Train classification models with bug fixes

[auroracramer]

US8K

• L3 embedding / linear model / music set
• L3 embedding / melspec1 / music set
• L3 embedding / melspec2 / music set
• L3 embedding / linear model / env set
• L3 embedding / melspec1 / env set
• L3 embedding / melspec2 / env set

ESC-50

• L3 embedding / linear model / music set
• L3 embedding / melspec1 / music set
• L3 embedding / melspec2 / music set
• L3 embedding / linear model / env set
• L3 embedding / melspec1 / env set
• L3 embedding / melspec2 / env set

DCASE 2013

• L3 embedding / linear model / music set
• L3 embedding / melspec1 / music set
• L3 embedding / melspec2 / music set
• L3 embedding / linear model / env set
• L3 embedding / melspec1 / env set
• L3 embedding / melspec2 / env set

Instructions

To train the classifier, make edit or make a copy of jobs/classifier-train-array.sbatch In there, you'll want to make the following changes:

1. Change the email to your email
2. Change the anaconda environment name
3. Change SRCDIR to whatever directory the l3embedding repository is in
4. Change FEATURES_DIR to the directory where the features are
5. Change OUTPUT_DIR to your general experiment output directory. A subdirectory called classifier will be created here for the classifier output.
6. Change MODEL_DIR to the type of model you want to train. This should remain mlp
7. Change FEATURE_MODE to the type of feature preprocessing you want to perform. Options are framewise or stats. This should remain framewise
8. Change GOOGLE_DEV_APP_NAME and GSHEET_ID if you want to sync the results with a Google Docs spreadsheet.
9. Add --non-overlap as an argument if you want to do a run with no overlapping frames. But this shouldn't be necessary for this set of experiments.
10. Add --parameter-search-no-valid-fold to use 4 folds for training, and then perform a 85-15 split to get a validation set. This is present in the template. This should remain set.
11. Add --parameter-search-valid-ratio <float> to change the validation ratio, when --parameter-search-no-valid-fold is used. This is present in the template. This should remain set
12. Add any additional command line arguments to the script call. Take a look at 06_train_classifier.py for the options. This will not be necessary for this set of experiments though. This is only mentioned for posterity.
13. For US8K/ESC50: Run sbatch --array=1-<NUM_FOLDS> classifier-train-array.sbatch

• <NUM_FOLDS> is 10 for US8K, 5 for ESC-50.

14. For DCASE2013 run sbatch --array=2-2 classifier-train-array.sbatch.
15. Repeat 13. or 14. (depending on the dataset) 2 times for US8K, 5 times for ESC-50, and once for DCASE (though we should check for DCASE)