espnet2.asr.frontend.espnet_ssl.Featureizer

source

class espnet2.asr.frontend.espnet_ssl.Featureizer(num_layers: int)

Bases: Module

Simplified S3PRL-style featurizer.

Outputs a learned weighted sum of input layers. Original code by Leo Yang (2022) in the S3PRL library. https://github.com/s3prl/s3prl/blob/main/s3prl/nn/upstream.py

forward(all_hs: List[FloatTensor])

Forward function.

• Parameters:all_hs (List *[*torch.FloatTensor ]) – List[ (batch_size, seq_len, hidden_size) ]
• Returns:
  1. The weighted-sum result, (batch_size, seq_len, hidden_size)