【模型整体的结构1】2021-04-28

#五个部分

结构图

#1模型参数（超参数）

data

{'name': 'aishell', 'vocab': 'egs/aishell/data/text2.txt',

'batch_size': 8, 'dataset_type': 'online', 'model_unit': 'chars', 'train': {'feat': ['egs/aishell/data/train/wav.scp'], 'text': ['egs/aishell/data/train/character.txt']}, 'test': {'feat': ['egs/aishell/data/test/wav.scp'], 'text': ['egs/aishell/data/test/character.txt']}, 'num_mel_bins': 40, 'normalization': True, 'spec_augment': True, 'speed_perturb': False, 'volume_perturb': False, 'gaussian_noise': 0.0, 'num_workers': 2, 'spec_augment_config': {'freq_mask_num': 2, 'time_mask_num': 5, 'freq_mask_rate': 0.3, 'time_mask_rate': 0.05}}

model

{'type': 'speech2text', 'frontend_type': 'conv',

'frontend': {'input_size': 40,

'output_size': 256,

'in_channel': 1,

'mid_channel': 64,

'out_channel': 128,

'kernel_size': [[3, 3], [3, 3]],

'stride': [2, 2],

'dropout': 0.0,

'act_func_type': 'relu',

 'front_end_layer_norm': False},

'encoder_type': 'transformer',

'encoder': {'d_model': 256, 'n_heads': 4, 'd_ff': 1024, 'n_blocks': 4, 'pos_dropout': 0.0, 'slf_attn_dropout': 0.0, 'ffn_dropout': 0.0, 'residual_dropout': 0.1, 'normalize_before': False, 'concat_after': False, 'activation': 'glu', 'relative_positional': False},

'decoder_type': 'transformer', 'decoder': {'vocab_size': 4233, 'd_model': 256, 'n_heads': 4, 'd_ff': 1024, 'memory_dim': 256, 'n_blocks': 4, 'pos_dropout': 0.0, 'slf_attn_dropout': 0.0, 'src_attn_dropout': 0.0, 'ffn_dropout': 0.0, 'residual_dropout': 0.1, 'activation': 'glu', 'normalize_before': False, 'concat_after': False, 'share_embedding': True},

'ctc_weight': 0.1, 'encoder_output_size': 256, 'smoothing': 0.1}

train

{'optimizer_type': 'adam', 'optimizer': {'lr': 0.001, 'betas': [0.9, 0.98], 'eps': 1e-09, 'weight_decay': 1e-06, 'amsgrad': False}, 'scheduler_type': 'transformer', 'scheduler': {'model_size': 256, 'warmup_steps': 12000, 'factor': 1.0}, 'clip_grad': 5, 'epochs': 200, 'accum_steps': 4, 'grad_noise': 0.0, 'load_model': False, 'save_name': 'transformer_baseline'}

#2