Key Word Spotting

Created on November 11|Last edited on November 13
Comment
﻿
Train initial model:path: saved/models/kws_sheila/1110_220958/model_acc_0.97_epoch_90.pth
epoch: 100
'mean_val_loss': 0.086
'mean_val_acc':  0.968, 
'mean_val_FA':    0.015, 
'mean_val_FR':   0.016, 
'au_fa_fr':            0.0003
Test model:
Num params:              564080
'val_time_inference': 0.0026 sec
Model size in mb:      2.152
config:
    config = {
        'verbosity': 2,
        'name': "train",
        'log_step': 50,
        'exper_name': f"kws_{key_word}",
        'key_word': key_word,
        'batch_size': 256,
        'len_epoch': 200,
        'learning_rate': 3e-4,
        'weight_decay': 1e-5,
        'num_epochs': 100,
        'n_mels': 40,  # number of mels for melspectrogram
        'kernel_size': (20, 5),  # size of kernel for convolution layer in CRNN
        'stride': (8, 2),  # size of stride for convolution layer in CRNN
        'hidden_size': 128,  # size of hidden representation in GRU
        'gru_num_layers': 2,  # number of GRU layers in CRNN
        'gru_num_dirs': 2,  # number of directions in GRU (2 if bidirectional)
        'num_classes': 2,  # number of classes (2 for "no word" or "sheila is in audio")
        'sample_rate': 16000,
        'device': device.__str__()
    }
Columns [train | val | other]
﻿
Run: treasured-universe-161
﻿
﻿
Add a comment