def __init__(self, args):
super().__init__(None)
self.w2v_encoder = Wav2VecEncoder(args)
self.is_v0_arch = not args.adaptor_proj
self.w2v_proj_ln = None
if not self.is_v0_arch and self.w2v_encoder.proj is not None:
self.w2v_proj_ln = LayerNorm(args.decoder_embed_dim)
self.adaptor = self.build_adaptor(args)
self.num_updates = 0
self.freezing_updates = args.w2v_freezing_updates
self.finetuning_params = args.finetune_w2v_params
for k, p in self.w2v_encoder.w2v_model.named_parameters():
p.requires_grad = need_finetuning(self.finetuning_params, k)
def __init__(self, args):
super().__init__(None)
self.w2v_encoder = Wav2VecEncoder(args)
encoder_out_dim = self.w2v_encoder.w2v_model.encoder.embedding_dim
# Projection + 8x shrinking
self.adaptor = Conv1dAdaptor(encoder_out_dim,
args.decoder_embed_dim,
n_layers=args.adaptor_n_layers,
kernel_size=args.adaptor_kernel_size,
stride=args.adaptor_stride,
add_layernorm=args.adaptor_layernorm)
for k, p in self.w2v_encoder.w2v_model.named_parameters():
# Freeze pretrained models by default
if safe_hasattr(args, 'finetune_w2v_params'
) and XMTransformerModel.finetune_params(
args.finetune_w2v_params, k):
p.requires_grad = True
else:
p.requires_grad = False