|
|
@@ -0,0 +1,403 @@
|
|
|
+"""Positional Encoding Module."""
|
|
|
+
|
|
|
+import math
|
|
|
+
|
|
|
+import torch
|
|
|
+import torch.nn as nn
|
|
|
+from funasr.modules.embedding import (
|
|
|
+ LegacyRelPositionalEncoding, PositionalEncoding, RelPositionalEncoding,
|
|
|
+ ScaledPositionalEncoding, StreamPositionalEncoding)
|
|
|
+from funasr.modules.subsampling import (
|
|
|
+ Conv2dSubsampling, Conv2dSubsampling2, Conv2dSubsampling6,
|
|
|
+ Conv2dSubsampling8)
|
|
|
+from funasr.modules.subsampling_without_posenc import \
|
|
|
+ Conv2dSubsamplingWOPosEnc
|
|
|
+
|
|
|
+from funasr.export.models.language_models.subsampling import (
|
|
|
+ OnnxConv2dSubsampling, OnnxConv2dSubsampling2, OnnxConv2dSubsampling6,
|
|
|
+ OnnxConv2dSubsampling8)
|
|
|
+
|
|
|
+
|
|
|
+def get_pos_emb(pos_emb, max_seq_len=512, use_cache=True):
|
|
|
+ if isinstance(pos_emb, LegacyRelPositionalEncoding):
|
|
|
+ return OnnxLegacyRelPositionalEncoding(pos_emb, max_seq_len, use_cache)
|
|
|
+ elif isinstance(pos_emb, ScaledPositionalEncoding):
|
|
|
+ return OnnxScaledPositionalEncoding(pos_emb, max_seq_len, use_cache)
|
|
|
+ elif isinstance(pos_emb, RelPositionalEncoding):
|
|
|
+ return OnnxRelPositionalEncoding(pos_emb, max_seq_len, use_cache)
|
|
|
+ elif isinstance(pos_emb, PositionalEncoding):
|
|
|
+ return OnnxPositionalEncoding(pos_emb, max_seq_len, use_cache)
|
|
|
+ elif isinstance(pos_emb, StreamPositionalEncoding):
|
|
|
+ return OnnxStreamPositionalEncoding(pos_emb, max_seq_len, use_cache)
|
|
|
+ elif (isinstance(pos_emb, nn.Sequential) and len(pos_emb) == 0) or (
|
|
|
+ isinstance(pos_emb, Conv2dSubsamplingWOPosEnc)
|
|
|
+ ):
|
|
|
+ return pos_emb
|
|
|
+ else:
|
|
|
+ raise ValueError("Embedding model is not supported.")
|
|
|
+
|
|
|
+
|
|
|
+class Embedding(nn.Module):
|
|
|
+ def __init__(self, model, max_seq_len=512, use_cache=True):
|
|
|
+ super().__init__()
|
|
|
+ self.model = model
|
|
|
+ if not isinstance(model, nn.Embedding):
|
|
|
+ if isinstance(model, Conv2dSubsampling):
|
|
|
+ self.model = OnnxConv2dSubsampling(model)
|
|
|
+ self.model.out[-1] = get_pos_emb(model.out[-1], max_seq_len)
|
|
|
+ elif isinstance(model, Conv2dSubsampling2):
|
|
|
+ self.model = OnnxConv2dSubsampling2(model)
|
|
|
+ self.model.out[-1] = get_pos_emb(model.out[-1], max_seq_len)
|
|
|
+ elif isinstance(model, Conv2dSubsampling6):
|
|
|
+ self.model = OnnxConv2dSubsampling6(model)
|
|
|
+ self.model.out[-1] = get_pos_emb(model.out[-1], max_seq_len)
|
|
|
+ elif isinstance(model, Conv2dSubsampling8):
|
|
|
+ self.model = OnnxConv2dSubsampling8(model)
|
|
|
+ self.model.out[-1] = get_pos_emb(model.out[-1], max_seq_len)
|
|
|
+ else:
|
|
|
+ self.model[-1] = get_pos_emb(model[-1], max_seq_len)
|
|
|
+
|
|
|
+ def forward(self, x, mask=None):
|
|
|
+ if mask is None:
|
|
|
+ return self.model(x)
|
|
|
+ else:
|
|
|
+ return self.model(x, mask)
|
|
|
+
|
|
|
+
|
|
|
+def _pre_hook(
|
|
|
+ state_dict,
|
|
|
+ prefix,
|
|
|
+ local_metadata,
|
|
|
+ strict,
|
|
|
+ missing_keys,
|
|
|
+ unexpected_keys,
|
|
|
+ error_msgs,
|
|
|
+):
|
|
|
+ """Perform pre-hook in load_state_dict for backward compatibility.
|
|
|
+
|
|
|
+ Note:
|
|
|
+ We saved self.pe until v.0.5.2 but we have omitted it later.
|
|
|
+ Therefore, we remove the item "pe" from `state_dict` for backward compatibility.
|
|
|
+
|
|
|
+ """
|
|
|
+ k = prefix + "pe"
|
|
|
+ if k in state_dict:
|
|
|
+ state_dict.pop(k)
|
|
|
+
|
|
|
+
|
|
|
+class OnnxPositionalEncoding(torch.nn.Module):
|
|
|
+ """Positional encoding.
|
|
|
+
|
|
|
+ Args:
|
|
|
+ d_model (int): Embedding dimension.
|
|
|
+ dropout_rate (float): Dropout rate.
|
|
|
+ max_seq_len (int): Maximum input length.
|
|
|
+ reverse (bool): Whether to reverse the input position. Only for
|
|
|
+ the class LegacyRelPositionalEncoding. We remove it in the current
|
|
|
+ class RelPositionalEncoding.
|
|
|
+ """
|
|
|
+
|
|
|
+ def __init__(self, model, max_seq_len=512, reverse=False, use_cache=True):
|
|
|
+ """Construct an PositionalEncoding object."""
|
|
|
+ super(OnnxPositionalEncoding, self).__init__()
|
|
|
+ self.d_model = model.d_model
|
|
|
+ self.reverse = reverse
|
|
|
+ self.max_seq_len = max_seq_len
|
|
|
+ self.xscale = math.sqrt(self.d_model)
|
|
|
+ self._register_load_state_dict_pre_hook(_pre_hook)
|
|
|
+ self.pe = model.pe
|
|
|
+ self.use_cache = use_cache
|
|
|
+ self.model = model
|
|
|
+ if self.use_cache:
|
|
|
+ self.extend_pe()
|
|
|
+ else:
|
|
|
+ self.div_term = torch.exp(
|
|
|
+ torch.arange(0, self.d_model, 2, dtype=torch.float32)
|
|
|
+ * -(math.log(10000.0) / self.d_model)
|
|
|
+ )
|
|
|
+
|
|
|
+ def extend_pe(self):
|
|
|
+ """Reset the positional encodings."""
|
|
|
+ pe_length = len(self.pe[0])
|
|
|
+ if self.max_seq_len < pe_length:
|
|
|
+ self.pe = self.pe[:, : self.max_seq_len]
|
|
|
+ else:
|
|
|
+ self.model.extend_pe(torch.tensor(0.0).expand(1, self.max_seq_len))
|
|
|
+ self.pe = self.model.pe
|
|
|
+
|
|
|
+ def _add_pe(self, x):
|
|
|
+ """Computes positional encoding"""
|
|
|
+ if self.reverse:
|
|
|
+ position = torch.arange(
|
|
|
+ x.size(1) - 1, -1, -1.0, dtype=torch.float32
|
|
|
+ ).unsqueeze(1)
|
|
|
+ else:
|
|
|
+ position = torch.arange(0, x.size(1), dtype=torch.float32).unsqueeze(1)
|
|
|
+
|
|
|
+ x = x * self.xscale
|
|
|
+ x[:, :, 0::2] += torch.sin(position * self.div_term)
|
|
|
+ x[:, :, 1::2] += torch.cos(position * self.div_term)
|
|
|
+ return x
|
|
|
+
|
|
|
+ def forward(self, x: torch.Tensor):
|
|
|
+ """Add positional encoding.
|
|
|
+
|
|
|
+ Args:
|
|
|
+ x (torch.Tensor): Input tensor (batch, time, `*`).
|
|
|
+
|
|
|
+ Returns:
|
|
|
+ torch.Tensor: Encoded tensor (batch, time, `*`).
|
|
|
+ """
|
|
|
+ if self.use_cache:
|
|
|
+ x = x * self.xscale + self.pe[:, : x.size(1)]
|
|
|
+ else:
|
|
|
+ x = self._add_pe(x)
|
|
|
+ return x
|
|
|
+
|
|
|
+
|
|
|
+class OnnxScaledPositionalEncoding(OnnxPositionalEncoding):
|
|
|
+ """Scaled positional encoding module.
|
|
|
+
|
|
|
+ See Sec. 3.2 https://arxiv.org/abs/1809.08895
|
|
|
+
|
|
|
+ Args:
|
|
|
+ d_model (int): Embedding dimension.
|
|
|
+ dropout_rate (float): Dropout rate.
|
|
|
+ max_seq_len (int): Maximum input length.
|
|
|
+
|
|
|
+ """
|
|
|
+
|
|
|
+ def __init__(self, model, max_seq_len=512, use_cache=True):
|
|
|
+ """Initialize class."""
|
|
|
+ super().__init__(model, max_seq_len, use_cache=use_cache)
|
|
|
+ self.alpha = torch.nn.Parameter(torch.tensor(1.0))
|
|
|
+
|
|
|
+ def reset_parameters(self):
|
|
|
+ """Reset parameters."""
|
|
|
+ self.alpha.data = torch.tensor(1.0)
|
|
|
+
|
|
|
+ def _add_pe(self, x):
|
|
|
+ """Computes positional encoding"""
|
|
|
+ if self.reverse:
|
|
|
+ position = torch.arange(
|
|
|
+ x.size(1) - 1, -1, -1.0, dtype=torch.float32
|
|
|
+ ).unsqueeze(1)
|
|
|
+ else:
|
|
|
+ position = torch.arange(0, x.size(1), dtype=torch.float32).unsqueeze(1)
|
|
|
+
|
|
|
+ x = x * self.alpha
|
|
|
+ x[:, :, 0::2] += torch.sin(position * self.div_term)
|
|
|
+ x[:, :, 1::2] += torch.cos(position * self.div_term)
|
|
|
+ return x
|
|
|
+
|
|
|
+ def forward(self, x):
|
|
|
+ """Add positional encoding.
|
|
|
+
|
|
|
+ Args:
|
|
|
+ x (torch.Tensor): Input tensor (batch, time, `*`).
|
|
|
+
|
|
|
+ Returns:
|
|
|
+ torch.Tensor: Encoded tensor (batch, time, `*`).
|
|
|
+
|
|
|
+ """
|
|
|
+ if self.use_cache:
|
|
|
+ x = x + self.alpha * self.pe[:, : x.size(1)]
|
|
|
+ else:
|
|
|
+ x = self._add_pe(x)
|
|
|
+ return x
|
|
|
+
|
|
|
+
|
|
|
+class OnnxLegacyRelPositionalEncoding(OnnxPositionalEncoding):
|
|
|
+ """Relative positional encoding module (old version).
|
|
|
+
|
|
|
+ Details can be found in https://github.com/espnet/espnet/pull/2816.
|
|
|
+
|
|
|
+ See : Appendix B in https://arxiv.org/abs/1901.02860
|
|
|
+
|
|
|
+ Args:
|
|
|
+ d_model (int): Embedding dimension.
|
|
|
+ dropout_rate (float): Dropout rate.
|
|
|
+ max_seq_len (int): Maximum input length.
|
|
|
+
|
|
|
+ """
|
|
|
+
|
|
|
+ def __init__(self, model, max_seq_len=512, use_cache=True):
|
|
|
+ """Initialize class."""
|
|
|
+ super().__init__(model, max_seq_len, reverse=True, use_cache=use_cache)
|
|
|
+
|
|
|
+ def _get_pe(self, x):
|
|
|
+ """Computes positional encoding"""
|
|
|
+ if self.reverse:
|
|
|
+ position = torch.arange(
|
|
|
+ x.size(1) - 1, -1, -1.0, dtype=torch.float32
|
|
|
+ ).unsqueeze(1)
|
|
|
+ else:
|
|
|
+ position = torch.arange(0, x.size(1), dtype=torch.float32).unsqueeze(1)
|
|
|
+
|
|
|
+ pe = torch.zeros(x.shape)
|
|
|
+ pe[:, :, 0::2] += torch.sin(position * self.div_term)
|
|
|
+ pe[:, :, 1::2] += torch.cos(position * self.div_term)
|
|
|
+ return pe
|
|
|
+
|
|
|
+ def forward(self, x):
|
|
|
+ """Compute positional encoding.
|
|
|
+
|
|
|
+ Args:
|
|
|
+ x (torch.Tensor): Input tensor (batch, time, `*`).
|
|
|
+
|
|
|
+ Returns:
|
|
|
+ torch.Tensor: Encoded tensor (batch, time, `*`).
|
|
|
+ torch.Tensor: Positional embedding tensor (1, time, `*`).
|
|
|
+
|
|
|
+ """
|
|
|
+ x = x * self.xscale
|
|
|
+ if self.use_cache:
|
|
|
+ pos_emb = self.pe[:, : x.size(1)]
|
|
|
+ else:
|
|
|
+ pos_emb = self._get_pe(x)
|
|
|
+ return x, pos_emb
|
|
|
+
|
|
|
+
|
|
|
+class OnnxRelPositionalEncoding(torch.nn.Module):
|
|
|
+ """Relative positional encoding module (new implementation).
|
|
|
+ Details can be found in https://github.com/espnet/espnet/pull/2816.
|
|
|
+ See : Appendix B in https://arxiv.org/abs/1901.02860
|
|
|
+ Args:
|
|
|
+ d_model (int): Embedding dimension.
|
|
|
+ dropout_rate (float): Dropout rate.
|
|
|
+ max_seq_len (int): Maximum input length.
|
|
|
+ """
|
|
|
+
|
|
|
+ def __init__(self, model, max_seq_len=512, use_cache=True):
|
|
|
+ """Construct an PositionalEncoding object."""
|
|
|
+ super(OnnxRelPositionalEncoding, self).__init__()
|
|
|
+ self.d_model = model.d_model
|
|
|
+ self.xscale = math.sqrt(self.d_model)
|
|
|
+ self.pe = None
|
|
|
+ self.use_cache = use_cache
|
|
|
+ if self.use_cache:
|
|
|
+ self.extend_pe(torch.tensor(0.0).expand(1, max_seq_len))
|
|
|
+ else:
|
|
|
+ self.div_term = torch.exp(
|
|
|
+ torch.arange(0, self.d_model, 2, dtype=torch.float32)
|
|
|
+ * -(math.log(10000.0) / self.d_model)
|
|
|
+ )
|
|
|
+
|
|
|
+ def extend_pe(self, x):
|
|
|
+ """Reset the positional encodings."""
|
|
|
+ if self.pe is not None and self.pe.size(1) >= x.size(1) * 2 - 1:
|
|
|
+ # self.pe contains both positive and negative parts
|
|
|
+ # the length of self.pe is 2 * input_len - 1
|
|
|
+ if self.pe.dtype != x.dtype or self.pe.device != x.device:
|
|
|
+ self.pe = self.pe.to(dtype=x.dtype, device=x.device)
|
|
|
+ return
|
|
|
+ # Suppose `i` means to the position of query vecotr and `j` means the
|
|
|
+ # position of key vector. We use position relative positions when keys
|
|
|
+ # are to the left (i>j) and negative relative positions otherwise (i<j).
|
|
|
+ pe_positive = torch.zeros(x.size(1), self.d_model)
|
|
|
+ pe_negative = torch.zeros(x.size(1), self.d_model)
|
|
|
+ position = torch.arange(0, x.size(1), dtype=torch.float32).unsqueeze(1)
|
|
|
+ div_term = torch.exp(
|
|
|
+ torch.arange(0, self.d_model, 2, dtype=torch.float32)
|
|
|
+ * -(math.log(10000.0) / self.d_model)
|
|
|
+ )
|
|
|
+ pe_positive[:, 0::2] = torch.sin(position * div_term)
|
|
|
+ pe_positive[:, 1::2] = torch.cos(position * div_term)
|
|
|
+ pe_negative[:, 0::2] = torch.sin(-1 * position * div_term)
|
|
|
+ pe_negative[:, 1::2] = torch.cos(-1 * position * div_term)
|
|
|
+
|
|
|
+ # Reserve the order of positive indices and concat both positive and
|
|
|
+ # negative indices. This is used to support the shifting trick
|
|
|
+ # as in https://arxiv.org/abs/1901.02860
|
|
|
+ pe_positive = torch.flip(pe_positive, [0]).unsqueeze(0)
|
|
|
+ pe_negative = pe_negative[1:].unsqueeze(0)
|
|
|
+ pe = torch.cat([pe_positive, pe_negative], dim=1)
|
|
|
+ self.pe = pe.to(device=x.device, dtype=x.dtype)
|
|
|
+
|
|
|
+ def _get_pe(self, x):
|
|
|
+ pe_positive = torch.zeros(x.size(1), self.d_model)
|
|
|
+ pe_negative = torch.zeros(x.size(1), self.d_model)
|
|
|
+ theta = (
|
|
|
+ torch.arange(0, x.size(1), dtype=torch.float32).unsqueeze(1) * self.div_term
|
|
|
+ )
|
|
|
+ pe_positive[:, 0::2] = torch.sin(theta)
|
|
|
+ pe_positive[:, 1::2] = torch.cos(theta)
|
|
|
+ pe_negative[:, 0::2] = -1 * torch.sin(theta)
|
|
|
+ pe_negative[:, 1::2] = torch.cos(theta)
|
|
|
+
|
|
|
+ # Reserve the order of positive indices and concat both positive and
|
|
|
+ # negative indices. This is used to support the shifting trick
|
|
|
+ # as in https://arxiv.org/abs/1901.02860
|
|
|
+ pe_positive = torch.flip(pe_positive, [0]).unsqueeze(0)
|
|
|
+ pe_negative = pe_negative[1:].unsqueeze(0)
|
|
|
+ return torch.cat([pe_positive, pe_negative], dim=1)
|
|
|
+
|
|
|
+ def forward(self, x: torch.Tensor, use_cache=True):
|
|
|
+ """Add positional encoding.
|
|
|
+ Args:
|
|
|
+ x (torch.Tensor): Input tensor (batch, time, `*`).
|
|
|
+ Returns:
|
|
|
+ torch.Tensor: Encoded tensor (batch, time, `*`).
|
|
|
+ """
|
|
|
+ x = x * self.xscale
|
|
|
+ if self.use_cache:
|
|
|
+ pos_emb = self.pe[
|
|
|
+ :,
|
|
|
+ self.pe.size(1) // 2 - x.size(1) + 1 : self.pe.size(1) // 2 + x.size(1),
|
|
|
+ ]
|
|
|
+ else:
|
|
|
+ pos_emb = self._get_pe(x)
|
|
|
+ return x, pos_emb
|
|
|
+
|
|
|
+
|
|
|
+class OnnxStreamPositionalEncoding(torch.nn.Module):
|
|
|
+ """Streaming Positional encoding."""
|
|
|
+
|
|
|
+ def __init__(self, model, max_seq_len=5000, use_cache=True):
|
|
|
+ """Construct an PositionalEncoding object."""
|
|
|
+ super(StreamPositionalEncoding, self).__init__()
|
|
|
+ self.use_cache = use_cache
|
|
|
+ self.d_model = model.d_model
|
|
|
+ self.xscale = model.xscale
|
|
|
+ self.pe = model.pe
|
|
|
+ self.use_cache = use_cache
|
|
|
+ self.max_seq_len = max_seq_len
|
|
|
+ if self.use_cache:
|
|
|
+ self.extend_pe()
|
|
|
+ else:
|
|
|
+ self.div_term = torch.exp(
|
|
|
+ torch.arange(0, self.d_model, 2, dtype=torch.float32)
|
|
|
+ * -(math.log(10000.0) / self.d_model)
|
|
|
+ )
|
|
|
+ self._register_load_state_dict_pre_hook(_pre_hook)
|
|
|
+
|
|
|
+ def extend_pe(self):
|
|
|
+ """Reset the positional encodings."""
|
|
|
+ pe_length = len(self.pe[0])
|
|
|
+ if self.max_seq_len < pe_length:
|
|
|
+ self.pe = self.pe[:, : self.max_seq_len]
|
|
|
+ else:
|
|
|
+ self.model.extend_pe(self.max_seq_len)
|
|
|
+ self.pe = self.model.pe
|
|
|
+
|
|
|
+ def _add_pe(self, x, start_idx):
|
|
|
+ position = torch.arange(start_idx, x.size(1), dtype=torch.float32).unsqueeze(1)
|
|
|
+ x = x * self.xscale
|
|
|
+ x[:, :, 0::2] += torch.sin(position * self.div_term)
|
|
|
+ x[:, :, 1::2] += torch.cos(position * self.div_term)
|
|
|
+ return x
|
|
|
+
|
|
|
+ def forward(self, x: torch.Tensor, start_idx: int = 0):
|
|
|
+ """Add positional encoding.
|
|
|
+
|
|
|
+ Args:
|
|
|
+ x (torch.Tensor): Input tensor (batch, time, `*`).
|
|
|
+
|
|
|
+ Returns:
|
|
|
+ torch.Tensor: Encoded tensor (batch, time, `*`).
|
|
|
+
|
|
|
+ """
|
|
|
+ if self.use_cache:
|
|
|
+ return x * self.xscale + self.pe[:, start_idx : start_idx + x.size(1)]
|
|
|
+ else:
|
|
|
+ return self._add_pe(x, start_idx)
|
游雁