Armak/SED

1training:2  #batch size: [strong, weak, unlabel]3  batch_size: [24, 24, 32]4  batch_size_val: 645  n_epochs_warmup: 50 # num epochs used for exponential warmup6  num_workers: 8 # change according to your cpu7  n_epochs: 150 # max num epochs8  early_stop_patience: 150 # Same as number of epochs by default, so no early stopping used9  accumulate_batches: 110  gradient_clip: 0. # 0 no gradient clipping11  median_window: 7 # length of median filter used to smooth prediction in inference (nb of output frames)12  val_thresholds: [0.5] # thresholds used to compute f1 intersection in validation.13  n_test_thresholds: 50 # number of thresholds used to compute psds in test14  backend: dp # pytorch lightning backend, ddp, dp or None15  validation_interval: 1 # perform validation every X epoch, 1 default16  seed: 4217  deterministic: False18  precision: 3219  mixup: soft # Soft mixup gives the ratio of the mix to the labels, hard mixup gives a 1 to every label present.20  obj_metric_strong_type: intersection21  enable_progress_bar: True22  weak_split: 0.9 # split weak data in train and validation23  consistency_loss: mse24  ema_factor: 0.99925  const_max: 226scaler:27  statistic: instance # instance or dataset-wide statistic28  normtype: minmax # minmax or standard or mean normalization29  dims: [1, 2] # dimensions over which normalization is applied30  savepath: ./scaler.ckpt # path to scaler checkpoint31data: # change with your paths if different.32  # NOTE: if you have data in 44kHz only then synth_folder will be the path where33  # resampled data will be placed.34  synth_folder: "/mnt/d/DESED_dataset/dcase_synth/audio/train/synthetic21_train/soundscapes_16k/"35  synth_tsv:  "/mnt/d/DESED_dataset/dcase_synth/metadata/train/synthetic21_train/soundscapes.tsv"36  strong_folder: "/mnt/d/DESED_dataset/audio/train/strong_label_real_16k/"37  strong_tsv: "/mnt/d/DESED_dataset/metadata/train/audioset_strong.tsv"38  weak_folder: "/mnt/d/DESED_dataset/audio/train/weak_16k/"39  weak_tsv: "/mnt/d/DESED_dataset/metadata/train/weak.tsv"40  unlabeled_folder: "/mnt/d/DESED_dataset/audio/train/unlabel_in_domain_16k/"41  unlabeled_tsv: "/mnt/d/DESED_dataset/metadata/train/unlabel_in_domain.tsv"42  # synth_val_folder: "/mnt/d/DESED_dataset/dcase_synth/audio/validation/synthetic21_validation/soundscapes_16k/"43  # synth_val_folder_44k: "/mnt/d/DESED_dataset/dcase_synth/audio/validation/synthetic21_validation/soundscapes/"44  # synth_val_tsv:  "/mnt/d/DESED_dataset/dcase_synth/metadata/validation/synthetic21_validation/soundscapes.tsv"45  # synth_val_dur: "/mnt/d/DESED_dataset/dcase_synth/metadata/validation/synthetic21_validation/durations.tsv"46  synth_val_folder: "/mnt/d/DESED_dataset/audio/validation/validation_16k/"47  synth_val_tsv:  "/mnt/d/DESED_dataset/metadata/validation/validation.tsv"48  synth_val_dur: "/mnt/d/DESED_dataset/metadata/validation/validation_durations.tsv"49  # test_folder: "/mnt/d/DESED_dataset/audio/validation/validation_16k/"50  # test_tsv: "/mnt/d/DESED_dataset/metadata/validation/validation.tsv"51  # test_dur: "/mnt/d/DESED_dataset/metadata/validation/validation_durations.tsv"52  test_folder: "/mnt/d/DESED_dataset/audio/test/public/"53  test_tsv: "/mnt/d/DESED_dataset/metadata/test/public.tsv"54  test_dur: "/mnt/d/DESED_dataset/metadata/test/test_durations.tsv"55  eval_folder: "/mnt/d/DESED_dataset/audio/eval21_16k"56  audio_max_len: 1057  fs: 1600058  net_subsample: 159opt:60  lr: 0.00161feats:62  n_mels: 12863  n_filters: 204864  hop_length: 25665  n_window: 204866  sample_rate: 1600067  f_min: 068  f_max: 800069net:70  dropout: 0.571  n_layer_RNN: 272  n_in_channel: 173  nclass: 1074  n_RNN_cell: 12875  activation: Relu76  rnn_type: BGRU77  kernel_size: [3, 3, 3, 3, 3, 3, 3]78  padding: [1, 1, 1, 1, 1, 1, 1]79  stride: [1, 1, 1, 1, 1, 1, 1]80  nb_filters: [ 16, 32, 64, 128, 128, 128, 128 ]81  pooling: [ [ 1, 2 ], [ 1, 2 ], [ 1, 2 ], [ 1, 2 ], [ 1, 2 ], [ 1, 2 ], [ 1, 2 ] ]82  dropout_recurrent: 083  attention: True84 85