L test 320 | 151_2p3_Pl_mc_AKGC417L test 321 | 151_2p3_Pr_mc_AKGC417L test 322 | 151_2p3_Tc_mc_AKGC417L test 323 | 151_2p4_Al_mc_AKGC417L test 324 | 151_2p4_Ar_mc_AKGC417L test 325 | 151_2p4_Ll_mc_AKGC417L test 326 | 151_2p4_Lr_mc_AKGC417L test 327 | 151_2p4_Pl_mc_AKGC417L test 328 | 151_2p4_Pr_mc_AKGC417L test 329 | 151_2p4_Tc_mc_AKGC417L test 330 | 151_3p2_Al_mc_AKGC417L test 331 | 151_3p2_Ar_mc_AKGC417L test 332 | 151_3p2_Lr_mc_AKGC417L test 333 | 151_3p2_Pl_mc_AKGC417L test 334 | 151_3p2_Pr_mc_AKGC417L test 335 | 151_3p2_Tc_mc_AKGC417L test 336 | 151_3p3_Ll_mc_AKGC417L test 337 | 152_1b1_Al_sc_Meditron train 338 | 153_1b1_Al_sc_Meditron train 339 | 154_1b3_Al_mc_AKGC417L train 340 | 154_1b3_Ar_mc_AKGC417L train 341 | 154_1b3_Ll_mc_AKGC417L train 342 | 154_1b3_Lr_mc_AKGC417L train 343 | 154_1b3_Pl_mc_AKGC417L train 344 | 154_1b3_Pr_mc_AKGC417L train 345 | 154_1b3_Tc_mc_AKGC417L train 346 | 154_2b4_Al_mc_AKGC417L train 347 | 154_2b4_Ar_mc_AKGC417L train 348 | 154_2b4_Ll_mc_AKGC417L train 349 | 154_2b4_Lr_mc_AKGC417L train 350 | 154_2b4_Pl_mc_AKGC417L train 351 | 154_2b4_Pr_mc_AKGC417L train 352 | 154_2b4_Tc_mc_AKGC417L train 353 | 154_3b3_Al_mc_AKGC417L train 354 | 154_3b3_Ar_mc_AKGC417L train 355 | 154_3b3_Ll_mc_AKGC417L train 356 | 154_4b4_Al_mc_AKGC417L train 357 | 154_4b4_Ar_mc_AKGC417L train 358 | 154_4b4_Ll_mc_AKGC417L train 359 | 154_4b4_Lr_mc_AKGC417L train 360 | 154_4b4_Pl_mc_AKGC417L train 361 | 154_4b4_Pr_mc_AKGC417L train 362 | 155_2b1_Al_mc_LittC2SE train 363 | 156_2b3_Al_mc_AKGC417L test 364 | 156_2b3_Ar_mc_AKGC417L train 365 | 156_2b3_Ll_mc_AKGC417L train 366 | 156_2b3_Lr_mc_AKGC417L test 367 | 156_2b3_Pl_mc_AKGC417L test 368 | 156_2b3_Pr_mc_AKGC417L train 369 | 156_5b3_Al_mc_AKGC417L train 370 | 156_5b3_Ar_mc_AKGC417L test 371 | 156_5b3_Ll_mc_AKGC417L test 372 | 156_5b3_Lr_mc_AKGC417L train 373 | 156_5b3_Pl_mc_AKGC417L test 374 | 156_5b3_Pr_mc_AKGC417L train 375 | 156_8b3_Al_mc_AKGC417L test 376 | 156_8b3_Ar_mc_AKGC417L train 377 | 156_8b3_Ll_mc_AKGC417L train 378 | 156_8b3_Lr_mc_AKGC417L test 379 | 156_8b3_Pl_mc_AKGC417L train 380 | 157_1b1_Al_sc_Meditron train 381 | 157_1b1_Ar_sc_Meditron train 382 | 157_1b1_Lr_sc_Meditron train 383 | 157_1b1_Pl_sc_Meditron train 384 | 157_1b1_Pr_sc_Meditron train 385 | 158_1b3_Ar_mc_LittC2SE train 386 | 158_1p2_Al_mc_AKGC417L train 387 | 158_1p2_Ar_mc_AKGC417L train 388 | 158_1p2_Ll_mc_AKGC417L train 389 | 158_1p2_Lr_mc_AKGC417L train 390 | 158_1p2_Pl_mc_AKGC417L train 391 | 158_1p2_Pr_mc_AKGC417L train 392 | 158_1p2_Tc_mc_AKGC417L train 393 | 158_1p3_Al_mc_AKGC417L train 394 | 158_1p3_Ar_mc_AKGC417L train 395 | 158_1p3_Ll_mc_AKGC417L train 396 | 158_1p3_Lr_mc_AKGC417L train 397 | 158_1p3_Pl_mc_AKGC417L train 398 | 158_1p3_Pr_mc_AKGC417L train 399 | 158_1p3_Tc_mc_AKGC417L train 400 | 158_1p4_Al_mc_AKGC417L train 401 | 158_1p4_Ar_mc_AKGC417L train 402 | 158_1p4_Lr_mc_AKGC417L train 403 | 158_1p4_Pl_mc_AKGC417L train 404 | 158_1p4_Pr_mc_AKGC417L train 405 | 158_1p4_Tc_mc_AKGC417L train 406 | 158_2p2_Ar_mc_AKGC417L train 407 | 158_2p3_Lr_mc_AKGC417L train 408 | 158_2p3_Tc_mc_AKGC417L train 409 | 159_1b1_Al_sc_Meditron train 410 | 159_1b1_Ar_sc_Meditron train 411 | 159_1b1_Ll_sc_Meditron train 412 | 159_1b1_Pr_sc_Meditron train 413 | 160_1b2_Al_mc_AKGC417L test 414 | 160_1b2_Ar_mc_AKGC417L test 415 | 160_1b2_Lr_mc_AKGC417L test 416 | 160_1b2_Pl_mc_AKGC417L test 417 | 160_1b2_Pr_mc_AKGC417L test 418 | 160_1b2_Tc_mc_AKGC417L test 419 | 160_1b3_Al_mc_AKGC417L test 420 | 160_1b3_Ar_mc_AKGC417L test 421 | 160_1b3_Lr_mc_AKGC417L test 422 | 160_1b3_Pl_mc_AKGC417L test 423 | 160_1b3_Pr_mc_AKGC417L test 424 | 160_1b3_Tc_mc_AKGC417L test 425 | 160_1b4_Al_mc_AKGC417L test 426 | 160_1b4_Ar_mc_AKGC417L test 427 | 160_1b4_Lr_mc_AKGC417L test 428 | 160_1b4_Pl_mc_AKGC417L test 429 | 160_1b4_Pr_mc_AKGC417L test 430 | 160_1b4_Tc_mc_AKGC417L test 431 | 160_2b3_Lr_mc_AKGC417L test 432 | 160_2b4_Ar_mc_AKGC417L test 433 | 160_2b4_Pl_mc_AKGC417L test 434 | 160_2b4_Pr_mc_AKGC417L test 435 | 160_2b4_Tc_mc_AKGC417L test 436 | 161_1b1_Al_sc_Meditron train 437 | 161_1b1_Pl_sc_Meditron train 438 | 162_1b2_Al_mc_AKGC417L train 439 | 162_1b2_Ar_mc_AKGC417L train 440 | 162_1b2_Ll_mc_AKGC417L train 441 | 162_1b2_Lr_mc_AKGC417L train 442 | 162_1b2_Pl_mc_AKGC417L train 443 | 162_1b2_Pr_mc_AKGC417L train 444 | 162_1b2_Tc_mc_AKGC417L train 445 | 162_2b2_Al_mc_AKGC417L train 446 | 162_2b2_Ar_mc_AKGC417L train 447 | 162_2b2_Pl_mc_AKGC417L train 448 | 162_2b2_Pr_mc_AKGC417L train 449 | 162_2b2_Tc_mc_AKGC417L train 450 | 162_2b3_Al_mc_AKGC417L train 451 | 162_2b3_Ar_mc_AKGC417L train 452 | 162_2b3_Lr_mc_AKGC417L train 453 | 162_2b3_Pl_mc_AKGC417L train 454 | 162_2b3_Pr_mc_AKGC417L train 455 | 162_2b3_Tc_mc_AKGC417L train 456 | 162_2b4_Al_mc_AKGC417L train 457 | 162_2b4_Ar_mc_AKGC417L train 458 | 162_2b4_Lr_mc_AKGC417L train 459 | 162_2b4_Pl_mc_AKGC417L train 460 | 162_2b4_Pr_mc_AKGC417L train 461 | 162_2b4_Tc_mc_AKGC417L train 462 | 163_2b2_Al_mc_AKGC417L train 463 | 163_2b2_Ar_mc_AKGC417L train 464 | 163_2b2_Ll_mc_AKGC417L train 465 | 163_2b2_Lr_mc_AKGC417L train 466 | 163_2b2_Pl_mc_AKGC417L train 467 | 163_2b2_Pr_mc_AKGC417L train 468 | 163_2b2_Tc_mc_AKGC417L train 469 | 163_8b3_Al_mc_AKGC417L train 470 | 163_8b3_Ar_mc_AKGC417L train 471 | 163_8b3_Ll_mc_AKGC417L train 472 | 163_8b3_Lr_mc_AKGC417L train 473 | 163_8b3_Pl_mc_AKGC417L train 474 | 163_8b3_Pr_mc_AKGC417L train 475 | 164_1b1_Ll_sc_Meditron train 476 | 165_1b1_Ar_sc_Meditron test 477 | 165_1b1_Pl_sc_Meditron test 478 | 165_1b1_Pr_sc_Meditron test 479 | 166_1p1_Al_sc_Meditron train 480 | 166_1p1_Ar_sc_Meditron train 481 | 166_1p1_Ll_sc_Meditron train 482 | 166_1p1_Pl_sc_Meditron train 483 | 166_1p1_Pr_sc_Meditron train 484 | 167_1b1_Al_sc_Meditron train 485 | 167_1b1_Pr_sc_Meditron train 486 | 168_1b1_Al_sc_Meditron train 487 | 169_1b1_Lr_sc_Meditron test 488 | 169_1b2_Ll_sc_Meditron test 489 | 170_1b2_Al_mc_AKGC417L test 490 | 170_1b2_Ar_mc_AKGC417L test 491 | 170_1b2_Lr_mc_AKGC417L test 492 | 170_1b2_Pl_mc_AKGC417L test 493 | 170_1b2_Pr_mc_AKGC417L test 494 | 170_1b2_Tc_mc_AKGC417L test 495 | 170_1b3_Al_mc_AKGC417L test 496 | 170_1b3_Ar_mc_AKGC417L test 497 | 170_1b3_Ll_mc_AKGC417L test 498 | 170_1b3_Lr_mc_AKGC417L test 499 | 170_1b3_Pl_mc_AKGC417L test 500 | 170_1b3_Pr_mc_AKGC417L test 501 | 170_1b3_Tc_mc_AKGC417L test 502 | 170_1b4_Al_mc_AKGC417L test 503 | 170_1b4_Ar_mc_AKGC417L test 504 | 170_1b4_Lr_mc_AKGC417L test 505 | 170_1b4_Pl_mc_AKGC417L test 506 | 170_1b4_Pr_mc_AKGC417L test 507 | 170_1b4_Tc_mc_AKGC417L test 508 | 170_2b2_Al_mc_AKGC417L test 509 | 170_2b2_Ar_mc_AKGC417L test 510 | 170_2b2_Lr_mc_AKGC417L test 511 | 170_2b2_Pl_mc_AKGC417L test 512 | 170_2b2_Pr_mc_AKGC417L test 513 | 170_2b2_Tc_mc_AKGC417L test 514 | 171_1b1_Al_sc_Meditron test 515 | 172_1b3_Al_mc_AKGC417L train 516 | 172_1b3_Ar_mc_AKGC417L train 517 | 172_1b3_Ll_mc_AKGC417L train 518 | 172_1b3_Lr_mc_AKGC417L train 519 | 172_1b3_Pl_mc_AKGC417L train 520 | 172_1b3_Pr_mc_AKGC417L train 521 | 172_1b3_Tc_mc_AKGC417L train 522 | 172_1b4_Al_mc_AKGC417L train 523 | 172_1b4_Ar_mc_AKGC417L train 524 | 172_1b4_Ll_mc_AKGC417L train 525 | 172_1b4_Lr_mc_AKGC417L train 526 | 172_1b4_Pl_mc_AKGC417L train 527 | 172_1b4_Pr_mc_AKGC417L train 528 | 172_1b4_Tc_mc_AKGC417L train 529 | 172_1b5_Al_mc_AKGC417L train 530 | 172_1b5_Ar_mc_AKGC417L train 531 | 172_1b5_Ll_mc_AKGC417L train 532 | 172_1b5_Lr_mc_AKGC417L train 533 | 172_1b5_Pl_mc_AKGC417L train 534 | 172_1b5_Pr_mc_AKGC417L train 535 | 172_1b5_Tc_mc_AKGC417L train 536 | 172_2b5_Al_mc_AKGC417L train 537 | 172_2b5_Ar_mc_AKGC417L train 538 | 172_2b5_Lr_mc_AKGC417L train 539 | 172_2b5_Pl_mc_AKGC417L train 540 | 172_2b5_Pr_mc_AKGC417L train 541 | 172_2b5_Tc_mc_AKGC417L train 542 | 173_1b1_Al_sc_Meditron test 543 | 174_1p2_Ar_mc_AKGC417L test 544 | 174_1p2_Ll_mc_AKGC417L test 545 | 174_1p2_Lr_mc_AKGC417L test 546 | 174_1p2_Pl_mc_AKGC417L test 547 | 174_1p2_Pr_mc_AKGC417L test 548 | 174_1p2_Tc_mc_AKGC417L test 549 | 174_1p3_Ar_mc_AKGC417L test 550 | 174_1p3_Ll_mc_AKGC417L test 551 | 174_1p3_Lr_mc_AKGC417L test 552 | 174_1p3_Pl_mc_AKGC417L test 553 | 174_1p3_Pr_mc_AKGC417L test 554 | 174_1p3_Tc_mc_AKGC417L test 555 | 174_1p4_Ar_mc_AKGC417L test 556 | 174_1p4_Ll_mc_AKGC417L test 557 | 174_1p4_Lr_mc_AKGC417L test 558 | 174_1p4_Pl_mc_AKGC417L test 559 | 174_1p4_Pr_mc_AKGC417L test 560 | 174_1p4_Tc_mc_AKGC417L test 561 | 174_2p3_Al_mc_AKGC417L test 562 | 174_2p3_Ar_mc_AKGC417L test 563 | 174_2p3_Pl_mc_AKGC417L test 564 | 174_2p3_Pr_mc_AKGC417L test 565 | 174_2p3_Tc_mc_AKGC417L test 566 | 175_1b1_Al_sc_Litt3200 test 567 | 175_1b1_Ar_sc_Litt3200 test 568 | 175_1b1_Ll_sc_Litt3200 test 569 | 175_1b1_Lr_sc_Litt3200 test 570 | 175_1b1_Pl_sc_Litt3200 test 571 | 175_1b1_Pr_sc_Litt3200 test 572 | 176_1b3_Al_mc_AKGC417L test 573 | 176_1b3_Ar_mc_AKGC417L test 574 | 176_1b3_Ll_mc_AKGC417L test 575 | 176_1b3_Lr_mc_AKGC417L test 576 | 176_1b3_Pl_mc_AKGC417L test 577 | 176_1b3_Pr_mc_AKGC417L test 578 | 176_1b3_Tc_mc_AKGC417L test 579 | 176_1b4_Al_mc_AKGC417L test 580 | 176_1b4_Ar_mc_AKGC417L test 581 | 176_1b4_Ll_mc_AKGC417L test 582 | 176_1b4_Lr_mc_AKGC417L test 583 | 176_1b4_Pl_mc_AKGC417L test 584 | 176_1b4_Pr_mc_AKGC417L test 585 | 176_1b4_Tc_mc_AKGC417L test 586 | 176_2b3_Al_mc_AKGC417L test 587 | 176_2b3_Ar_mc_AKGC417L test 588 | 176_2b3_Ll_mc_AKGC417L test 589 | 176_2b3_Lr_mc_AKGC417L test 590 | 176_2b3_Pl_mc_AKGC417L test 591 | 176_2b3_Pr_mc_AKGC417L test 592 | 176_2b3_Tc_mc_AKGC417L test 593 | 177_1b2_Al_mc_AKGC417L test 594 | 177_1b2_Ar_mc_AKGC417L test 595 | 177_1b2_Lr_mc_AKGC417L test 596 | 177_1b2_Pl_mc_AKGC417L test 597 | 177_1b2_Pr_mc_AKGC417L test 598 | 177_1b2_Tc_mc_AKGC417L test 599 | 177_1b4_Al_mc_AKGC417L test 600 | 177_1b4_Ar_mc_AKGC417L test 601 | 177_1b4_Lr_mc_AKGC417L test 602 | 177_1b4_Pl_mc_AKGC417L test 603 | 177_1b4_Pr_mc_AKGC417L test 604 | 177_1b4_Tc_mc_AKGC417L test 605 | 177_2b4_Al_mc_AKGC417L test 606 | 177_2b4_Lr_mc_AKGC417L test 607 | 177_2b4_Pl_mc_AKGC417L test 608 | 177_2b4_Pr_mc_AKGC417L test 609 | 177_2b4_Tc_mc_AKGC417L test 610 | 178_1b2_Al_mc_AKGC417L test 611 | 178_1b2_Ar_mc_AKGC417L test 612 | 178_1b2_Lr_mc_AKGC417L test 613 | 178_1b2_Pl_mc_AKGC417L test 614 | 178_1b2_Pr_mc_AKGC417L test 615 | 178_1b2_Tc_mc_AKGC417L test 616 | 178_1b3_Al_mc_AKGC417L test 617 | 178_1b3_Ar_mc_AKGC417L test 618 | 178_1b3_Lr_mc_AKGC417L test 619 | 178_1b3_Pl_mc_AKGC417L test 620 | 178_1b3_Pr_mc_AKGC417L test 621 | 178_1b3_Tc_mc_AKGC417L test 622 | 178_1b6_Al_mc_AKGC417L test 623 | 178_1b6_Ar_mc_AKGC417L test 624 | 178_1b6_Ll_mc_AKGC417L test 625 | 178_1b6_Lr_mc_AKGC417L test 626 | 178_1b6_Pl_mc_AKGC417L test 627 | 178_1b6_Pr_mc_AKGC417L test 628 | 178_1b6_Tc_mc_AKGC417L test 629 | 178_2b2_Al_mc_AKGC417L test 630 | 178_2b2_Ar_mc_AKGC417L test 631 | 178_2b2_Lr_mc_AKGC417L test 632 | 178_2b2_Pr_mc_AKGC417L test 633 | 178_2b2_Tc_mc_AKGC417L test 634 | 179_1b1_Al_sc_Meditron train 635 | 179_1b1_Tc_sc_Meditron train 636 | 180_1b4_Al_mc_AKGC417L train 637 | 180_1b4_Ar_mc_AKGC417L train 638 | 180_1b4_Lr_mc_AKGC417L train 639 | 180_1b4_Pl_mc_AKGC417L train 640 | 180_1b4_Pr_mc_AKGC417L train 641 | 181_1b1_Ar_mc_LittC2SE train 642 | 181_1b1_Tc_mc_LittC2SE train 643 | 181_1b2_Ar_mc_LittC2SE train 644 | 181_1b3_Tc_mc_LittC2SE train 645 | 182_1b1_Tc_sc_Meditron test 646 | 183_1b1_Pl_sc_Meditron train 647 | 183_1b1_Tc_sc_Meditron train 648 | 184_1b1_Ar_sc_Meditron train 649 | 185_1b1_Al_sc_Litt3200 test 650 | 185_1b1_Ar_sc_Litt3200 test 651 | 185_1b1_Ll_sc_Litt3200 test 652 | 185_1b1_Lr_sc_Litt3200 test 653 | 185_1b1_Pl_sc_Litt3200 test 654 | 185_1b1_Pr_sc_Litt3200 test 655 | 186_2b2_Al_mc_AKGC417L train 656 | 186_2b2_Ar_mc_AKGC417L train 657 | 186_2b2_Lr_mc_AKGC417L train 658 | 186_2b2_Pl_mc_AKGC417L train 659 | 186_2b2_Pr_mc_AKGC417L train 660 | 186_2b2_Tc_mc_AKGC417L train 661 | 186_2b3_Al_mc_AKGC417L train 662 | 186_2b3_Ar_mc_AKGC417L train 663 | 186_2b3_Lr_mc_AKGC417L train 664 | 186_2b3_Pl_mc_AKGC417L train 665 | 186_2b3_Pr_mc_AKGC417L train 666 | 186_2b3_Tc_mc_AKGC417L train 667 | 186_2b4_Al_mc_AKGC417L train 668 | 186_2b4_Ar_mc_AKGC417L train 669 | 186_2b4_Lr_mc_AKGC417L train 670 | 186_2b4_Pl_mc_AKGC417L train 671 | 186_2b4_Pr_mc_AKGC417L train 672 | 186_2b4_Tc_mc_AKGC417L train 673 | 186_3b3_Al_mc_AKGC417L train 674 | 186_3b3_Ar_mc_AKGC417L train 675 | 186_3b3_Lr_mc_AKGC417L train 676 | 186_3b3_Pl_mc_AKGC417L train 677 | 186_3b3_Pr_mc_AKGC417L train 678 | 186_3b3_Tc_mc_AKGC417L train 679 | 187_1b1_Ll_sc_Meditron test 680 | 188_1b1_Al_sc_Meditron train 681 | 188_1b1_Ar_sc_Meditron train 682 | 188_1b1_Pl_sc_Meditron train 683 | 188_1b1_Tc_sc_Meditron train 684 | 189_1b2_Lr_mc_LittC2SE train 685 | 190_1b1_Tc_sc_Meditron train 686 | 191_2b1_Pl_mc_LittC2SE train 687 | 191_2b1_Pr_mc_LittC2SE train 688 | 191_2b2_Tc_mc_LittC2SE train 689 | 192_2b1_Al_mc_LittC2SE train 690 | 192_2b1_Ar_mc_LittC2SE train 691 | 192_2b2_Al_mc_LittC2SE train 692 | 192_2b2_Ar_mc_LittC2SE train 693 | 192_2b3_Al_mc_LittC2SE train 694 | 192_2b3_Ar_mc_LittC2SE train 695 | 193_1b2_Al_mc_AKGC417L train 696 | 193_1b2_Ar_mc_AKGC417L train 697 | 193_1b2_Ll_mc_AKGC417L train 698 | 193_1b2_Pl_mc_AKGC417L train 699 | 193_1b2_Pr_mc_AKGC417L train 700 | 193_1b2_Tc_mc_AKGC417L train 701 | 193_1b4_Lr_mc_AKGC417L train 702 | 193_7b3_Al_mc_AKGC417L train 703 | 193_7b3_Ar_mc_AKGC417L train 704 | 193_7b3_Ll_mc_AKGC417L train 705 | 193_7b3_Lr_mc_AKGC417L train 706 | 193_7b3_Pl_mc_AKGC417L train 707 | 193_7b3_Pr_mc_AKGC417L train 708 | 193_7b3_Tc_mc_AKGC417L train 709 | 194_1b1_Lr_sc_Meditron test 710 | 194_1b1_Pr_sc_Meditron test 711 | 195_1b1_Al_sc_Litt3200 test 712 | 195_1b1_Ar_sc_Litt3200 test 713 | 195_1b1_Ll_sc_Litt3200 test 714 | 195_1b1_Lr_sc_Litt3200 test 715 | 195_1b1_Pl_sc_Litt3200 test 716 | 195_1b1_Pr_sc_Litt3200 test 717 | 196_1b1_Pr_sc_Meditron train 718 | 197_1b1_Al_sc_Meditron train 719 | 197_1b1_Tc_sc_Meditron train 720 | 198_1b5_Al_mc_AKGC417L test 721 | 198_1b5_Ar_mc_AKGC417L test 722 | 198_1b5_Ll_mc_AKGC417L test 723 | 198_1b5_Lr_mc_AKGC417L test 724 | 198_1b5_Pl_mc_AKGC417L test 725 | 198_1b5_Pr_mc_AKGC417L test 726 | 198_1b5_Tc_mc_AKGC417L test 727 | 198_6p1_Al_mc_AKGC417L test 728 | 198_6p1_Ar_mc_AKGC417L test 729 | 198_6p1_Ll_mc_AKGC417L test 730 | 198_6p1_Lr_mc_AKGC417L test 731 | 198_6p1_Pl_mc_AKGC417L test 732 | 198_6p1_Pr_mc_AKGC417L test 733 | 198_6p1_Tc_mc_AKGC417L test 734 | 199_2b1_Ll_mc_LittC2SE train 735 | 199_2b3_Ll_mc_LittC2SE train 736 | 200_2p2_Al_mc_AKGC417L train 737 | 200_2p2_Ar_mc_AKGC417L train 738 | 200_2p2_Lr_mc_AKGC417L train 739 | 200_2p2_Pl_mc_AKGC417L train 740 | 200_2p2_Pr_mc_AKGC417L train 741 | 200_2p2_Tc_mc_AKGC417L train 742 | 200_2p3_Al_mc_AKGC417L train 743 | 200_2p3_Ar_mc_AKGC417L train 744 | 200_2p3_Lr_mc_AKGC417L train 745 | 200_2p3_Pl_mc_AKGC417L train 746 | 200_2p3_Pr_mc_AKGC417L train 747 | 200_2p3_Tc_mc_AKGC417L train 748 | 200_2p4_Al_mc_AKGC417L train 749 | 200_2p4_Ar_mc_AKGC417L train 750 | 200_2p4_Lr_mc_AKGC417L train 751 | 200_2p4_Pl_mc_AKGC417L train 752 | 200_2p4_Pr_mc_AKGC417L train 753 | 200_2p4_Tc_mc_AKGC417L train 754 | 200_3p4_Al_mc_AKGC417L train 755 | 200_3p4_Ar_mc_AKGC417L train 756 | 200_3p4_Pl_mc_AKGC417L train 757 | 200_3p4_Pr_mc_AKGC417L train 758 | 200_3p4_Tc_mc_AKGC417L train 759 | 201_1b1_Al_sc_Meditron train 760 | 201_1b1_Ar_sc_Meditron train 761 | 201_1b2_Al_sc_Meditron train 762 | 201_1b2_Ar_sc_Meditron train 763 | 201_1b3_Al_sc_Meditron train 764 | 201_1b3_Ar_sc_Meditron train 765 | 202_1b1_Ar_sc_Meditron test 766 | 203_1p2_Al_mc_AKGC417L train 767 | 203_1p2_Ar_mc_AKGC417L train 768 | 203_1p2_Lr_mc_AKGC417L train 769 | 203_1p2_Pl_mc_AKGC417L train 770 | 203_1p2_Pr_mc_AKGC417L train 771 | 203_1p2_Tc_mc_AKGC417L train 772 | 203_1p3_Al_mc_AKGC417L train 773 | 203_1p3_Ar_mc_AKGC417L train 774 | 203_1p3_Pl_mc_AKGC417L train 775 | 203_1p3_Pr_mc_AKGC417L train 776 | 203_1p3_Tc_mc_AKGC417L train 777 | 203_1p4_Al_mc_AKGC417L train 778 | 203_1p4_Ar_mc_AKGC417L train 779 | 203_1p4_Pl_mc_AKGC417L train 780 | 203_1p4_Pr_mc_AKGC417L train 781 | 203_1p4_Tc_mc_AKGC417L train 782 | 203_2p3_Al_mc_AKGC417L train 783 | 203_2p3_Ar_mc_AKGC417L train 784 | 203_2p3_Pl_mc_AKGC417L train 785 | 203_2p3_Pr_mc_AKGC417L train 786 | 203_2p3_Tc_mc_AKGC417L train 787 | 204_2b5_Al_mc_AKGC417L train 788 | 204_2b5_Ar_mc_AKGC417L train 789 | 204_2b5_Ll_mc_AKGC417L train 790 | 204_7p5_Al_mc_AKGC417L train 791 | 204_7p5_Ar_mc_AKGC417L train 792 | 204_7p5_Ll_mc_AKGC417L train 793 | 204_7p5_Lr_mc_AKGC417L train 794 | 204_7p5_Pr_mc_AKGC417L train 795 | 204_7p5_Tc_mc_AKGC417L train 796 | 205_1b3_Al_mc_AKGC417L test 797 | 205_1b3_Ar_mc_AKGC417L test 798 | 205_1b3_Ll_mc_AKGC417L test 799 | 205_1b3_Lr_mc_AKGC417L test 800 | 205_1b3_Pl_mc_AKGC417L test 801 | 205_1b3_Pr_mc_AKGC417L test 802 | 205_2b2_Pr_mc_AKGC417L test 803 | 205_2b3_Al_mc_AKGC417L test 804 | 205_2b3_Ar_mc_AKGC417L test 805 | 205_2b3_Ll_mc_AKGC417L test 806 | 205_2b4_Pl_mc_AKGC417L test 807 | 205_3b4_Al_mc_AKGC417L test 808 | 205_3b4_Ar_mc_AKGC417L test 809 | 205_3b4_Pl_mc_AKGC417L test 810 | 205_3b4_Pr_mc_AKGC417L test 811 | 205_4b2_Al_mc_AKGC417L test 812 | 205_4b2_Ar_mc_AKGC417L test 813 | 205_4b2_Lr_mc_AKGC417L test 814 | 205_4b2_Pl_mc_AKGC417L test 815 | 205_4b2_Pr_mc_AKGC417L test 816 | 206_1b1_Ar_sc_Meditron train 817 | 206_1b1_Lr_sc_Meditron train 818 | 206_1b1_Pl_sc_Meditron train 819 | 207_2b2_Al_mc_AKGC417L test 820 | 207_2b2_Ar_mc_AKGC417L test 821 | 207_2b2_Pl_mc_AKGC417L test 822 | 207_2b2_Pr_mc_AKGC417L test 823 | 207_2b2_Tc_mc_AKGC417L test 824 | 207_2b3_Al_mc_AKGC417L test 825 | 207_2b3_Ar_mc_AKGC417L test 826 | 207_2b3_Pl_mc_AKGC417L test 827 | 207_2b3_Pr_mc_AKGC417L test 828 | 207_2b3_Tc_mc_AKGC417L test 829 | 207_2b4_Al_mc_AKGC417L test 830 | 207_2b4_Ar_mc_AKGC417L test 831 | 207_2b4_Pl_mc_AKGC417L test 832 | 207_2b4_Pr_mc_AKGC417L test 833 | 207_2b4_Tc_mc_AKGC417L test 834 | 207_3b2_Al_mc_AKGC417L test 835 | 207_3b2_Ar_mc_AKGC417L test 836 | 207_3b2_Lr_mc_AKGC417L test 837 | 207_3b2_Pl_mc_AKGC417L test 838 | 207_3b2_Pr_mc_AKGC417L test 839 | 207_3b2_Tc_mc_AKGC417L test 840 | 208_1b1_Ll_sc_Meditron train 841 | 209_1b1_Tc_sc_Meditron train 842 | 210_1b1_Al_sc_Meditron train 843 | 210_1b1_Ar_sc_Meditron train 844 | 211_1p2_Ar_mc_AKGC417L test 845 | 211_1p2_Pl_mc_AKGC417L test 846 | 211_1p2_Pr_mc_AKGC417L test 847 | 211_1p3_Ar_mc_AKGC417L test 848 | 211_1p5_Ar_mc_AKGC417L test 849 | 211_2p2_Tc_mc_AKGC417L test 850 | 211_2p3_Tc_mc_AKGC417L test 851 | 211_2p4_Tc_mc_AKGC417L test 852 | 212_2b2_Tc_mc_LittC2SE train 853 | 213_1p2_Al_mc_AKGC417L train 854 | 213_1p2_Ar_mc_AKGC417L train 855 | 213_1p2_Lr_mc_AKGC417L train 856 | 213_1p2_Pl_mc_AKGC417L train 857 | 213_1p2_Pr_mc_AKGC417L train 858 | 213_1p2_Tc_mc_AKGC417L train 859 | 213_1p3_Al_mc_AKGC417L train 860 | 213_1p3_Ar_mc_AKGC417L train 861 | 213_1p3_Pl_mc_AKGC417L train 862 | 213_1p3_Pr_mc_AKGC417L train 863 | 213_1p5_Al_mc_AKGC417L train 864 | 213_1p5_Ar_mc_AKGC417L train 865 | 213_1p5_Pl_mc_AKGC417L train 866 | 213_1p5_Pr_mc_AKGC417L train 867 | 213_1p5_Tc_mc_AKGC417L train 868 | 213_2p2_Al_mc_AKGC417L train 869 | 213_2p2_Ar_mc_AKGC417L train 870 | 213_2p2_Pl_mc_AKGC417L train 871 | 213_2p2_Pr_mc_AKGC417L train 872 | 213_2p2_Tc_mc_AKGC417L train 873 | 214_1b1_Ar_sc_Meditron test 874 | 215_1b2_Ar_sc_Meditron train 875 | 215_1b3_Tc_sc_Meditron train 876 | 216_1b1_Al_sc_Meditron test 877 | 216_1b1_Pl_sc_Meditron test 878 | 217_1b1_Tc_sc_Meditron train 879 | 218_1b1_Al_sc_Meditron test 880 | 218_1b1_Ar_sc_Meditron train 881 | 218_1b1_Lr_sc_Meditron train 882 | 218_1b1_Pl_sc_Meditron test 883 | 218_1b1_Pr_sc_Meditron test 884 | 218_1p1_Ar_sc_Litt3200 train 885 | 218_1p1_Pl_sc_Litt3200 test 886 | 218_1p1_Pr_sc_Litt3200 train 887 | 219_2b1_Ar_mc_LittC2SE train 888 | 219_2b1_Tc_mc_LittC2SE train 889 | 219_2b2_Ar_mc_LittC2SE train 890 | 219_2b2_Tc_mc_LittC2SE train 891 | 219_2b3_Tc_mc_LittC2SE train 892 | 220_1b1_Tc_mc_LittC2SE train 893 | 220_1b2_Al_mc_LittC2SE train 894 | 221_2b1_Al_mc_LittC2SE train 895 | 221_2b1_Ar_mc_LittC2SE train 896 | 221_2b1_Lr_mc_LittC2SE train 897 | 221_2b1_Pl_mc_LittC2SE train 898 | 221_2b2_Al_mc_LittC2SE train 899 | 221_2b2_Ar_mc_LittC2SE train 900 | 221_2b2_Lr_mc_LittC2SE train 901 | 221_2b2_Pl_mc_LittC2SE train 902 | 221_2b3_Al_mc_LittC2SE train 903 | 221_2b3_Ar_mc_LittC2SE train 904 | 221_2b3_Lr_mc_LittC2SE train 905 | 221_2b3_Pr_mc_LittC2SE train 906 | 222_1b1_Ar_sc_Meditron train 907 | 222_1b1_Lr_sc_Meditron train 908 | 222_1b1_Pr_sc_Meditron train 909 | 223_1b1_Pr_sc_Meditron test 910 | 223_1b1_Pl_sc_Meditron test 911 | 223_1b1_Ar_sc_Meditron test 912 | 223_1b1_Al_sc_Meditron test 913 | 223_1b1_Lr_sc_Meditron test 914 | 223_1b1_Ll_sc_Meditron test 915 | 224_1b1_Tc_sc_Meditron test 916 | 224_1b2_Al_sc_Meditron test 917 | 225_1b1_Pl_sc_Meditron test 918 | 226_1b1_Al_sc_Meditron train 919 | 226_1b1_Ll_sc_Meditron train 920 | 226_1b1_Pl_sc_LittC2SE train -------------------------------------------------------------------------------- /data/icbhi_dataset/patient_diagnosis.txt: -------------------------------------------------------------------------------- 1 | 101 URTI 2 | 102 Healthy 3 | 103 Asthma 4 | 104 COPD 5 | 105 URTI 6 | 106 COPD 7 | 107 COPD 8 | 108 LRTI 9 | 109 COPD 10 | 110 COPD 11 | 111 Bronchiectasis 12 | 112 COPD 13 | 113 COPD 14 | 114 COPD 15 | 115 LRTI 16 | 116 Bronchiectasis 17 | 117 COPD 18 | 118 COPD 19 | 119 URTI 20 | 120 COPD 21 | 121 Healthy 22 | 122 Pneumonia 23 | 123 Healthy 24 | 124 COPD 25 | 125 Healthy 26 | 126 Healthy 27 | 127 Healthy 28 | 128 COPD 29 | 129 URTI 30 | 130 COPD 31 | 131 URTI 32 | 132 COPD 33 | 133 COPD 34 | 134 COPD 35 | 135 Pneumonia 36 | 136 Healthy 37 | 137 URTI 38 | 138 COPD 39 | 139 COPD 40 | 140 Pneumonia 41 | 141 COPD 42 | 142 COPD 43 | 143 Healthy 44 | 144 Healthy 45 | 145 COPD 46 | 146 COPD 47 | 147 COPD 48 | 148 URTI 49 | 149 Bronchiolitis 50 | 150 URTI 51 | 151 COPD 52 | 152 Healthy 53 | 153 Healthy 54 | 154 COPD 55 | 155 COPD 56 | 156 COPD 57 | 157 COPD 58 | 158 COPD 59 | 159 Healthy 60 | 160 COPD 61 | 161 Bronchiolitis 62 | 162 COPD 63 | 163 COPD 64 | 164 URTI 65 | 165 URTI 66 | 166 COPD 67 | 167 Bronchiolitis 68 | 168 Bronchiectasis 69 | 169 Bronchiectasis 70 | 170 COPD 71 | 171 Healthy 72 | 172 COPD 73 | 173 Bronchiolitis 74 | 174 COPD 75 | 175 COPD 76 | 176 COPD 77 | 177 COPD 78 | 178 COPD 79 | 179 Healthy 80 | 180 COPD 81 | 181 COPD 82 | 182 Healthy 83 | 183 Healthy 84 | 184 Healthy 85 | 185 COPD 86 | 186 COPD 87 | 187 Healthy 88 | 188 URTI 89 | 189 COPD 90 | 190 URTI 91 | 191 Pneumonia 92 | 192 COPD 93 | 193 COPD 94 | 194 Healthy 95 | 195 COPD 96 | 196 Bronchiectasis 97 | 197 URTI 98 | 198 COPD 99 | 199 COPD 100 | 200 COPD 101 | 201 Bronchiectasis 102 | 202 Healthy 103 | 203 COPD 104 | 204 COPD 105 | 205 COPD 106 | 206 Bronchiolitis 107 | 207 COPD 108 | 208 Healthy 109 | 209 Healthy 110 | 210 URTI 111 | 211 COPD 112 | 212 COPD 113 | 213 COPD 114 | 214 Healthy 115 | 215 Bronchiectasis 116 | 216 Bronchiolitis 117 | 217 Healthy 118 | 218 COPD 119 | 219 Pneumonia 120 | 220 COPD 121 | 221 COPD 122 | 222 COPD 123 | 223 COPD 124 | 224 Healthy 125 | 225 Healthy 126 | 226 Pneumonia -------------------------------------------------------------------------------- /data/icbhi_dataset/patient_list_foldwise.txt: -------------------------------------------------------------------------------- 1 | 158 0 2 | 193 0 3 | 177 0 4 | 170 0 5 | 180 0 6 | 211 0 7 | 147 0 8 | 107 0 9 | 162 0 10 | 156 0 11 | 146 0 12 | 200 0 13 | 138 0 14 | 160 0 15 | 203 0 16 | 204 0 17 | 172 0 18 | 207 0 19 | 163 0 20 | 205 0 21 | 213 0 22 | 114 0 23 | 130 0 24 | 154 0 25 | 186 0 26 | 184 0 27 | 153 1 28 | 115 1 29 | 224 1 30 | 223 1 31 | 201 1 32 | 218 1 33 | 127 1 34 | 137 1 35 | 215 1 36 | 161 1 37 | 206 1 38 | 101 1 39 | 168 1 40 | 131 1 41 | 216 1 42 | 120 1 43 | 188 1 44 | 167 1 45 | 210 1 46 | 197 1 47 | 183 1 48 | 152 1 49 | 173 1 50 | 108 1 51 | 208 1 52 | 105 2 53 | 110 2 54 | 116 2 55 | 196 2 56 | 182 2 57 | 222 2 58 | 166 2 59 | 209 2 60 | 144 2 61 | 111 2 62 | 165 2 63 | 148 2 64 | 164 2 65 | 159 2 66 | 121 2 67 | 157 2 68 | 217 2 69 | 123 2 70 | 169 2 71 | 179 2 72 | 190 2 73 | 125 2 74 | 129 2 75 | 225 2 76 | 136 2 77 | 118 3 78 | 185 3 79 | 112 3 80 | 124 3 81 | 104 3 82 | 195 3 83 | 175 3 84 | 212 3 85 | 140 3 86 | 219 3 87 | 132 3 88 | 142 3 89 | 220 3 90 | 122 3 91 | 191 3 92 | 128 3 93 | 226 3 94 | 141 3 95 | 103 3 96 | 134 3 97 | 117 3 98 | 192 3 99 | 106 3 100 | 155 3 101 | 199 3 102 | 174 4 103 | 145 4 104 | 151 4 105 | 176 4 106 | 178 4 107 | 133 4 108 | 198 4 109 | 214 4 110 | 149 4 111 | 143 4 112 | 187 4 113 | 202 4 114 | 119 4 115 | 194 4 116 | 126 4 117 | 150 4 118 | 171 4 119 | 102 4 120 | 109 4 121 | 113 4 122 | 139 4 123 | 189 4 124 | 181 4 125 | 221 4 126 | 135 4 127 | -------------------------------------------------------------------------------- /main.py: -------------------------------------------------------------------------------- 1 | from copy import deepcopy 2 | import os 3 | import sys 4 | import json 5 | import warnings 6 | warnings.filterwarnings("ignore") 7 | 8 | import math 9 | import time 10 | import random 11 | import pickle 12 | import argparse 13 | import numpy as np 14 | 15 | import logging 16 | logging.disable(logging.INFO) # disable INFO and DEBUG logging everywhere 17 | # or 18 | logging.disable(logging.WARNING) # disable WARNING, INFO and DEBUG logging everywhere 19 | 20 | import torch 21 | import torch.nn as nn 22 | import torch.optim as optim 23 | import torch.backends.cudnn as cudnn 24 | from torchvision import transforms 25 | from transformers import set_seed 26 | 27 | from util.icbhi_dataset import ICBHIDataset 28 | from util.icbhi_util import get_score 29 | from util.misc import adjust_learning_rate, warmup_learning_rate, set_optimizer, update_moving_average 30 | from util.misc import AverageMeter, accuracy, save_model, update_json 31 | 32 | from transformers import ClapProcessor 33 | Processor = ClapProcessor.from_pretrained("laion/clap-htsat-unfused", sampling_rate=48000) 34 | 35 | import os 36 | os.environ["TOKENIZERS_PARALLELISM"] = "false" 37 | #CUBLAS_WORKSPACE_CONFIG=:16:8 38 | 39 | def parse_args(): 40 | parser = argparse.ArgumentParser('argument for supervised training') 41 | 42 | parser.add_argument('--seed', type=int, default=0) 43 | parser.add_argument('--print_freq', type=int, default=10) 44 | parser.add_argument('--save_freq', type=int, default=100) 45 | parser.add_argument('--save_dir', type=str, default='./save') ## have to change 46 | parser.add_argument('--tag', type=str, default='', 47 | help='tag for experiment name') 48 | parser.add_argument('--resume', type=str, default=None, 49 | help='path of model checkpoint to resume') 50 | parser.add_argument('--eval', action='store_true', 51 | help='only evaluation with pretrained encoder and classifier') 52 | parser.add_argument('--two_cls_eval', action='store_true', 53 | help='evaluate with two classes') 54 | 55 | # optimization 56 | parser.add_argument('--optimizer', type=str, default='adam') 57 | parser.add_argument('--epochs', type=int, default=50) 58 | parser.add_argument('--learning_rate', type=float, default=5e-5) 59 | parser.add_argument('--lr_decay_epochs', type=str, default='120,160') 60 | parser.add_argument('--lr_decay_rate', type=float, default=0.1) 61 | parser.add_argument('--weight_decay', type=float, default=1e-4) 62 | parser.add_argument('--momentum', type=float, default=0.9) 63 | parser.add_argument('--cosine', action='store_true', 64 | help='using cosine annealing') 65 | parser.add_argument('--warm', action='store_true', 66 | help='warm-up for large batch training') 67 | parser.add_argument('--warm_epochs', type=int, default=0, 68 | help='warmup epochs') 69 | parser.add_argument('--weighted_loss', action='store_true', 70 | help='weighted cross entropy loss (higher weights on abnormal class)') 71 | # dataset 72 | parser.add_argument('--dataset', type=str, default='icbhi') 73 | parser.add_argument('--data_folder', type=str, default='./data/') 74 | parser.add_argument('--batch_size', type=int, default=128) 75 | parser.add_argument('--num_workers', type=int, default=8) 76 | # icbhi dataset 77 | parser.add_argument('--class_split', type=str, default='lungsound', 78 | help='lungsound: (normal, crackles, wheezes, both), diagnosis: (healthy, chronic diseases, non-chronic diseases)') 79 | parser.add_argument('--n_cls', type=int, default=0, 80 | help='set k-way classification problem for class') 81 | parser.add_argument('--test_fold', type=str, default='official', choices=['official', '0', '1', '2', '3', '4'], 82 | help='test fold to use official 60-40 split or 80-20 split from RespireNet') 83 | parser.add_argument('--sample_rate', type=int, default=16000, 84 | help='sampling rate when load audio data, and it denotes the number of samples per one second') 85 | parser.add_argument('--desired_length', type=int, default=8, 86 | help='fixed length size of individual cycle') 87 | parser.add_argument('--pad_types', type=str, default='repeat', 88 | help='zero: zero-padding, repeat: padding with duplicated samples, aug: padding with augmented samples') 89 | 90 | # model 91 | parser.add_argument('--model', type=str, default='ast') 92 | parser.add_argument('--model_type', type=str, default='ClapAudioModelWithProjection', choices=['ClapAudioModelWithProjection', 'ClapModel']) 93 | parser.add_argument('--test_drop_key', action='store_true') 94 | parser.add_argument('--test_drop_key_prob', type=float, default=0.1) 95 | parser.add_argument('--test_unknown_all', action='store_true') 96 | parser.add_argument('--test_bmi', action='store_true') 97 | parser.add_argument('--test_wrong_label', action='store_true') 98 | parser.add_argument('--analysis', action='store_true') 99 | parser.add_argument('--clap_final', type=str, default='concat', choices=['concat', 'add']) 100 | 101 | # Meta for generate descriptions 102 | parser.add_argument('--meta_mode', type=str, default='none', 103 | help='the meta information for selecting', choices=['age', 'sex', 'loc', 'dev', 'age_sex', 'age_loc', 'age_dev', 104 | 'sex_loc', 'sex_dev', 'loc_dev', 'age_sex_loc', 'age_sex_dev', 'age_loc_dev', 'sex_loc_dev', 'all']) 105 | 106 | parser.add_argument('--pretrained', action='store_true') 107 | parser.add_argument('--pretrained_ckpt', type=str, default=None, 108 | help='path to pre-trained encoder model') 109 | parser.add_argument('--ma_update', action='store_true', 110 | help='whether to use moving average update for model') 111 | parser.add_argument('--ma_beta', type=float, default=0, 112 | help='moving average value') 113 | parser.add_argument('--method', type=str, default='ce') 114 | 115 | 116 | 117 | args = parser.parse_args() 118 | 119 | iterations = args.lr_decay_epochs.split(',') 120 | args.lr_decay_epochs = list([]) 121 | for it in iterations: 122 | args.lr_decay_epochs.append(int(it)) 123 | args.model_name = '{}_{}'.format(args.dataset, args.model) if args.meta_mode == 'none' else '{}_{}_{}'.format(args.dataset, args.model, args.meta_mode) 124 | if args.tag: 125 | args.model_name += '_{}'.format(args.tag) 126 | 127 | args.save_folder = os.path.join(args.save_dir, args.model_name) 128 | os.makedirs(args.save_folder, exist_ok=True) 129 | 130 | if args.warm: 131 | args.warmup_from = args.learning_rate * 0.1 132 | args.warm_epochs = 10 133 | if args.cosine: 134 | eta_min = args.learning_rate * (args.lr_decay_rate ** 3) 135 | args.warmup_to = eta_min + (args.learning_rate - eta_min) * ( 136 | 1 + math.cos(math.pi * args.warm_epochs / args.epochs)) / 2 137 | else: 138 | args.warmup_to = args.learning_rate 139 | 140 | 141 | if args.dataset == 'icbhi': 142 | if args.class_split == 'lungsound': 143 | if args.n_cls == 4: 144 | args.cls_list = ['normal', 'crackle', 'wheeze', 'both'] 145 | elif args.n_cls == 2: 146 | args.cls_list = ['normal', 'abnormal'] 147 | 148 | elif args.class_split == 'diagnosis': 149 | if args.n_cls == 3: 150 | args.cls_list = ['healthy', 'chronic_diseases', 'non-chronic_diseases'] 151 | elif args.n_cls == 2: 152 | args.cls_list = ['healthy', 'unhealthy'] 153 | else: 154 | raise NotImplementedError 155 | 156 | else: 157 | raise NotImplementedError 158 | 159 | 160 | return args 161 | 162 | 163 | def set_loader(args): 164 | if args.dataset == 'icbhi': 165 | 166 | train_dataset = ICBHIDataset(train_flag=True, transform=None, args=args, print_flag=True) 167 | val_dataset = ICBHIDataset(train_flag=False, transform=None, args=args, print_flag=True) 168 | 169 | else: 170 | raise NotImplemented 171 | 172 | train_loader = torch.utils.data.DataLoader(train_dataset, batch_size=args.batch_size, shuffle=True, 173 | num_workers=args.num_workers, pin_memory=True, drop_last=True) 174 | val_loader = torch.utils.data.DataLoader(val_dataset, batch_size=args.batch_size, shuffle=False, 175 | num_workers=args.num_workers, pin_memory=True) 176 | 177 | return train_loader, val_loader, args 178 | 179 | 180 | def set_model(args): 181 | kwargs = {} 182 | 183 | if args.model_type == 'ClapAudioModelWithProjection': 184 | from models.clap import PretrainedCLAPWithProjection 185 | model = PretrainedCLAPWithProjection(args.model, 512) 186 | elif args.model_type == 'ClapModel': 187 | from models.clap import PretrainedCLAP 188 | model = PretrainedCLAP(args.model, 512) 189 | 190 | if args.model_type == 'ClapModel': 191 | classifier = nn.Linear(model.final_feat_dim * 2, args.n_cls) 192 | else: 193 | classifier = nn.Linear(model.final_feat_dim, args.n_cls) 194 | 195 | projector = nn.Identity() 196 | 197 | criterion = nn.CrossEntropyLoss() 198 | 199 | # load SSL pretrained checkpoint for linear evaluation 200 | if args.pretrained and args.pretrained_ckpt is not None: 201 | ckpt = torch.load(args.pretrained_ckpt, map_location='cpu') 202 | state_dict = ckpt['model'] 203 | 204 | # HOTFIX: always use dataparallel during SSL pretraining 205 | new_state_dict = {} 206 | for k, v in state_dict.items(): 207 | if "module." in k: 208 | k = k.replace("module.", "") 209 | if "backbone." in k: 210 | k = k.replace("backbone.", "") 211 | if not 'mlp_head' in k: #del mlp_head 212 | new_state_dict[k] = v 213 | state_dict = new_state_dict 214 | model.load_state_dict(state_dict, strict=False) 215 | 216 | if ckpt.get('classifier', None) is not None: 217 | classifier.load_state_dict(ckpt['classifier'], strict=True) 218 | 219 | print('pretrained model loaded from: {}'.format(args.pretrained_ckpt)) 220 | 221 | criterion = [criterion.cuda()] 222 | 223 | if torch.cuda.device_count() > 1: 224 | model = torch.nn.DataParallel(model) 225 | 226 | model.cuda() 227 | classifier.cuda() 228 | projector.cuda() 229 | 230 | optim_params = list(model.parameters()) + list(classifier.parameters()) + list(projector.parameters()) 231 | optimizer = set_optimizer(args, optim_params) 232 | 233 | return model, classifier, projector, criterion, optimizer 234 | 235 | def train(train_loader, model, classifier, projector, criterion, optimizer, epoch, args, scaler=None): 236 | model.train() 237 | classifier.train() 238 | projector.train() 239 | 240 | batch_time = AverageMeter() 241 | data_time = AverageMeter() 242 | losses = AverageMeter() 243 | top1 = AverageMeter() 244 | 245 | end = time.time() 246 | for idx, (images, labels) in enumerate(train_loader): 247 | # data load 248 | data_time.update(time.time() - end) 249 | images = images.cuda(non_blocking=True) 250 | if args.model_type == 'ClapModel': 251 | class_labels = labels[0].cuda(non_blocking=True) 252 | meta_texts = labels[1].cuda(non_blocking=True) 253 | meta_masks = labels[2].cuda(non_blocking=True) 254 | labels = class_labels 255 | else: 256 | labels = labels.cuda(non_blocking=True) 257 | bsz = labels.shape[0] 258 | 259 | if args.ma_update: 260 | # store the previous iter checkpoint 261 | with torch.no_grad(): 262 | ma_ckpt = [deepcopy(model.state_dict()), deepcopy(classifier.state_dict()), deepcopy(projector.state_dict())] 263 | alpha = None 264 | 265 | warmup_learning_rate(args, epoch, idx, len(train_loader), optimizer) 266 | 267 | with torch.cuda.amp.autocast(): 268 | if args.model_type == 'ClapAudioModelWithProjection': 269 | features = model(images, args=args, alpha=alpha, training=True) 270 | elif args.model_type == 'ClapModel': 271 | features = model((meta_texts, meta_masks, images), args=args, alpha=alpha, training=True) 272 | output = classifier(features) 273 | loss = criterion[0](output, labels) 274 | 275 | losses.update(loss.item(), bsz) 276 | [acc1], _ = accuracy(output[:bsz], labels, topk=(1,)) 277 | top1.update(acc1[0], bsz) 278 | 279 | optimizer.zero_grad() 280 | scaler.scale(loss).backward() 281 | scaler.step(optimizer) 282 | scaler.update() 283 | 284 | # measure elapsed time 285 | batch_time.update(time.time() - end) 286 | end = time.time() 287 | 288 | if args.ma_update: 289 | with torch.no_grad(): 290 | # exponential moving average update 291 | model = update_moving_average(args.ma_beta, model, ma_ckpt[0]) 292 | classifier = update_moving_average(args.ma_beta, classifier, ma_ckpt[1]) 293 | projector = update_moving_average(args.ma_beta, projector, ma_ckpt[2]) 294 | 295 | # print info 296 | if (idx + 1) % args.print_freq == 0: 297 | print('Train: [{0}][{1}/{2}]\t' 298 | 'BT {batch_time.val:.3f} ({batch_time.avg:.3f})\t' 299 | 'DT {data_time.val:.3f} ({data_time.avg:.3f})\t' 300 | 'loss {loss.val:.3f} ({loss.avg:.3f})\t' 301 | 'Acc@1 {top1.val:.3f} ({top1.avg:.3f})'.format( 302 | epoch, idx + 1, len(train_loader), batch_time=batch_time, 303 | data_time=data_time, loss=losses, top1=top1)) 304 | sys.stdout.flush() 305 | 306 | return losses.avg, top1.avg 307 | 308 | def validate(val_loader, model, classifier, criterion, args, best_acc, best_model=None): 309 | save_bool = False 310 | model.eval() 311 | 312 | classifier.eval() 313 | 314 | batch_time = AverageMeter() 315 | losses = AverageMeter() 316 | top1 = AverageMeter() 317 | hits, counts = [0.0] * args.n_cls, [0.0] * args.n_cls 318 | 319 | with torch.no_grad(): 320 | end = time.time() 321 | for idx, (images, labels) in enumerate(val_loader): 322 | 323 | images = images.cuda(non_blocking=True) 324 | if args.model_type == 'ClapModel': 325 | class_labels = labels[0].cuda(non_blocking=True) 326 | meta_texts = labels[1].cuda(non_blocking=True) 327 | meta_masks = labels[2].cuda(non_blocking=True) 328 | labels = class_labels 329 | else: 330 | labels = labels.cuda(non_blocking=True) 331 | 332 | bsz = labels.shape[0] 333 | 334 | with torch.cuda.amp.autocast(): 335 | if args.model in ['laion/clap-htsat-unfused']: 336 | if args.model_type == 'ClapAudioModelWithProjection': 337 | features = model(images, args=args, training=False) 338 | elif args.model_type == 'ClapModel': 339 | features = model((meta_texts, meta_masks, images), args=args, training=False) 340 | else: 341 | features = model(images, args=args, training=False) 342 | 343 | output = classifier(features) 344 | loss = criterion[0](output, labels) 345 | 346 | losses.update(loss.item(), bsz) 347 | [acc1], _ = accuracy(output, labels, topk=(1,)) 348 | top1.update(acc1[0], bsz) 349 | 350 | _, preds = torch.max(output, 1) 351 | for idx in range(preds.shape[0]): 352 | counts[labels[idx].item()] += 1.0 353 | if not args.two_cls_eval: 354 | if preds[idx].item() == labels[idx].item(): 355 | hits[labels[idx].item()] += 1.0 356 | else: # only when args.n_cls == 4 357 | if labels[idx].item() == 0 and preds[idx].item() == labels[idx].item(): 358 | hits[labels[idx].item()] += 1.0 359 | elif labels[idx].item() != 0 and preds[idx].item() > 0: # abnormal 360 | hits[labels[idx].item()] += 1.0 361 | 362 | sp, se, sc = get_score(hits, counts) 363 | 364 | batch_time.update(time.time() - end) 365 | end = time.time() 366 | 367 | if (idx + 1) % args.print_freq == 0: 368 | print('Test: [{0}/{1}]\t' 369 | 'Time {batch_time.val:.3f} ({batch_time.avg:.3f})\t' 370 | 'Loss {loss.val:.4f} ({loss.avg:.4f})\t' 371 | 'Acc@1 {top1.val:.3f} ({top1.avg:.3f})'.format( 372 | idx + 1, len(val_loader), batch_time=batch_time, 373 | loss=losses, top1=top1)) 374 | 375 | if sc > best_acc[-1] and se > 5: 376 | save_bool = True 377 | best_acc = [sp, se, sc] 378 | best_model = [deepcopy(model.state_dict()), deepcopy(classifier.state_dict())] 379 | 380 | print(' * S_p: {:.2f}, S_e: {:.2f}, Score: {:.2f} (Best S_p: {:.2f}, S_e: {:.2f}, Score: {:.2f})'.format(sp, se, sc, best_acc[0], best_acc[1], best_acc[-1])) 381 | print(' * Acc@1 {top1.avg:.2f}'.format(top1=top1)) 382 | print(' * Loss is {loss.avg:.2f} '.format(loss=losses)) 383 | 384 | return best_acc, best_model, save_bool 385 | 386 | 387 | def main(): 388 | args = parse_args() 389 | with open(os.path.join(args.save_folder, 'train_args.json'), 'w') as f: 390 | json.dump(vars(args), f, indent=4) 391 | 392 | #torch.autograd.set_detect_anomaly(True) 393 | 394 | # fix seed 395 | random.seed(args.seed) 396 | np.random.seed(args.seed) 397 | torch.manual_seed(args.seed) 398 | torch.cuda.manual_seed(args.seed) 399 | cudnn.deterministic = True 400 | cudnn.benchmark = True 401 | 402 | #cudnn.benchmark = False 403 | #torch.use_deterministic_algorithms(True, warn_only=True) 404 | set_seed(args.seed) 405 | #os.environ['PYTHONHASHSEED'] = str(args.seed) 406 | 407 | best_model = None 408 | if args.dataset == 'icbhi': 409 | best_acc = [0, 0, 0] # Specificity, Sensitivity, Score 410 | 411 | train_loader, val_loader, args = set_loader(args) 412 | model, classifier, projector, criterion, optimizer = set_model(args) 413 | 414 | print('# of params', sum(p.numel() for p in model.parameters() if p.requires_grad)) 415 | 416 | if args.resume: 417 | if os.path.isfile(args.resume): 418 | print("=> loading checkpoint '{}'".format(args.resume)) 419 | checkpoint = torch.load(args.resume) 420 | args.start_epoch = checkpoint['epoch'] 421 | model.load_state_dict(checkpoint['model']) 422 | optimizer.load_state_dict(checkpoint['optimizer']) 423 | args.start_epoch += 1 424 | print("=> loaded checkpoint '{}' (epoch {})".format(args.resume, checkpoint['epoch'])) 425 | else: 426 | print("=> no checkpoint found at '{}'".format(args.resume)) 427 | else: 428 | args.start_epoch = 1 429 | 430 | # use mix_precision: 431 | scaler = torch.cuda.amp.GradScaler() 432 | 433 | print('*' * 20) 434 | 435 | if not args.eval: 436 | print('Experiments {} start'.format(args.model_name)) 437 | print('Training for {} epochs on {} dataset'.format(args.epochs, args.dataset)) 438 | 439 | for epoch in range(args.start_epoch, args.epochs+1): 440 | adjust_learning_rate(args, optimizer, epoch) 441 | 442 | # train for one epoch 443 | time1 = time.time() 444 | loss, acc = train(train_loader, model, classifier, projector, criterion, optimizer, epoch, args, scaler) 445 | time2 = time.time() 446 | print('Train epoch {}, total time {:.2f}, accuracy:{:.2f}'.format(epoch, time2-time1, acc)) 447 | 448 | # eval for one epoch 449 | best_acc, best_model, save_bool = validate(val_loader, model, classifier, criterion, args, best_acc, best_model) 450 | 451 | # save a checkpoint of model and classifier when the best score is updated 452 | if save_bool: 453 | save_file = os.path.join(args.save_folder, 'best_epoch_{}.pth'.format(epoch)) 454 | print('Best ckpt is modified with Score = {:.2f} when Epoch = {}'.format(best_acc[2], epoch)) 455 | 456 | # save a checkpoint of classifier with the best accuracy or score 457 | save_file = os.path.join(args.save_folder, 'best.pth') 458 | model.load_state_dict(best_model[0]) 459 | classifier.load_state_dict(best_model[1]) 460 | save_model(model, optimizer, args, epoch, save_file, classifier) 461 | else: 462 | print('Testing the pretrained checkpoint on {} dataset'.format(args.dataset)) 463 | best_acc, _, _ = validate(val_loader, model, classifier, criterion, args, best_acc) 464 | 465 | print('{} finished'.format(args.model_name)) 466 | update_json('%s' % args.model_name, best_acc, path=os.path.join(args.save_dir, 'results.json')) 467 | 468 | if __name__ == '__main__': 469 | main() 470 | -------------------------------------------------------------------------------- /models/clap.py: -------------------------------------------------------------------------------- 1 | from transformers import ClapModel, ClapAudioModelWithProjection 2 | import torch 3 | import torch.nn as nn 4 | #from transformers import set_seed 5 | #set_seed(1) 6 | class PretrainedCLAPWithProjection(nn.Module): 7 | def __init__(self, pretrained_name, final_feat_dim): 8 | super().__init__() 9 | 10 | self.pretrained = pretrained_name 11 | self.audio_features = ClapAudioModelWithProjection.from_pretrained(pretrained_name) 12 | self.final_feat_dim = final_feat_dim 13 | 14 | def forward(self, x, args=None, alpha=None, training=False): 15 | x = self.audio_features(x) 16 | return x.audio_embeds 17 | 18 | 19 | 20 | class PretrainedCLAP(nn.Module): 21 | def __init__(self, pretrained_name, final_feat_dim): 22 | super().__init__() 23 | 24 | self.pretrained = pretrained_name 25 | self.audio_features = ClapModel.from_pretrained(pretrained_name) 26 | self.final_feat_dim = final_feat_dim 27 | 28 | def forward(self, x, args=None, alpha=None, training=False): 29 | text_inputs, attention_mask, audio_inputs = x 30 | x = self.audio_features(input_ids=text_inputs, attention_mask=attention_mask, input_features=audio_inputs) 31 | 32 | text_embeds = x.text_embeds 33 | audio_embeds = x.audio_embeds 34 | if args.clap_final == 'concat': 35 | return torch.cat((text_embeds, audio_embeds), dim=-1) 36 | elif args.clap_final == 'add': 37 | return (text_embeds * args.te_alpha) + (audio_embeds * (1 - args.te_alpha)) -------------------------------------------------------------------------------- /requirements.txt: -------------------------------------------------------------------------------- 1 | tqdm 2 | pandas 3 | cmapy 4 | transformers -------------------------------------------------------------------------------- /scripts/eval_bts.sh: -------------------------------------------------------------------------------- 1 | 2 | MODEL="laion/clap-htsat-unfused" 3 | SEED="1" 4 | 5 | for s in $SEED 6 | do 7 | for m in $MODEL 8 | do 9 | TAG="BTS_eval" 10 | CUDA_VISIBLE_DEVICES=0 python main.py --tag $TAG \ 11 | --dataset icbhi \ 12 | --seed $s \ 13 | --class_split lungsound \ 14 | --n_cls 4 \ 15 | --epochs 50 \ 16 | --batch_size 8 \ 17 | --optimizer adam \ 18 | --learning_rate 5e-5 \ 19 | --weight_decay 1e-6 \ 20 | --cosine \ 21 | --sample_rate 48000 \ 22 | --model $m \ 23 | --model_type ClapModel \ 24 | --meta_mode all \ 25 | --test_fold official \ 26 | --pad_types repeat \ 27 | --ma_update \ 28 | --ma_beta 0.5 \ 29 | --method ce \ 30 | --print_freq 100 \ 31 | --eval \ 32 | --pretrained \ 33 | --pretrained_ckpt /home2/jw/workspace/mcl/BTS/save/icbhi_laion/clap-htsat-unfused_ce_all_BTS_bs8_lr5e-5_ep50_seed1_meta_all/best.pth 34 | 35 | done 36 | done 37 | -------------------------------------------------------------------------------- /scripts/icbhi_audio-clap_ce.sh: -------------------------------------------------------------------------------- 1 | 2 | MODEL="laion/clap-htsat-unfused" 3 | SEED="1 2 3 4 5" 4 | 5 | for s in $SEED 6 | do 7 | for m in $MODEL 8 | do 9 | TAG="audio-clap_bs8_lr5e-5_ep50_seed${s}" 10 | CUDA_VISIBLE_DEVICES=0 python main.py --tag $TAG \ 11 | --dataset icbhi \ 12 | --seed $s \ 13 | --class_split lungsound \ 14 | --n_cls 4 \ 15 | --epochs 50 \ 16 | --batch_size 8 \ 17 | --optimizer adam \ 18 | --learning_rate 5e-5 \ 19 | --weight_decay 1e-6 \ 20 | --cosine \ 21 | --sample_rate 48000 \ 22 | --model $m \ 23 | --model_type ClapAudioModelWithProjection \ 24 | --test_fold official \ 25 | --pad_types repeat \ 26 | --ma_update \ 27 | --ma_beta 0.5 \ 28 | --method ce \ 29 | --print_freq 100 30 | 31 | # only for evaluation, add the following arguments 32 | # --eval \ 33 | # --pretrained \ 34 | # --pretrained_ckpt ./save/icbhi_ast_ce_bs8_lr5e-5_ep50_seed1/best.pth 35 | 36 | done 37 | done 38 | -------------------------------------------------------------------------------- /scripts/icbhi_bts_meta_all.sh: -------------------------------------------------------------------------------- 1 | 2 | MODEL="laion/clap-htsat-unfused" 3 | SEED="1 2 3 4 5" 4 | 5 | for s in $SEED 6 | do 7 | for m in $MODEL 8 | do 9 | TAG="BTS_bs8_lr5e-5_ep50_seed${s}_meta_all" 10 | CUDA_VISIBLE_DEVICES=0 python main.py --tag $TAG \ 11 | --dataset icbhi \ 12 | --seed $s \ 13 | --class_split lungsound \ 14 | --n_cls 4 \ 15 | --epochs 50 \ 16 | --batch_size 8 \ 17 | --optimizer adam \ 18 | --learning_rate 5e-5 \ 19 | --weight_decay 1e-6 \ 20 | --cosine \ 21 | --sample_rate 48000 \ 22 | --model $m \ 23 | --model_type ClapModel \ 24 | --meta_mode all \ 25 | --test_fold official \ 26 | --pad_types repeat \ 27 | --ma_update \ 28 | --ma_beta 0.5 \ 29 | --method ce \ 30 | --print_freq 100 31 | 32 | # only for evaluation, add the following arguments 33 | # --eval \ 34 | # --pretrained \ 35 | # --pretrained_ckpt ./save/icbhi_ast_ce_bs8_lr5e-5_ep50_seed1/best.pth 36 | 37 | done 38 | done 39 | -------------------------------------------------------------------------------- /util/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/kaen2891/bts/9a3d6a3bb3057588da1739e5734fef2c13222e42/util/__init__.py -------------------------------------------------------------------------------- /util/icbhi_dataset.py: -------------------------------------------------------------------------------- 1 | from curses import meta 2 | import os 3 | import pickle 4 | import random 5 | import numpy as np 6 | import pandas as pd 7 | from tqdm import tqdm 8 | from glob import glob 9 | import random 10 | 11 | import torch 12 | from torch.utils.data import Dataset 13 | from copy import deepcopy 14 | 15 | from .icbhi_util import get_annotations, get_individual_cycles_torchaudio, get_meta_text_descriptions 16 | from transformers import ClapProcessor 17 | 18 | Processor = ClapProcessor.from_pretrained("laion/clap-htsat-unfused", sampling_rate=48000) 19 | 20 | class ICBHIDataset(Dataset): 21 | def __init__(self, train_flag, transform, args, print_flag=True, mean_std=False): 22 | data_folder = os.path.join(args.data_folder, 'icbhi_dataset/audio_test_data') 23 | folds_file = os.path.join(args.data_folder, 'icbhi_dataset/patient_list_foldwise.txt') 24 | official_folds_file = os.path.join(args.data_folder, 'icbhi_dataset/official_split.txt') 25 | test_fold = args.test_fold 26 | 27 | self.data_folder = data_folder 28 | self.train_flag = train_flag 29 | self.split = 'train' if train_flag else 'test' 30 | self.transform = transform 31 | self.args = args 32 | self.mean_std = mean_std 33 | 34 | # parameters for spectrograms 35 | self.sample_rate = args.sample_rate 36 | 37 | cache_path = './data/training.pt' if self.train_flag else './data/test.pt' 38 | if not os.path.isfile(cache_path): 39 | 40 | 41 | # ========================================================================== 42 | """ get ICBHI dataset meta information """ 43 | # store stethoscope device information for each file or patient 44 | device_id, self.device_to_id = 0, {} 45 | self.device_id_to_patient, self.file_to_device = {}, {} 46 | self.device_to_id = {'Meditron': 0, 'LittC2SE': 1, 'Litt3200': 2, 'AKGC417L': 3} 47 | self.device_id_to_patient = {0: [], 1: [], 2: [], 3: []} 48 | 49 | filenames = sorted(glob(data_folder+'/*')) #--> 1840: 920 wav + 920 txt 50 | filenames = set(f.strip().split('/')[-1].split('.')[0] for f in filenames if '.wav' in f or '.txt' in f) 51 | filenames = sorted(list(set(filenames))) #--> 920 52 | 53 | for f in filenames: 54 | f += '.wav' 55 | # get the total number of devices from original dataset (icbhi dataset has 4 stethoscope devices) 56 | device = f.strip().split('_')[-1].split('.')[0] #-->Meditron / LittC2SE / Litt3200 / AKGC417L 57 | # get the device information for each wav file 58 | self.file_to_device[f.strip().split('.')[0]] = self.device_to_id[device] 59 | 60 | pat_id = f.strip().split('_')[0] # 101 61 | if pat_id not in self.device_id_to_patient[self.device_to_id[device]]: 62 | self.device_id_to_patient[self.device_to_id[device]].append(pat_id) #0: ['101'] 63 | 64 | # store all metadata (age, sex, adult_BMI, child_weight, child_height, device_index) 65 | self.file_to_metadata = {} 66 | meta_file = pd.read_csv(os.path.join(args.data_folder, 'icbhi_dataset/metadata.txt'), names=['age', 'sex', 'adult_BMI', 'child_weight', 'child_height', 'chest_location'], delimiter= '\t') 67 | 68 | meta_file['chest_location'].replace({'Tc':0, 'Al':1, 'Ar':2, 'Pl':3, 'Pr':4, 'Ll':5, 'Lr':6}, inplace=True) # Tc --> 0, Al -> 1, ... 69 | for f in filenames: 70 | pat_idx = int(f.strip().split('_')[0]) 71 | info = list(meta_file.loc[pat_idx]) 72 | info[1] = 0 if info[1] == 'M' else 1 # --> Man:0, Woman:1 73 | 74 | info = np.array(info) 75 | for idx in np.argwhere(np.isnan(info)): 76 | info[idx] = -1 77 | 78 | self.file_to_metadata[f] = torch.tensor(np.append(info, self.file_to_device[f.strip()])) #age, sex, adult_BMI, child_weight, child_height, chest_location, device 79 | 80 | 81 | """ train-test split based on train_flag and test_fold """ 82 | if test_fold in ['0', '1', '2', '3', '4']: # from RespireNet, 80-20% split 83 | 84 | patient_dict = {} 85 | all_patients = open(folds_file).read().splitlines() 86 | for line in all_patients: 87 | idx, fold = line.strip().split(' ') 88 | if train_flag and int(fold) != int(test_fold): 89 | patient_dict[idx] = fold 90 | elif train_flag == False and int(fold) == int(test_fold): 91 | patient_dict[idx] = fold 92 | 93 | if print_flag: 94 | print('*' * 20) 95 | print('Train and test 80-20% split with test_fold {}'.format(test_fold)) 96 | print('Patience number in {} dataset: {}'.format(self.split, len(patient_dict))) 97 | else: 98 | patient_dict = {} 99 | all_fpath = open(official_folds_file).read().splitlines() 100 | 101 | for line in all_fpath: 102 | fpath, fold = line.strip().split('\t') #--> fpath: '101_1b1_Al_sc_Meditron' / fold --> test 103 | if train_flag and fold == 'train': # using for training set 104 | patient_dict[fpath] = fold 105 | elif not train_flag and fold == 'test': # using for test set 106 | patient_dict[fpath] = fold 107 | 108 | if print_flag: 109 | print('*' * 20) 110 | print('Train and test 60-40% split with test_fold {}'.format(test_fold)) 111 | print('File number in {} dataset: {}'.format(self.split, len(patient_dict))) 112 | 113 | annotation_dict = get_annotations(args, data_folder) 114 | 115 | self.filenames = [] 116 | for f in filenames: 117 | # for 'official' test_fold, two patient dataset contain both train and test samples 118 | idx = f.split('_')[0] if test_fold in ['0', '1', '2', '3', '4'] else f 119 | 120 | if idx in patient_dict: 121 | self.filenames.append(f) 122 | 123 | self.audio_data = [] # each sample is a tuple with (audio_data, label, metadata) 124 | 125 | if print_flag: 126 | print('*' * 20) 127 | print("Extracting individual breathing cycles..") 128 | 129 | self.cycle_list = [] 130 | 131 | audio_image_array = [] 132 | label_array = [] 133 | meta_str_array = [] 134 | 135 | for idx, filename in enumerate(self.filenames): 136 | sample_data = get_individual_cycles_torchaudio(args, annotation_dict[filename], data_folder, filename, self.sample_rate, args.n_cls) 137 | for samples in sample_data: 138 | data1, data2 = samples[0], samples[1] # audio_image, label 139 | audio_image_array.append(data1.squeeze(0).numpy()) 140 | label_array.append(data2) 141 | meta_str_array.append(self.file_to_metadata[filename]) 142 | 143 | 144 | cycles_with_labels = [(data[0], data[1]) for data in sample_data] 145 | self.cycle_list.extend(cycles_with_labels) 146 | 147 | audio_image_array = np.array(audio_image_array) 148 | for sample in self.cycle_list: 149 | self.audio_data.append(sample) 150 | 151 | self.class_nums = np.zeros(args.n_cls) 152 | 153 | for sample in self.audio_data: 154 | self.class_nums[sample[1]] += 1 155 | 156 | self.class_ratio = self.class_nums / sum(self.class_nums) * 100 157 | 158 | if print_flag: 159 | print('[Preprocessed {} dataset information]'.format(self.split)) 160 | print('total number of audio data: {}'.format(len(self.audio_data))) 161 | print('*' * 25) 162 | print('For the Label Distribution') 163 | for i, (n, p) in enumerate(zip(self.class_nums, self.class_ratio)): 164 | print('Class {} {:<9}: {:<4} ({:.1f}%)'.format(i, '('+args.cls_list[i]+')', int(n), p)) 165 | 166 | """ convert fbank """ 167 | self.audio_images = [] 168 | inputs = Processor(audios=audio_image_array, return_tensors="pt") 169 | audio_inputs = inputs["input_features"] 170 | for audio, label, meta in zip(audio_inputs, label_array, meta_str_array): 171 | self.audio_images.append((audio, label, meta)) 172 | 173 | if self.train_flag: 174 | torch.save(self.audio_images, './data/training.pt') 175 | else: 176 | torch.save(self.audio_images, './data/test.pt') 177 | 178 | else: 179 | if self.train_flag: 180 | self.audio_images = torch.load('./data/training.pt') 181 | else: 182 | self.audio_images = torch.load('./data/test.pt') 183 | 184 | 185 | def __getitem__(self, index): 186 | audio_image, label, meta_str = self.audio_images[index][0], self.audio_images[index][1], self.audio_images[index][2] 187 | if self.args.model_type in ['ClapModel']: 188 | meta_sentence = get_meta_text_descriptions(meta_str, self.args, training=self.train_flag) 189 | inputs = Processor.tokenizer(text=meta_sentence, return_tensors="pt", padding="max_length", max_length=64) 190 | text_inputs = inputs['input_ids'] 191 | attn_masks = inputs['attention_mask'] 192 | 193 | if self.transform is not None: 194 | audio_image = self.transform(audio_image) 195 | 196 | if self.args.model_type == 'ClapModel': 197 | return audio_image.squeeze(1), (label, text_inputs.squeeze(0), attn_masks.squeeze(0)) 198 | else: 199 | return audio_image.squeeze(1), label 200 | 201 | def __len__(self): 202 | return len(self.audio_images) -------------------------------------------------------------------------------- /util/icbhi_util.py: -------------------------------------------------------------------------------- 1 | from collections import namedtuple 2 | import os 3 | import math 4 | import random 5 | import pandas as pd 6 | import numpy as np 7 | from tqdm import tqdm 8 | from glob import glob 9 | 10 | import torch 11 | import torchaudio 12 | from torchaudio import transforms as T 13 | from .meta_description import generate_meta_description 14 | 15 | __all__ = ['get_annotations', 'get_individual_cycles_torchaudio', 'generate_mel_spectrogram', 'generate_fbank', 'get_score'] 16 | 17 | 18 | # ========================================================================== 19 | """ ICBHI dataset information """ 20 | def _extract_lungsound_annotation(file_name, data_folder): 21 | tokens = file_name.strip().split('_') 22 | recording_info = pd.DataFrame(data = [tokens], columns = ['Patient Number', 'Recording index', 'Chest location','Acquisition mode','Recording equipment']) 23 | recording_annotations = pd.read_csv(os.path.join(data_folder, file_name + '.txt'), names = ['Start', 'End', 'Crackles', 'Wheezes'], delimiter= '\t') 24 | 25 | return recording_info, recording_annotations 26 | 27 | 28 | def get_annotations(args, data_folder): 29 | if args.class_split == 'lungsound' or args.class_split in ['lungsound_meta', 'meta']: # 4-class 30 | filenames = sorted(glob(data_folder+'/*')) #--> 1840: 920 wav + 920 txt 31 | filenames = set(f.strip().split('/')[-1].split('.')[0] for f in filenames if '.txt' in f) 32 | filenames = sorted(list(set(filenames))) #--> 920 33 | 34 | annotation_dict = {} 35 | for f in filenames: 36 | info, ann = _extract_lungsound_annotation(f, data_folder) 37 | annotation_dict[f] = ann 38 | 39 | elif args.class_split == 'diagnosis': 40 | filenames = sorted(glob(data_folder+'/*')) #--> 1840: 920 wav + 920 txt 41 | filenames = set(f.strip().split('/')[-1].split('.')[0] for f in filenames if '.txt' in f) 42 | filenames = sorted(list(set(filenames))) #--> 920 43 | tmp = pd.read_csv(os.path.join(args.data_folder, 'icbhi_dataset/patient_diagnosis.txt'), names=['Disease'], delimiter='\t') 44 | 45 | annotation_dict = {} 46 | for f in filenames: 47 | info, ann = _extract_lungsound_annotation(f, data_folder) 48 | ann.drop(['Crackles', 'Wheezes'], axis=1, inplace=True) 49 | 50 | disease = tmp.loc[int(f.strip().split('_')[0]), 'Disease'] 51 | ann['Disease'] = disease 52 | 53 | annotation_dict[f] = ann 54 | 55 | return annotation_dict 56 | 57 | """ data preprocessing """ 58 | 59 | def _get_lungsound_label(crackle, wheeze, n_cls, args): 60 | if n_cls == 4 or args.method in ['mscl']: 61 | if crackle == 0 and wheeze == 0: 62 | return 0 63 | elif crackle == 1 and wheeze == 0: 64 | return 1 65 | elif crackle == 0 and wheeze == 1: 66 | return 2 67 | elif crackle == 1 and wheeze == 1: 68 | return 3 69 | 70 | elif n_cls == 2: 71 | if crackle == 0 and wheeze == 0: 72 | return 0 73 | else: 74 | return 1 75 | 76 | 77 | def _get_diagnosis_label(disease, n_cls): 78 | if n_cls == 3: 79 | if disease in ['COPD', 'Bronchiectasis', 'Asthma']: 80 | return 1 81 | elif disease in ['URTI', 'LRTI', 'Pneumonia', 'Bronchiolitis']: 82 | return 2 83 | else: 84 | return 0 85 | 86 | elif n_cls == 2: 87 | if disease == 'Healthy': 88 | return 0 89 | else: 90 | return 1 91 | 92 | 93 | def _slice_data_torchaudio(start, end, data, sample_rate): 94 | """ 95 | SCL paper.. 96 | sample_rate denotes how many sample points for one second 97 | """ 98 | max_ind = data.shape[1] 99 | start_ind = min(int(start * sample_rate), max_ind) 100 | end_ind = min(int(end * sample_rate), max_ind) 101 | 102 | return data[:, start_ind: end_ind] 103 | 104 | 105 | def cut_pad_sample_torchaudio(data, args): 106 | fade_samples_ratio = 16 107 | fade_samples = int(args.sample_rate / fade_samples_ratio) 108 | fade_out = T.Fade(fade_in_len=0, fade_out_len=fade_samples, fade_shape='linear') 109 | target_duration = args.desired_length * args.sample_rate 110 | 111 | if data.shape[-1] > target_duration: 112 | data = data[..., :target_duration] 113 | if data.dim() == 1: 114 | data = data.unsqueeze(0) 115 | else: 116 | if args.pad_types == 'zero': 117 | tmp = torch.zeros(1, target_duration, dtype=torch.float32) 118 | diff = target_duration - data.shape[-1] 119 | tmp[..., diff//2:data.shape[-1]+diff//2] = data 120 | data = tmp 121 | elif args.pad_types == 'repeat': 122 | ratio = math.ceil(target_duration / data.shape[-1]) 123 | data = data.repeat(1, ratio) 124 | data = data[..., :target_duration] 125 | data = fade_out(data) 126 | 127 | return data 128 | 129 | def get_meta_text_descriptions(metadata, args, training=False): 130 | age = int(metadata[0]) 131 | sex = int(metadata[1]) 132 | loc = int(metadata[5]) 133 | dev = int(metadata[6]) 134 | 135 | 136 | if age >= 19: 137 | age_str = 'adult' 138 | age_tmp = 0 139 | else: 140 | age_str = 'pediatric' 141 | age_tmp = 1 142 | 143 | if sex == 0: 144 | sex_str = 'male' 145 | else: 146 | sex_str = 'female' 147 | #'Tc':0, 'Al':1, 'Ar':2, 'Pl':3, 'Pr':4, 'Ll':5, 'Lr':6 148 | if loc == 0: 149 | loc_str = 'trachea' 150 | elif loc == 1: 151 | loc_str = 'left anterior chest' 152 | elif loc == 2: 153 | loc_str = 'right anterior chest' 154 | elif loc == 3: 155 | loc_str = 'left posterior chest' 156 | elif loc == 4: 157 | loc_str = 'right posterior chest' 158 | elif loc == 5: 159 | loc_str = 'left lateral chest' 160 | elif loc == 6: 161 | loc_str = 'right lateral chest' 162 | 163 | #Meditron': 0, 'LittC2SE': 1, 'Litt3200': 2, 'AKGC417L': 3 164 | if dev == 0: 165 | dev_str = 'Meditron' 166 | elif dev == 1: 167 | dev_str = 'LittC2SE' 168 | elif dev == 2: 169 | dev_str = 'Litt3200' 170 | elif dev == 3: 171 | dev_str = 'AKGC417L' 172 | 173 | age_str_list = [] 174 | sex_str_list = [] 175 | loc_str_list = [] 176 | dev_str_list = [] 177 | 178 | if args.test_wrong_label: 179 | age_str_list.append('adult') 180 | age_str_list.append('pediatric') 181 | 182 | sex_str_list.append('male') 183 | sex_str_list.append('female') 184 | 185 | loc_str_list.append('trachea') 186 | loc_str_list.append('left anterior chest') 187 | loc_str_list.append('right anterior chest') 188 | loc_str_list.append('left posterior chest') 189 | loc_str_list.append('right posterior chest') 190 | loc_str_list.append('left lateral chest') 191 | loc_str_list.append('right lateral chest') 192 | 193 | dev_str_list.append('Meditron') 194 | dev_str_list.append('LittC2SE') 195 | dev_str_list.append('Litt3200') 196 | dev_str_list.append('AKGC417L') 197 | 198 | ## 199 | age_str_list.pop(age_tmp) 200 | sex_str_list.pop(sex) 201 | loc_str_list.pop(loc) 202 | dev_str_list.pop(dev) 203 | 204 | age_str = random.choice(age_str_list) 205 | sex_str = random.choice(sex_str_list) 206 | loc_str = random.choice(loc_str_list) 207 | dev_str = random.choice(dev_str_list) 208 | 209 | 210 | if training: 211 | if args.meta_mode == 'age': 212 | meta_dict = {"age": age_str, "sex": None, "loc": None, "dev": None} 213 | elif args.meta_mode == 'sex': 214 | meta_dict = {"age": None, "sex": sex_str, "loc": None, "dev": None} 215 | elif args.meta_mode == 'loc': 216 | meta_dict = {"age": None, "sex": None, "loc": loc_str, "dev": None} 217 | elif args.meta_mode == 'dev': 218 | meta_dict = {"age": None, "sex": None, "loc": None, "dev": dev_str} 219 | elif args.meta_mode == 'age_sex': 220 | meta_dict = {"age": age_str, "sex": sex_str, "loc": None, "dev": None} 221 | elif args.meta_mode == 'age_loc': 222 | meta_dict = {"age": age_str, "sex": None, "loc": loc_str, "dev": None} 223 | elif args.meta_mode == 'age_dev': 224 | meta_dict = {"age": age_str, "sex": None, "loc": None, "dev": dev_str} 225 | elif args.meta_mode == 'sex_loc': 226 | meta_dict = {"age": None, "sex": sex_str, "loc": loc_str, "dev": None} 227 | elif args.meta_mode == 'sex_dev': 228 | meta_dict = {"age": None, "sex": sex_str, "loc": None, "dev": dev_str} 229 | elif args.meta_mode == 'loc_dev': 230 | meta_dict = {"age": None, "sex": None, "loc": loc_str, "dev": dev_str} 231 | elif args.meta_mode == 'age_sex_loc': 232 | meta_dict = {"age": age_str, "sex": sex_str, "loc": loc_str, "dev": None} 233 | elif args.meta_mode == 'age_sex_dev': 234 | meta_dict = {"age": age_str, "sex": sex_str, "loc": None, "dev": dev_str} 235 | elif args.meta_mode == 'age_loc_dev': 236 | meta_dict = {"age": age_str, "sex": None, "loc": loc_str, "dev": dev_str} 237 | elif args.meta_mode == 'sex_loc_dev': 238 | meta_dict = {"age": None, "sex": sex_str, "loc": loc_str, "dev": dev_str} 239 | elif args.meta_mode == 'all': 240 | meta_dict = {"age": age_str, "sex": sex_str, "loc": loc_str, "dev": dev_str} 241 | 242 | output_str = generate_meta_description(**meta_dict) 243 | return output_str 244 | 245 | else: 246 | if args.test_drop_key: 247 | drop_key = False 248 | if random.random() < args.test_drop_key_prob: 249 | drop_key = True 250 | 251 | if args.meta_mode == 'age': 252 | meta_dict = {"age": 'unknown' if drop_key else age_str, "sex": None, "loc": None, "dev": None} 253 | elif args.meta_mode == 'sex': 254 | meta_dict = {"age": None, "sex": 'unknown' if drop_key else sex_str, "loc": None, "dev": None} 255 | elif args.meta_mode == 'loc': 256 | meta_dict = {"age": None, "sex": None, "loc": 'unknown' if drop_key else loc_str, "dev": None} 257 | elif args.meta_mode == 'dev': 258 | meta_dict = {"age": None, "sex": None, "loc": None, "dev": 'unknown' if drop_key else dev_str} 259 | 260 | elif args.meta_mode == 'age_sex': 261 | if drop_key: 262 | dice = random.randint(0, 1) 263 | if dice == 0: 264 | age_str = 'unknown' 265 | else: 266 | sex_str = 'unknown' 267 | meta_dict = {"age": age_str, "sex": sex_str, "loc": None, "dev": None} 268 | 269 | elif args.meta_mode == 'age_loc': 270 | if drop_key: 271 | dice = random.randint(0, 1) 272 | if dice == 0: 273 | age_str = 'unknown' 274 | else: 275 | loc_str = 'unknown' 276 | meta_dict = {"age": age_str, "sex": None, "loc": loc_str, "dev": None} 277 | 278 | elif args.meta_mode == 'age_dev': 279 | if drop_key: 280 | dice = random.randint(0, 1) 281 | if dice == 0: 282 | age_str = 'unknown' 283 | else: 284 | dev_str = 'unknown' 285 | meta_dict = {"age": age_str, "sex": None, "loc": None, "dev": dev_str} 286 | 287 | elif args.meta_mode == 'sex_loc': 288 | if drop_key: 289 | dice = random.randint(0, 1) 290 | if dice == 0: 291 | sex_str = 'unknown' 292 | else: 293 | loc_str = 'unknown' 294 | meta_dict = {"age": None, "sex": sex_str, "loc": loc_str, "dev": None} 295 | 296 | elif args.meta_mode == 'sex_dev': 297 | if drop_key: 298 | dice = random.randint(0, 1) 299 | if dice == 0: 300 | sex_str = 'unknown' 301 | else: 302 | dev_str = 'unknown' 303 | meta_dict = {"age": None, "sex": sex_str, "loc": None, "dev": dev_str} 304 | 305 | elif args.meta_mode == 'loc_dev': 306 | if drop_key: 307 | dice = random.randint(0, 1) 308 | if dice == 0: 309 | loc_str = 'unknown' 310 | else: 311 | dev_str = 'unknown' 312 | meta_dict = {"age": None, "sex": None, "loc": loc_str, "dev": dev_str} 313 | 314 | elif args.meta_mode == 'age_sex_loc': 315 | if drop_key: 316 | dice = random.randint(0, 2) 317 | if dice == 0: 318 | age_str = 'unknown' 319 | elif dice == 1: 320 | sex_str = 'unknown' 321 | else: 322 | loc_str = 'unknown' 323 | meta_dict = {"age": age_str, "sex": sex_str, "loc": loc_str, "dev": None} 324 | 325 | elif args.meta_mode == 'age_sex_dev': 326 | if drop_key: 327 | dice = random.randint(0, 2) 328 | if dice == 0: 329 | age_str = 'unknown' 330 | elif dice == 1: 331 | sex_str = 'unknown' 332 | else: 333 | dev_str = 'unknown' 334 | meta_dict = {"age": age_str, "sex": sex_str, "loc": None, "dev": dev_str} 335 | 336 | elif args.meta_mode == 'age_loc_dev': 337 | if drop_key: 338 | dice = random.randint(0, 2) 339 | if dice == 0: 340 | age_str = 'unknown' 341 | elif dice == 1: 342 | loc_str = 'unknown' 343 | else: 344 | dev_str = 'unknown' 345 | meta_dict = {"age": age_str, "sex": None, "loc": loc_str, "dev": dev_str} 346 | 347 | elif args.meta_mode == 'sex_loc_dev': 348 | if drop_key: 349 | dice = random.randint(0, 2) 350 | if dice == 0: 351 | sex_str = 'unknown' 352 | elif dice == 1: 353 | loc_str = 'unknown' 354 | else: 355 | dev_str = 'unknown' 356 | meta_dict = {"age": None, "sex": sex_str, "loc": loc_str, "dev": dev_str} 357 | 358 | elif args.meta_mode == 'all': 359 | if drop_key: 360 | dice = random.randint(0, 3) 361 | if dice == 0: 362 | age_str = 'unknown' 363 | elif dice == 1: 364 | sex_str = 'unknown' 365 | elif dice == 2: 366 | loc_str = 'unknown' 367 | else: 368 | dev_str = 'unknown' 369 | meta_dict = {"age": age_str, "sex": sex_str, "loc": loc_str, "dev": dev_str} 370 | 371 | elif args.test_unknown_all: 372 | return 'No description.' 373 | 374 | else: 375 | if args.meta_mode == 'age': 376 | meta_dict = {"age": age_str, "sex": None, "loc": None, "dev": None} 377 | elif args.meta_mode == 'sex': 378 | meta_dict = {"age": None, "sex": sex_str, "loc": None, "dev": None} 379 | elif args.meta_mode == 'loc': 380 | meta_dict = {"age": None, "sex": None, "loc": loc_str, "dev": None} 381 | elif args.meta_mode == 'dev': 382 | meta_dict = {"age": None, "sex": None, "loc": None, "dev": dev_str} 383 | elif args.meta_mode == 'age_sex': 384 | meta_dict = {"age": age_str, "sex": sex_str, "loc": None, "dev": None} 385 | elif args.meta_mode == 'age_loc': 386 | meta_dict = {"age": age_str, "sex": None, "loc": loc_str, "dev": None} 387 | elif args.meta_mode == 'age_dev': 388 | meta_dict = {"age": age_str, "sex": None, "loc": None, "dev": dev_str} 389 | elif args.meta_mode == 'sex_loc': 390 | meta_dict = {"age": None, "sex": sex_str, "loc": loc_str, "dev": None} 391 | elif args.meta_mode == 'sex_dev': 392 | meta_dict = {"age": None, "sex": sex_str, "loc": None, "dev": dev_str} 393 | elif args.meta_mode == 'loc_dev': 394 | meta_dict = {"age": None, "sex": None, "loc": loc_str, "dev": dev_str} 395 | elif args.meta_mode == 'age_sex_loc': 396 | meta_dict = {"age": age_str, "sex": sex_str, "loc": loc_str, "dev": None} 397 | elif args.meta_mode == 'age_sex_dev': 398 | meta_dict = {"age": age_str, "sex": sex_str, "loc": None, "dev": dev_str} 399 | elif args.meta_mode == 'age_loc_dev': 400 | meta_dict = {"age": age_str, "sex": None, "loc": loc_str, "dev": dev_str} 401 | elif args.meta_mode == 'sex_loc_dev': 402 | meta_dict = {"age": None, "sex": sex_str, "loc": loc_str, "dev": dev_str} 403 | elif args.meta_mode == 'all': 404 | meta_dict = {"age": age_str, "sex": sex_str, "loc": loc_str, "dev": dev_str} 405 | 406 | output_str = generate_meta_description(**meta_dict) 407 | 408 | if args.test_bmi: 409 | adult_bmi = float(metadata[2]) 410 | child_weight = float(metadata[3]) 411 | child_height = float(metadata[4]) 412 | 413 | if adult_bmi == -1.0: 414 | bmi = round(child_weight / ((child_height * 0.01) ** 2), 1) 415 | else: 416 | bmi = adult_bmi 417 | bmi_sentence = ' The BMI of the patient was {}.'.format(bmi) 418 | output_str += bmi_sentence 419 | 420 | return output_str 421 | 422 | 423 | def get_individual_cycles_torchaudio(args, recording_annotations, data_folder, filename, sample_rate, n_cls): 424 | sample_data = [] 425 | fpath = os.path.join(data_folder, filename+'.wav') 426 | data, sr = torchaudio.load(fpath) 427 | 428 | if sr != sample_rate: 429 | resample = T.Resample(sr, sample_rate) 430 | data = resample(data) 431 | 432 | fade_samples_ratio = 16 433 | fade_samples = int(sample_rate / fade_samples_ratio) 434 | fade = T.Fade(fade_in_len=fade_samples, fade_out_len=fade_samples, fade_shape='linear') 435 | data = fade(data) 436 | 437 | for idx in recording_annotations.index: 438 | row = recording_annotations.loc[idx] 439 | 440 | start = row['Start'] # start time (second) 441 | end = row['End'] # end time (second) 442 | audio_chunk = _slice_data_torchaudio(start, end, data, sample_rate) 443 | 444 | if args.class_split == 'lungsound': 445 | crackles = row['Crackles'] 446 | wheezes = row['Wheezes'] 447 | label = _get_lungsound_label(crackles, wheezes, n_cls, args) 448 | #meta_str = get_meta_infor(metadata, args) 449 | 450 | #sample_data.append((audio_chunk, _get_lungsound_label(crackles, wheezes, n_cls, args), meta_str)) 451 | sample_data.append((audio_chunk, _get_lungsound_label(crackles, wheezes, n_cls, args))) 452 | 453 | elif args.class_split == 'diagnosis': 454 | disease = row['Disease'] 455 | sample_data.append((audio_chunk, _get_diagnosis_label(disease, n_cls))) 456 | 457 | padded_sample_data = [] 458 | for data, label in sample_data: 459 | data = cut_pad_sample_torchaudio(data, args) # --> resample to [1, 128000] --> 8 seconds 460 | padded_sample_data.append((data, label)) 461 | 462 | return padded_sample_data 463 | 464 | 465 | def generate_fbank(args, audio, sample_rate, n_mels=128): 466 | """ 467 | use torchaudio library to convert mel fbank for AST model 468 | """ 469 | assert sample_rate == 16000, 'input audio sampling rate must be 16kHz' 470 | fbank = torchaudio.compliance.kaldi.fbank(audio, htk_compat=True, sample_frequency=sample_rate, use_energy=False, window_type='hanning', num_mel_bins=n_mels, dither=0.0, frame_shift=10) 471 | 472 | if args.model in ['ast']: 473 | mean, std = -4.2677393, 4.5689974 474 | else: 475 | mean, std = fbank.mean(), fbank.std() 476 | fbank = (fbank - mean) / (std * 2) # mean / std 477 | fbank = fbank.unsqueeze(-1).numpy() 478 | return fbank 479 | 480 | 481 | # ========================================================================== 482 | 483 | 484 | # ========================================================================== 485 | """ evaluation metric """ 486 | def get_score(hits, counts, pflag=False): 487 | # normal accuracy 488 | sp = hits[0] / (counts[0] + 1e-10) * 100 489 | # abnormal accuracy 490 | se = sum(hits[1:]) / (sum(counts[1:]) + 1e-10) * 100 491 | sc = (sp + se) / 2.0 492 | 493 | if pflag: 494 | # print("************* Metrics ******************") 495 | print("S_p: {}, S_e: {}, Score: {}".format(sp, se, sc)) 496 | 497 | return sp, se, sc 498 | # ========================================================================== 499 | -------------------------------------------------------------------------------- /util/meta_description.py: -------------------------------------------------------------------------------- 1 | from itertools import product 2 | 3 | import pandas as pd 4 | 5 | 6 | def generate_meta_description(age=None, sex=None, loc=None, dev=None): 7 | # set up the prefix 8 | if loc or dev: 9 | output_sent = "This sound was recorded" 10 | else: 11 | output_sent = "This patient is" 12 | 13 | # prepare the main descriptions 14 | if loc: 15 | s_loc = "an unknown region" if loc == "unknown" else f"the {loc}" 16 | s_loc = f"from {s_loc}" 17 | 18 | if age and sex: 19 | if age == "unknown": 20 | s_age_sex = f"a {sex} patient of unknown age" 21 | elif sex == "unknown": 22 | s_age_sex = "an " if age == "adult" else "a " 23 | s_age_sex += f"{age} patient of unknown sex" 24 | elif age == "pediatric": 25 | s_age_sex = f"a {sex} {age} patient" 26 | else: 27 | s_age_sex = f"an {age} {sex} patient" 28 | elif age: 29 | if age == "unknown": 30 | s_age_sex = "a patient of unknown age" 31 | else: 32 | s_age_sex = "an " if age == "adult" else "a " 33 | s_age_sex += f"{age} patient" 34 | elif sex: 35 | if sex == "unknown": 36 | s_age_sex = "a patient of unknown sex" 37 | else: 38 | s_age_sex = f"a {sex} patient" 39 | 40 | if dev: 41 | if dev == "unknown": 42 | s_dev = "an unknown device" 43 | else: 44 | s_dev_prefix = "an" if dev == "AKGC417L" else "a" 45 | s_dev_suffix = "microphone" if dev == "AKGC417L" else "stethoscope" 46 | s_dev = f"{s_dev_prefix} {dev} {s_dev_suffix}" 47 | 48 | # combine the descriptions 49 | if "s_loc" in locals(): 50 | output_sent += f" {s_loc}" 51 | if "s_age_sex" in locals(): 52 | s_age_sex = s_age_sex.replace("of", "with") 53 | 54 | if "s_age_sex" in locals(): 55 | if "s_loc" not in locals() and "s_dev" not in locals(): 56 | output_sent += f" {s_age_sex}" 57 | elif "s_loc" in locals(): 58 | output_sent += f" of {s_age_sex}" 59 | else: 60 | output_sent += f" from {s_age_sex}" 61 | 62 | if "s_age_sex" in locals() and "s_dev" in locals(): 63 | output_sent += f", using {s_dev}" 64 | elif "s_dev" in locals(): 65 | output_sent += f" with {s_dev}" 66 | 67 | output_sent += "." 68 | 69 | return output_sent 70 | 71 | 72 | def generate_clap_meta_description( 73 | age=None, sex=None, loc=None, dev=None, sex_debiasing=True 74 | ): 75 | output_sent = "" 76 | 77 | # prep the descriptions about the patient (age & sex) 78 | if age and sex: 79 | if age == "unknown": 80 | s_patient = f"A {sex} patient of unknown age" 81 | elif sex == "unknown": 82 | s_patient = "An " if age == "adult" else "A " 83 | s_patient += f"{age} patient of unknown sex" 84 | elif age == "pediatric": 85 | s_patient = f"A {sex} {age} patient" 86 | else: 87 | s_patient = f"An {age} {sex} patient" 88 | elif age: 89 | if age == "unknown": 90 | s_patient = "A patient of unknown age" 91 | else: 92 | s_patient = "An " if age == "adult" else "A " 93 | s_patient += f"{age} patient" 94 | elif sex: 95 | if sex == "unknown": 96 | s_patient = "A patient of unknown sex" 97 | else: 98 | s_patient = f"A {sex} patient" 99 | else: 100 | s_patient = "A patient" 101 | 102 | # prep descriptions about the location 103 | if loc: 104 | s_loc = "an unknown region" if loc == "unknown" else f"the {loc}" 105 | s_loc = f"body sounds recorded from {s_loc}" 106 | else: 107 | s_loc = "body sounds recorded" 108 | 109 | if sex_debiasing: 110 | s_loc = f"their {s_loc}" 111 | else: 112 | if sex == "male": 113 | s_loc = f"his {s_loc}" 114 | elif sex == "female": 115 | s_loc = f"her {s_loc}" 116 | else: 117 | s_loc = f"their {s_loc}" 118 | 119 | # prep descriptions about the device 120 | if dev: 121 | if dev == "unknown": 122 | s_dev = "an unknown device" 123 | else: 124 | s_dev_prefix = "an" if dev == "AKGC417L" else "a" 125 | s_dev_suffix = "microphone" if dev == "AKGC417L" else "stethoscope" 126 | s_dev = f"{s_dev_prefix} {dev} {s_dev_suffix}" 127 | 128 | # combine the descriptions 129 | output_sent += f"{s_patient} had {s_loc}" 130 | 131 | if "s_dev" in locals(): 132 | if loc: 133 | output_sent += f", with {s_dev}" 134 | else: 135 | output_sent += f" with {s_dev}" 136 | 137 | output_sent += "." 138 | 139 | return output_sent 140 | 141 | ''' 142 | # set values for each variable 143 | age_values = ["adult", "pediatric", "unknown", None] 144 | sex_values = ["male", "female", "unknown", None] 145 | loc_values = [ 146 | "trachea", 147 | "left anterior chest", 148 | "right anterior chest", 149 | "left posterior chest", 150 | "right posterior chest", 151 | "left lateral chest", 152 | "right lateral chest", 153 | "unknown", 154 | None, 155 | ] 156 | dev_values = ["Meditron", "LittC2SE", "Litt3200", "AKGC417L", "unknown", None] 157 | 158 | # generate all possible combinations 159 | meta_combinations = list(product(age_values, sex_values, loc_values, dev_values)) 160 | # remove combinations with more than 1 "unknown" 161 | meta_combinations = [comb for comb in meta_combinations if comb.count("unknown") <= 1] 162 | meta_combinations = [ 163 | comb for comb in meta_combinations if comb != (None, None, None, None) 164 | ] 165 | 166 | # # generate all meta descriptions 167 | # meta_descriptions = [] 168 | # for comb in meta_combinations: 169 | # meta_dict = {"age": comb[0], "sex": comb[1], "loc": comb[2], "dev": comb[3]} 170 | # meta_description = generate_meta_description(**meta_dict) 171 | # meta_dict["meta_description"] = meta_description 172 | # meta_descriptions.append(meta_dict) 173 | 174 | # meta_descriptions_df = pd.DataFrame(meta_descriptions) 175 | # meta_descriptions_df.to_excel("meta_descriptions.xlsx", index=False) 176 | 177 | # generate all meta descriptions 178 | output_descriptions = [] 179 | for comb in meta_combinations: 180 | meta_dict = {"age": comb[0], "sex": comb[1], "loc": comb[2], "dev": comb[3]} 181 | output_dict = dict(meta_dict) 182 | 183 | # w/o sex debiasing (w/ "his", "her", "their") 184 | meta_description_wo_debiasing = generate_clap_meta_description( 185 | **meta_dict, sex_debiasing=False 186 | ) 187 | output_dict["meta_description_wo_debiasing"] = meta_description_wo_debiasing 188 | 189 | # w/ sex debiasing (only w/ "their") 190 | meta_description = generate_clap_meta_description(**meta_dict) 191 | output_dict["meta_description"] = meta_description 192 | 193 | output_descriptions.append(output_dict) 194 | 195 | output_descriptions_df = pd.DataFrame(output_descriptions) 196 | output_descriptions_df.to_excel("meta_descriptions_clap_style.xlsx", index=False) 197 | ''' 198 | -------------------------------------------------------------------------------- /util/misc.py: -------------------------------------------------------------------------------- 1 | import os 2 | import json 3 | import math 4 | import numpy as np 5 | 6 | import torch 7 | import torch.nn as nn 8 | import torch.optim as optim 9 | 10 | 11 | """ train util """ 12 | def adjust_learning_rate(args, optimizer, epoch): 13 | lr = args.learning_rate 14 | if args.cosine: 15 | eta_min = lr * (args.lr_decay_rate ** 3) 16 | lr = eta_min + (lr - eta_min) * ( 17 | 1 + math.cos(math.pi * epoch / args.epochs)) / 2 18 | else: 19 | steps = np.sum(epoch > np.asarray(args.lr_decay_epochs)) 20 | if steps > 0: 21 | lr = lr * (args.lr_decay_rate ** steps) 22 | 23 | for i, param_group in enumerate(optimizer.param_groups): 24 | param_group['lr'] = lr 25 | 26 | 27 | def adjust_lr_wd(args, optimizer, epoch): 28 | lr = args.learning_rate 29 | if args.cosine: 30 | eta_min = lr * (args.lr_decay_rate ** 3) 31 | lr = eta_min + (lr - eta_min) * ( 32 | 1 + math.cos(math.pi * epoch / args.epochs)) / 2 33 | else: 34 | steps = np.sum(epoch > np.asarray(args.lr_decay_epochs)) 35 | if steps > 0: 36 | lr = lr * (args.lr_decay_rate ** steps) 37 | 38 | wd = args.weight_decay 39 | if args.wd_scheduler: 40 | wd_min = args.weight_decay_end 41 | wd = wd_min + (wd - wd_min) * ( 42 | 1 + math.cos(math.pi * epoch / args.epochs)) / 2 43 | 44 | for i, param_group in enumerate(optimizer.param_groups): 45 | param_group['lr'] = lr 46 | if i == 0: # in case of DINO and ViT, only wd for regularized params 47 | param_group['weight_decay'] = wd 48 | 49 | 50 | def warmup_learning_rate(args, epoch, batch_id, total_batches, optimizer): 51 | if args.warm and epoch <= args.warm_epochs: 52 | p = (batch_id + (epoch - 1) * total_batches) / \ 53 | (args.warm_epochs * total_batches) 54 | lr = args.warmup_from + p * (args.warmup_to - args.warmup_from) 55 | 56 | for param_group in optimizer.param_groups: 57 | param_group['lr'] = lr 58 | 59 | 60 | def set_optimizer(args, optim_params=None, model=None): 61 | if model is not None: 62 | optim_params = model.parameters() 63 | 64 | if args.optimizer == 'sgd': 65 | optimizer = optim.SGD(optim_params, 66 | lr=args.learning_rate, 67 | momentum=args.momentum, 68 | weight_decay=args.weight_decay) 69 | elif args.optimizer == 'adam': 70 | optimizer = optim.Adam(optim_params, 71 | lr=args.learning_rate, 72 | weight_decay=args.weight_decay) 73 | else: 74 | raise NotImplemented 75 | 76 | return optimizer 77 | 78 | 79 | class MA: 80 | def __init__(self, beta): 81 | super().__init__() 82 | self.beta = beta 83 | 84 | def update_average(self, old, new): 85 | if old is None: 86 | return new 87 | return old * self.beta + (1 - self.beta) * new 88 | 89 | 90 | def update_moving_average(ma_beta, current_model, ma_ckpt): 91 | ma_updater = MA(ma_beta) 92 | new_state_dict = {} 93 | for (k1, current_params), (k2, ma_params) in zip(current_model.state_dict().items(), ma_ckpt.items()): 94 | assert k1 == k2 95 | old_weight, up_weight = ma_params.data, current_params.data 96 | new_state_dict[k1] = ma_updater.update_average(old_weight, up_weight) 97 | 98 | current_model.load_state_dict(new_state_dict) 99 | return current_model 100 | 101 | 102 | """ eval util """ 103 | class AverageMeter(object): 104 | """ Computes and stores the average and current value """ 105 | def __init__(self): 106 | self.reset() 107 | 108 | def reset(self): 109 | self.val = 0 110 | self.avg = 0 111 | self.sum = 0 112 | self.count = 0 113 | 114 | def update(self, val, n=1): 115 | self.val = val 116 | self.sum += val * n 117 | self.count += n 118 | self.avg = self.sum / self.count 119 | 120 | 121 | def accuracy(output, target, topk=(1,)): 122 | """Computes the accuracy over the k top predictions for the specified values of k""" 123 | with torch.no_grad(): 124 | n_cls = output.shape[1] 125 | valid_topk = [k for k in topk if k <= n_cls] 126 | 127 | maxk = max(valid_topk) 128 | bsz = target.size(0) 129 | 130 | _, pred = output.topk(maxk, 1, True, True) 131 | pred = pred.t() 132 | correct = pred.eq(target.view(1, -1).expand_as(pred)) 133 | 134 | res = [] 135 | for k in topk: 136 | if k in valid_topk: 137 | correct_k = correct[:k].reshape(-1).float().sum(0, keepdim=True) 138 | res.append(correct_k.mul_(100.0 / bsz)) 139 | else: res.append(torch.tensor([0.])) 140 | 141 | return res, bsz 142 | 143 | def save_model(model, optimizer, args, epoch, save_file, classifier): 144 | print('==> Saving...') 145 | state = { 146 | 'args': args, 147 | 'model': model.state_dict(), 148 | 'optimizer': optimizer.state_dict(), 149 | 'epoch': epoch, 150 | 'classifier': classifier.state_dict() 151 | } 152 | 153 | torch.save(state, save_file) 154 | del state 155 | 156 | 157 | def update_json(exp_name, acc, path='./save/results.json'): 158 | acc = [round(a, 2) for a in acc] 159 | if not os.path.exists(path): 160 | with open(path, 'w') as f: 161 | json.dump({}, f) 162 | 163 | with open(path, 'r', encoding="UTF-8") as f: 164 | result_dict = json.load(f) 165 | result_dict[exp_name] = acc 166 | 167 | with open(path, 'w') as f: 168 | json.dump(result_dict, f) 169 | 170 | print('best Score: {} (sp, se, sc)'.format(acc)) 171 | print('results updated to %s' % path) --------------------------------------------------------------------------------