Skip to content

Latest commit

 

History

History
 
 

PLG-MT_run

README.md

Couple Learning(PLG-MT)

Steps of training

  • open the config.py file to confirm parameter of mode is "".
  • execute python MT_train.py, Complete the training of the baseline model.
  • execute python PLG_process.py, Complete the generation of pseudo labels
  • open the config.py file to confirm parameter of mode is "_TbS".
  • execute python MT_train.py -sm "_TbS", Complete the training of the PLG-MT model.

Testing

mean teacher only

python TestModel.py -m "stored_data/MeanTeacher/model/baseline_best" -g ../dataset/public_eval/metadata/eval/public.tsv  \
-ga ../dataset/public_eval/audio/eval/public -s stored_data/PLG_MT_public_test/public_predictions.tsv

PLG-MT model

python TestModel.py -m "stored_data/MeanTeacher_TbS/model/baseline_best" -g ../dataset/public_eval/metadata/eval/public.tsv  \
-ga ../dataset/public_eval/audio/eval/public -s stored_data/mean_teacher_public_test/public_predictions.tsv

Results in paper

baseline model path : stored_data/models/trained_baseline_model

PLG-MT model path : stored_data/models/trained_PLG-MT_model

To reproduce the results of the paper run

python TestModel.py -m "stored_data/models/trained_PLG-MT_model" -g ../dataset/public_eval/metadata/eval/public.tsv  \
-ga ../dataset/public_eval/audio/eval/public -s stored_data/mean_teacher_public_test/public_predictions.tsv

Test result

public set

EB-F1 PSDS_1 PSDS_2 PSDS_3
baseline 37.12% 0.5900 0.5070 0.4209
PLG-MT 39.41% 0.6176 0.5552 0.4609

validation set

EB-F1 PSDS_1 PSDS_2 PSDS_3
baseline 32.39% 0.5831 0.4916 0.4098
PLG-MT 33.37% 0.6116 0.5420 0.4398

Supplementary experimental results

EB-F1(%) results with different methods on different test set

Validation public eval
CRNN 28.14 33.56
+ Mean Teacher(Baseline) 32.39 37.12
+ PLG 30.04 35.26
+ Mean Teacher + PLG 33.93 39.18

EB-F1(%) with different pseudo-labels and PLG models on validation set

FBCRNN MT-CRNN
Baseline 32.39 32.39
+ UPW 33.95 30.06
+ WPS 35.80 32.15
+ UPW 37.24 32.42
+ UPS + WPS 37.97 33.52
+ UPS + WPS + UPW 41.33 33.93

References

  • [1] L. Delphin-Poulat & C. Plapous, technical report, dcase 2019.
  • [2] Tarvainen, A. and Valpola, H., 2017. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. In Advances in neural information processing systems (pp. 1195-1204).