训练代码？

[bank010]

这个是lifeiteng/vall-e的模型参数：

VALLE( (ar_text_embedding): TokenEmbedding( (dropout): Dropout(p=0.0, inplace=False) (word_embeddings): Embedding(512, 1024) ) (nar_text_embedding): TokenEmbedding( (dropout): Dropout(p=0.0, inplace=False) (word_embeddings): Embedding(512, 1024) ) (ar_audio_embedding): TokenEmbedding( (dropout): Dropout(p=0.0, inplace=False) (word_embeddings): Embedding(1025, 1024) ) (ar_text_prenet): Identity() (ar_audio_prenet): Identity() (ar_text_position): SinePositionalEmbedding( (dropout): Dropout(p=0.1, inplace=False) ) (ar_audio_position): SinePositionalEmbedding( (dropout): Dropout(p=0.1, inplace=False) ) (ar_decoder): TransformerEncoder( (layers): ModuleList( (0-11): 12 x TransformerEncoderLayer( (self_attn): MultiheadAttention( (out_proj): NonDynamicallyQuantizableLinear(in_features=1024, out_features=1024, bias=True) ) (linear1): Linear(in_features=1024, out_features=4096, bias=True) (dropout): Dropout(p=0.1, inplace=False) (linear2): Linear(in_features=4096, out_features=1024, bias=True) (dropout1): Dropout(p=0.1, inplace=False) (dropout2): Dropout(p=0.1, inplace=False) (norm1): LayerNorm((1024,), eps=1e-05, elementwise_affine=True) (norm2): LayerNorm((1024,), eps=1e-05, elementwise_affine=True) ) ) (norm): LayerNorm((1024,), eps=1e-05, elementwise_affine=True) ) (ar_predict_layer): Linear(in_features=1024, out_features=1025, bias=False) (ar_accuracy_metric): MulticlassAccuracy() (nar_audio_embeddings): ModuleList( (0): TokenEmbedding( (dropout): Dropout(p=0.0, inplace=False) (word_embeddings): Embedding(1025, 1024) ) (1-7): 7 x TokenEmbedding( (dropout): Dropout(p=0.0, inplace=False) (word_embeddings): Embedding(1024, 1024) ) ) (nar_text_prenet): Identity() (nar_audio_prenet): Identity() (nar_text_position): SinePositionalEmbedding( (dropout): Dropout(p=0.0, inplace=False) ) (nar_audio_position): SinePositionalEmbedding( (dropout): Dropout(p=0.1, inplace=False) ) (nar_decoder): TransformerEncoder( (layers): ModuleList( (0-11): 12 x TransformerEncoderLayer( (self_attn): MultiheadAttention( (out_proj): NonDynamicallyQuantizableLinear(in_features=1024, out_features=1024, bias=True) ) (linear1): Linear(in_features=1024, out_features=4096, bias=True) (dropout): Dropout(p=0.1, inplace=False) (linear2): Linear(in_features=4096, out_features=1024, bias=True) (dropout1): Dropout(p=0.1, inplace=False) (dropout2): Dropout(p=0.1, inplace=False) (norm1): AdaptiveLayerNorm( (project_layer): Linear(in_features=1024, out_features=2048, bias=True) (norm): LayerNorm((1024,), eps=1e-05, elementwise_affine=True) ) (norm2): AdaptiveLayerNorm( (project_layer): Linear(in_features=1024, out_features=2048, bias=True) (norm): LayerNorm((1024,), eps=1e-05, elementwise_affine=True) ) ) ) (norm): AdaptiveLayerNorm( (project_layer): Linear(in_features=1024, out_features=2048, bias=True) (norm): LayerNorm((1024,), eps=1e-05, elementwise_affine=True) ) ) (nar_predict_layers): ModuleList( (0-6): 7 x Linear(in_features=1024, out_features=1024, bias=False) ) (nar_stage_embeddings): ModuleList( (0-6): 7 x TokenEmbedding( (dropout): Dropout(p=0.0, inplace=False) (word_embeddings): Embedding(1, 1024) ) ) (nar_accuracy_metric): MulticlassAccuracy() ) 这个是VALL-E-X的模型参数： VALLE( (ar_text_embedding): TokenEmbedding( (dropout): Dropout(p=0.0, inplace=False) (word_embeddings): Embedding(2048, 1024) ) (nar_text_embedding): TokenEmbedding( (dropout): Dropout(p=0.0, inplace=False) (word_embeddings): Embedding(2048, 1024) ) (ar_audio_embedding): TokenEmbedding( (dropout): Dropout(p=0.0, inplace=False) (word_embeddings): Embedding(1026, 1024) ) (ar_text_prenet): Identity() (ar_audio_prenet): Identity() (ar_text_position): SinePositionalEmbedding( (dropout): Dropout(p=0.1, inplace=False) ) (ar_audio_position): SinePositionalEmbedding( (dropout): Dropout(p=0.1, inplace=False) ) (ar_decoder): TransformerEncoder( (layers): ModuleList( (0-11): 12 x TransformerEncoderLayer( (self_attn): MultiheadAttention( (out_proj): NonDynamicallyQuantizableLinear(in_features=1024, out_features=1024, bias=True) ) (linear1): Linear(in_features=1024, out_features=4096, bias=True) (dropout): Dropout(p=0.1, inplace=False) (linear2): Linear(in_features=4096, out_features=1024, bias=True) (dropout1): Dropout(p=0.1, inplace=False) (dropout2): Dropout(p=0.1, inplace=False) (norm1): LayerNorm((1024,), eps=1e-05, elementwise_affine=True) (norm2): LayerNorm((1024,), eps=1e-05, elementwise_affine=True) ) ) (norm): LayerNorm((1024,), eps=1e-05, elementwise_affine=True) ) (ar_predict_layer): Linear(in_features=1024, out_features=1025, bias=False) (nar_audio_embeddings): ModuleList( (0): TokenEmbedding( (dropout): Dropout(p=0.0, inplace=False) (word_embeddings): Embedding(1025, 1024) ) (1-7): 7 x TokenEmbedding( (dropout): Dropout(p=0.0, inplace=False) (word_embeddings): Embedding(1024, 1024) ) ) (nar_text_prenet): Identity() (nar_audio_prenet): Identity() (nar_text_position): SinePositionalEmbedding( (dropout): Dropout(p=0.0, inplace=False) ) (nar_audio_position): SinePositionalEmbedding( (dropout): Dropout(p=0.1, inplace=False) ) (nar_decoder): TransformerEncoder( (layers): ModuleList( (0-11): 12 x TransformerEncoderLayer( (self_attn): MultiheadAttention( (out_proj): NonDynamicallyQuantizableLinear(in_features=1024, out_features=1024, bias=True) ) (linear1): Linear(in_features=1024, out_features=4096, bias=True) (dropout): Dropout(p=0.1, inplace=False) (linear2): Linear(in_features=4096, out_features=1024, bias=True) (dropout1): Dropout(p=0.1, inplace=False) (dropout2): Dropout(p=0.1, inplace=False) (norm1): AdaptiveLayerNorm( (project_layer): Linear(in_features=1024, out_features=2048, bias=True) (norm): LayerNorm((1024,), eps=1e-05, elementwise_affine=True) ) (norm2): AdaptiveLayerNorm( (project_layer): Linear(in_features=1024, out_features=2048, bias=True) (norm): LayerNorm((1024,), eps=1e-05, elementwise_affine=True) ) ) ) (norm): AdaptiveLayerNorm( (project_layer): Linear(in_features=1024, out_features=2048, bias=True) (norm): LayerNorm((1024,), eps=1e-05, elementwise_affine=True) ) ) (nar_predict_layers): ModuleList( (0-6): 7 x Linear(in_features=1024, out_features=1024, bias=False) ) (nar_stage_embeddings): ModuleList( (0-6): 7 x TokenEmbedding( (dropout): Dropout(p=0.0, inplace=False) (word_embeddings): Embedding(1, 1024) ) ) (ar_language_embedding): TokenEmbedding( (dropout): Dropout(p=0.0, inplace=False) (word_embeddings): Embedding(3, 1024) ) (nar_language_embedding): TokenEmbedding( (dropout): Dropout(p=0.0, inplace=False) (word_embeddings): Embedding(3, 1024) ) ) 是否可以修改vall-e的代码去训练vall-e-x？

[Plachtaa]

25