| | |
|---|
| | | import numpy as np |
|---|
| | | from funasr.modules.nets_utils import make_pad_mask |
|---|
| | | from funasr.modules.attention import MultiHeadedAttention, MultiHeadedAttentionSANM, MultiHeadedAttentionSANMwithMask |
|---|
| | | from funasr.modules.embedding import SinusoidalPositionEncoder |
|---|
| | | from funasr.modules.embedding import SinusoidalPositionEncoder, StreamSinusoidalPositionEncoder |
|---|
| | | from funasr.modules.layer_norm import LayerNorm |
|---|
| | | from funasr.modules.multi_layer_conv import Conv1dLinear |
|---|
| | | from funasr.modules.multi_layer_conv import MultiLayeredConv1d |
|---|
| | |
|---|
| | | self.embed = torch.nn.Linear(input_size, output_size) |
|---|
| | | elif input_layer == "pe": |
|---|
| | | self.embed = SinusoidalPositionEncoder() |
|---|
| | | elif input_layer == "pe_online": |
|---|
| | | self.embed = StreamSinusoidalPositionEncoder() |
|---|
| | | else: |
|---|
| | | raise ValueError("unknown input_layer: " + input_layer) |
|---|
| | | self.normalize_before = normalize_before |
|---|
| | |
|---|
| | | if self.embed is None: |
|---|
| | | xs_pad = xs_pad |
|---|
| | | else: |
|---|
| | | xs_pad = self.embed.forward_chunk(xs_pad, cache) |
|---|
| | | xs_pad = self.embed(xs_pad, cache) |
|---|
| | | |
|---|
| | | encoder_outs = self.encoders0(xs_pad, None, None, None, None) |
|---|
| | | xs_pad, masks = encoder_outs[0], encoder_outs[1] |
|---|
