revert RoPE

grad/encoder.py

@@ -1,10 +1,9 @@
 import math
 import torch
 
-from einops import rearrange
 from grad.base import BaseModule
 from grad.reversal import SpeakerClassifier
-from grad.utils import sequence_mask
+from grad.utils import sequence_mask, convert_pad_shape
 
 
 class LayerNorm(BaseModule):
@@ -77,67 +76,8 @@ def calc_mean_std(self, x, mask=None):
         return mn, sd
 
 
-class RotaryPositionalEmbeddings(BaseModule):
-    """
-    ## RoPE module
-    https://github.com/labmlai/annotated_deep_learning_paper_implementations
-    
-    Rotary encoding transforms pairs of features by rotating in the 2D plane.
-    That is, it organizes the $d$ features as $\frac{d}{2}$ pairs.
-    Each pair can be considered a coordinate in a 2D plane, and the encoding will rotate it
-    by an angle depending on the position of the token.
-    """
-    def __init__(self, d: int, base: int = 10_000):
-        r"""
-        * `d` is the number of features $d$
-        * `base` is the constant used for calculating $\Theta$
-        """
-        super().__init__()
-        self.base = base
-        self.d = int(d)
-        self.cos_cached = None
-        self.sin_cached = None
-
-    def _build_cache(self, x: torch.Tensor):
-        r"""
-        Cache $\cos$ and $\sin$ values
-        """
-        # Return if cache is already built
-        if self.cos_cached is not None and x.shape[0] <= self.cos_cached.shape[0]:
-            return
-        # Get sequence length
-        seq_len = x.shape[0]
-        theta = 1.0 / (self.base ** (torch.arange(0, self.d, 2).float() / self.d)).to(x.device)
-        # Create position indexes `[0, 1, ..., seq_len - 1]`
-        seq_idx = torch.arange(seq_len, device=x.device).float().to(x.device)
-        # Calculate the product of position index and $\theta_i$
-        idx_theta = torch.einsum("n,d->nd", seq_idx, theta)
-        # Concatenate so that for row $m$ we have
-        idx_theta2 = torch.cat([idx_theta, idx_theta], dim=1)
-        # Cache them
-        self.cos_cached = idx_theta2.cos()[:, None, None, :]
-        self.sin_cached = idx_theta2.sin()[:, None, None, :]
-
-    def _neg_half(self, x: torch.Tensor):
-        d_2 = self.d // 2
-        return torch.cat([-x[:, :, :, d_2:], x[:, :, :, :d_2]], dim=-1)
-
-    def forward(self, x: torch.Tensor):
-        """
-        * `x` is the Tensor at the head of a key or a query with shape `[seq_len, batch_size, n_heads, d]`
-        """
-        x = rearrange(x, "b h t d -> t b h d")
-        self._build_cache(x)
-        # Split the features, we can choose to apply rotary embeddings only to a partial set of features.
-        x_rope, x_pass = x[..., : self.d], x[..., self.d :]
-        # Calculate
-        neg_half_x = self._neg_half(x_rope)
-        x_rope = (x_rope * self.cos_cached[: x.shape[0]]) + (neg_half_x * self.sin_cached[: x.shape[0]])
-        return rearrange(torch.cat((x_rope, x_pass), dim=-1), "t b h d -> b h t d")
-
-
 class MultiHeadAttention(BaseModule):
-    def __init__(self, channels, out_channels, n_heads, 
+    def __init__(self, channels, out_channels, n_heads, window_size=None, 
                  heads_share=True, p_dropout=0.0, proximal_bias=False, 
                  proximal_init=False):
         super(MultiHeadAttention, self).__init__()
@@ -146,6 +86,7 @@ def __init__(self, channels, out_channels, n_heads,
         self.channels = channels
         self.out_channels = out_channels
         self.n_heads = n_heads
+        self.window_size = window_size
         self.heads_share = heads_share
         self.proximal_bias = proximal_bias
         self.p_dropout = p_dropout
@@ -155,11 +96,13 @@ def __init__(self, channels, out_channels, n_heads,
         self.conv_q = torch.nn.Conv1d(channels, channels, 1)
         self.conv_k = torch.nn.Conv1d(channels, channels, 1)
         self.conv_v = torch.nn.Conv1d(channels, channels, 1)
-
-        # from https://nn.labml.ai/transformers/rope/index.html
-        self.query_rotary_pe = RotaryPositionalEmbeddings(self.k_channels * 0.5)
-        self.key_rotary_pe = RotaryPositionalEmbeddings(self.k_channels * 0.5)
-
+        if window_size is not None:
+            n_heads_rel = 1 if heads_share else n_heads
+            rel_stddev = self.k_channels**-0.5
+            self.emb_rel_k = torch.nn.Parameter(torch.randn(n_heads_rel, 
+                             window_size * 2 + 1, self.k_channels) * rel_stddev)
+            self.emb_rel_v = torch.nn.Parameter(torch.randn(n_heads_rel, 
+                             window_size * 2 + 1, self.k_channels) * rel_stddev)
         self.conv_o = torch.nn.Conv1d(channels, out_channels, 1)
         self.drop = torch.nn.Dropout(p_dropout)
 
@@ -169,28 +112,31 @@ def __init__(self, channels, out_channels, n_heads,
             self.conv_k.weight.data.copy_(self.conv_q.weight.data)
             self.conv_k.bias.data.copy_(self.conv_q.bias.data)
         torch.nn.init.xavier_uniform_(self.conv_v.weight)
-
+        
     def forward(self, x, c, attn_mask=None):
         q = self.conv_q(x)
         k = self.conv_k(c)
         v = self.conv_v(c)
-
+        
         x, self.attn = self.attention(q, k, v, mask=attn_mask)
 
         x = self.conv_o(x)
         return x
 
     def attention(self, query, key, value, mask=None):
         b, d, t_s, t_t = (*key.size(), query.size(2))
-        query = rearrange(query, "b (h c) t-> b h t c", h=self.n_heads)
-        key = rearrange(key, "b (h c) t-> b h t c", h=self.n_heads)
-        value = rearrange(value, "b (h c) t-> b h t c", h=self.n_heads)
-
-        query = self.query_rotary_pe(query)
-        key = self.key_rotary_pe(key)
+        query = query.view(b, self.n_heads, self.k_channels, t_t).transpose(2, 3)
+        key = key.view(b, self.n_heads, self.k_channels, t_s).transpose(2, 3)
+        value = value.view(b, self.n_heads, self.k_channels, t_s).transpose(2, 3)
 
         scores = torch.matmul(query, key.transpose(-2, -1)) / math.sqrt(self.k_channels)
-
+        if self.window_size is not None:
+            assert t_s == t_t, "Relative attention is only available for self-attention."
+            key_relative_embeddings = self._get_relative_embeddings(self.emb_rel_k, t_s)
+            rel_logits = self._matmul_with_relative_keys(query, key_relative_embeddings)
+            rel_logits = self._relative_position_to_absolute_position(rel_logits)
+            scores_local = rel_logits / math.sqrt(self.k_channels)
+            scores = scores + scores_local
         if self.proximal_bias:
             assert t_s == t_t, "Proximal bias is only available for self-attention."
             scores = scores + self._attention_bias_proximal(t_s).to(device=scores.device, 
@@ -200,9 +146,52 @@ def attention(self, query, key, value, mask=None):
         p_attn = torch.nn.functional.softmax(scores, dim=-1)
         p_attn = self.drop(p_attn)
         output = torch.matmul(p_attn, value)
+        if self.window_size is not None:
+            relative_weights = self._absolute_position_to_relative_position(p_attn)
+            value_relative_embeddings = self._get_relative_embeddings(self.emb_rel_v, t_s)
+            output = output + self._matmul_with_relative_values(relative_weights, 
+                                                                value_relative_embeddings)
         output = output.transpose(2, 3).contiguous().view(b, d, t_t)
         return output, p_attn
 
+    def _matmul_with_relative_values(self, x, y):
+        ret = torch.matmul(x, y.unsqueeze(0))
+        return ret
+
+    def _matmul_with_relative_keys(self, x, y):
+        ret = torch.matmul(x, y.unsqueeze(0).transpose(-2, -1))
+        return ret
+
+    def _get_relative_embeddings(self, relative_embeddings, length):
+        pad_length = max(length - (self.window_size + 1), 0)
+        slice_start_position = max((self.window_size + 1) - length, 0)
+        slice_end_position = slice_start_position + 2 * length - 1
+        if pad_length > 0:
+            padded_relative_embeddings = torch.nn.functional.pad(
+                            relative_embeddings, convert_pad_shape([[0, 0], 
+                            [pad_length, pad_length], [0, 0]]))
+        else:
+            padded_relative_embeddings = relative_embeddings
+        used_relative_embeddings = padded_relative_embeddings[:,
+                                   slice_start_position:slice_end_position]
+        return used_relative_embeddings
+
+    def _relative_position_to_absolute_position(self, x):
+        batch, heads, length, _ = x.size()
+        x = torch.nn.functional.pad(x, convert_pad_shape([[0,0],[0,0],[0,0],[0,1]]))
+        x_flat = x.view([batch, heads, length * 2 * length])
+        x_flat = torch.nn.functional.pad(x_flat, convert_pad_shape([[0,0],[0,0],[0,length-1]]))
+        x_final = x_flat.view([batch, heads, length+1, 2*length-1])[:, :, :length, length-1:]
+        return x_final
+
+    def _absolute_position_to_relative_position(self, x):
+        batch, heads, length, _ = x.size()
+        x = torch.nn.functional.pad(x, convert_pad_shape([[0, 0], [0, 0], [0, 0], [0, length-1]]))
+        x_flat = x.view([batch, heads, length**2 + length*(length - 1)])
+        x_flat = torch.nn.functional.pad(x_flat, convert_pad_shape([[0, 0], [0, 0], [length, 0]]))
+        x_final = x_flat.view([batch, heads, length, 2*length])[:,:,:,1:]
+        return x_final
+
     def _attention_bias_proximal(self, length):
         r = torch.arange(length, dtype=torch.float32)
         diff = torch.unsqueeze(r, 0) - torch.unsqueeze(r, 1)
@@ -235,23 +224,24 @@ def forward(self, x, x_mask):
 
 class Encoder(BaseModule):
     def __init__(self, hidden_channels, filter_channels, n_heads, n_layers, 
-                 kernel_size=1, p_dropout=0.0, **kwargs):
+                 kernel_size=1, p_dropout=0.0, window_size=None, **kwargs):
         super(Encoder, self).__init__()
         self.hidden_channels = hidden_channels
         self.filter_channels = filter_channels
         self.n_heads = n_heads
         self.n_layers = n_layers
         self.kernel_size = kernel_size
         self.p_dropout = p_dropout
+        self.window_size = window_size
 
         self.drop = torch.nn.Dropout(p_dropout)
         self.attn_layers = torch.nn.ModuleList()
         self.norm_layers_1 = torch.nn.ModuleList()
         self.ffn_layers = torch.nn.ModuleList()
         self.norm_layers_2 = torch.nn.ModuleList()
         for _ in range(self.n_layers):
-            self.attn_layers.append(MultiHeadAttention(hidden_channels, hidden_channels, 
-                                                       n_heads, p_dropout=p_dropout))
+            self.attn_layers.append(MultiHeadAttention(hidden_channels, hidden_channels,
+                                    n_heads, window_size=window_size, p_dropout=p_dropout))
             self.norm_layers_1.append(LayerNorm(hidden_channels))
             self.ffn_layers.append(FFN(hidden_channels, hidden_channels,
                                        filter_channels, kernel_size, p_dropout=p_dropout))
@@ -278,7 +268,8 @@ def __init__(self, n_vecs, n_mels, n_embs,
                  n_heads=2,
                  n_layers=6,
                  kernel_size=3,
-                 p_dropout=0.1):
+                 p_dropout=0.1,
+                 window_size=4):
         super(TextEncoder, self).__init__()
         self.n_vecs = n_vecs
         self.n_mels = n_mels
@@ -289,6 +280,7 @@ def __init__(self, n_vecs, n_mels, n_embs,
         self.n_layers = n_layers
         self.kernel_size = kernel_size
         self.p_dropout = p_dropout
+        self.window_size = window_size
 
         self.prenet = ConvReluNorm(n_vecs,
                                    n_channels,
@@ -307,7 +299,8 @@ def __init__(self, n_vecs, n_mels, n_embs,
                                n_heads,
                                n_layers,
                                kernel_size,
-                               p_dropout)
+                               p_dropout,
+                               window_size=window_size)
 
         self.proj_m = torch.nn.Conv1d(n_channels + n_embs + n_embs, n_mels, 1)