train_hyp.yaml

# Generated 2023-02-12 from:
# /home/agorin/cssl_sound/hparams/ecapa_vgg.yaml
# yamllint disable
# File              : supclr_train.yaml
# Author            : Zhepei Wang <zhepeiw2@illinois.edu>
# Date              : 27.01.2022
# Last Modified Date: 31.03.2022
# Last Modified By  : Zhepei Wang <zhepeiw2@illinois.edu>


seed: 2022
__set_seed: !apply:torch.manual_seed [2022]
np_rng: !new:numpy.random.RandomState [2022]

resume_interrupt: false
resume_task_idx: 0
balanced_cry: false

time_stamp: 2023-02-12+21-11-02
experiment_name: ecapa_vgg
# output_folder: !ref results/<experiment_name>/<seed>
output_base: results #/home/agorin/datasets/results_cssl
output_folder: results/2023-02-12+21-11-02_seed_2022+ecapa_vgg
train_log: results/2023-02-12+21-11-02_seed_2022+ecapa_vgg/train_log.txt
save_folder: results/2023-02-12+21-11-02_seed_2022+ecapa_vgg/save

# Number of classes
n_classes: 308
num_tasks: 1
# cont learning setup
task_classes: &id001 !apply:utils.prepare_task_classes
  num_classes: 308
  num_tasks: 1
  seed: 2022
replay_num_keep: 0

use_mixup: false
mixup_alpha: 0.4
train_duration: 4.0

# Training parameters
number_of_epochs: 50
batch_size: 128
# lr: 0.001
# base_lr: 0.00000001
# max_lr: !ref <lr>
# step_size: 65000
warmup_epochs: 5
warmup_lr: 0.0
base_lr: 0.015
final_lr: 5e-09

# dataset
sample_rate: 16000

data_folder: /home/agorin/datasets/VGG-Sound
label_encoder_path: ./dataset/label_encoder_vggsound_ordered.txt
prepare_split_csv_fn: !name:dataset.prepare_vggsound2.prepare_split
  root_dir: /home/agorin/datasets/VGG-Sound
  output_dir: results/2023-02-12+21-11-02_seed_2022+ecapa_vgg/save
  task_classes: *id001
  train_split: 0.8
  seed: 2022

train_dataloader_opts:
  batch_size: 128
  num_workers: 8
  shuffle: true
  drop_last: true


valid_dataloader_opts:
  batch_size: 32
  num_workers: 8



# Experiment params
auto_mix_prec: false # True # True # Set it to True for mixed precision


# Feature parameters
n_mels: 80
left_frames: 0
right_frames: 0
deltas: false
amp_to_db: false
normalize: true
win_length: 25
hop_length: 10
n_fft: 400
f_min: 0
use_time_roll: false
use_freq_shift: false
emb_dim: 256
emb_norm_type: bn
proj_norm_type: bn

# augmentation
# time_domain_aug: !new:speechbrain.lobes.augment.TimeDomainSpecAugment
#     sample_rate: !ref <sample_rate>
#     # drop_chunk_count_high: 2
#     # drop_chunk_noise_factor: 0.05
#     speeds: [90, 95, 100, 105, 110]
#     drop_freq_count_high: 4
#     drop_chunk_count_high: 3
#     # drop_chunk_length_low: 1000
#     # drop_chunk_length_high: 5000
spec_domain_aug: !new:augmentation.TFAugmentation
  time_warp: true
  time_warp_window: 8
  freq_mask: true
  freq_mask_width: !tuple (0, 10)
  n_freq_mask: 2
  time_mask: true
  time_mask_width: !tuple (0, 10)
  n_time_mask: 2
  replace_with_zero: true
  time_roll: false
  time_roll_limit: !tuple (0, 200)
  freq_shift: false
  freq_shift_limit: !tuple (-10, 10)


# Functions
compute_features: &id002 !new:speechbrain.lobes.features.Fbank
  n_mels: 80
  left_frames: 0
  right_frames: 0
  deltas: false
  sample_rate: 16000
  n_fft: 400
  win_length: 25
  hop_length: 10
  f_min: 0

mean_var_norm: &id007 !new:speechbrain.processing.features.InputNormalization

  norm_type: sentence
  std_norm: false

embedding_model: &id003 !new:speechbrain.lobes.models.ECAPA_TDNN.ECAPA_TDNN
  input_size: 80
  channels: [1024, 1024, 1024, 1024, 3072]
  kernel_sizes: [5, 3, 3, 3, 1]
  dilations: [1, 2, 3, 4, 1]
  groups: [1, 1, 1, 1, 1]
  attention_channels: 128
  lin_neurons: 256

# embedding_model: !new:models.pann.Cnn14
# mel_bins: !ref <n_mels>
# emb_dim: !ref <emb_dim>
# norm_type: !ref <emb_norm_type>

projector: &id005 !new:models.modules.SimSiamProjector
  input_size: 256
  hidden_size: 256
  output_size: 256
  norm_type: bn

predictor: &id006 !new:models.modules.SimSiamPredictor
  input_size: 256
  hidden_size: 128
  norm_type: bn

classifier: &id004 !new:models.modules.Classifier
  input_size: 256
  output_size: 308

modules:
  compute_features: *id002
  embedding_model: *id003
  classifier: *id004
  projector: *id005
  predictor: *id006
  mean_var_norm: *id007
ssl_weight: 1.
compute_simclr_cost: !new:losses.SimCLRLoss
  tau: 0.5

sup_weight: 0.
compute_sup_cost: !new:losses.LogSoftmaxWithProbWrapper
  loss_fn: !new:torch.nn.Identity

dist_weight: 0
compute_dist_cost: !new:losses.SimCLRLoss
  tau: 0.5


acc_metric: !name:speechbrain.utils.Accuracy.AccuracyStats

# opt_class: !name:torch.optim.Adam
#     lr: !ref <base_lr>
#     weight_decay: 0.0005
#
# lr_scheduler_fn: !name:speechbrain.nnet.schedulers.CyclicLRScheduler
#     base_lr: !ref <final_lr>
#     max_lr: !ref <base_lr>
#     step_size: 888

opt_class: !name:torch.optim.SGD
  lr: 0.015
  weight_decay: 0.0005
  momentum: 0.9

lr_scheduler_fn: !name:schedulers.SimSiamCosineScheduler
  warmup_epochs: 5
  warmup_lr: 0.0
  num_epochs: 50
  base_lr: 0.015
  final_lr: 5e-09
  steps_per_epoch: 200
  constant_predictor_lr: true

epoch_counter_fn: !name:speechbrain.utils.epoch_loop.EpochCounter
  limit: 50

datapoint_counter: &id008 !new:utils.DatapointCounter

#prev_checkpointer: null
#prev_checkpointer: !new:speechbrain.utils.checkpoints.Checkpointer
#     checkpoints_dir: /home/agorin/vgg_offline/2022-04-13+23-33-21_seed_2022+ssl_offline/save/task0
# # Logging + checkpoints
checkpointer: !new:speechbrain.utils.checkpoints.Checkpointer
  checkpoints_dir: results/2023-02-12+21-11-02_seed_2022+ecapa_vgg/save

recoverables:
  embedding_model: *id003
  classifier: *id004
  projector: *id005
  predictor: *id006
  normalizer: *id007
  datapoint_counter: *id008
ssl_checkpoints_dir:      # /home/agorin/vgg_offline/2022-04-13+23-33-21_seed_2022+ssl_offline/save

train_logger: !new:speechbrain.utils.train_logger.FileTrainLogger
  save_file: results/2023-02-12+21-11-02_seed_2022+ecapa_vgg/train_log.txt

# wandb
use_wandb: false
train_log_frequency: 20
wandb_logger_fn: !name:utils.MyWandBLogger
  initializer: !name:wandb.init
  entity: CAL
  project: cssl_sound
  name: 2023-02-12+21-11-02+seed_2022+ecapa_vgg
  dir: results/2023-02-12+21-11-02_seed_2022+ecapa_vgg
  reinit: true
  yaml_config: hparams/vgg/supclr_train.yaml
  resume: false