add ability to use t5 relative positional bias, addressing #15

setup.py

@@ -3,7 +3,7 @@
 setup(
   name = 'soundstorm-pytorch',
   packages = find_packages(exclude=[]),
-  version = '0.0.17',
+  version = '0.0.18',
   license='MIT',
   description = 'SoundStorm - Efficient Parallel Audio Generation from Google Deepmind, in Pytorch',
   author = 'Phil Wang',

soundstorm_pytorch/attend.py

@@ -132,7 +132,8 @@ def forward(self, q, k, v, mask = None, attn_bias = None):
         kv_einsum_eq = 'b j d' if k.ndim == 3 else 'b h j d'
 
         if self.flash:
-            return self.flash_attn(q, k, v, mask = mask, attn_bias = attn_bias)
+            assert not exists(attn_bias)
+            return self.flash_attn(q, k, v, mask = mask)
 
         # similarity
 

soundstorm_pytorch/soundstorm.py

@@ -126,6 +126,64 @@ def rotate_half(x):
 def apply_rotary_pos_emb(pos, t):
     return (t * pos.cos()) + (rotate_half(t) * pos.sin())
 
+# t5 relative positional bias
+
+class T5RelativePositionBias(nn.Module):
+    def __init__(
+        self,
+        scale = 1.,
+        num_buckets = 32,
+        max_distance = 128,
+        heads = 8
+    ):
+        super().__init__()
+        self.scale = scale
+        self.num_buckets = num_buckets
+        self.max_distance = max_distance
+        self.relative_attention_bias = nn.Embedding(num_buckets, heads)
+
+    @staticmethod
+    def _relative_position_bucket(
+        relative_position,
+        num_buckets = 32,
+        max_distance = 128
+    ):
+        ret = 0
+        n = -relative_position
+
+        num_buckets //= 2
+        ret += (n < 0).long() * num_buckets
+        n = torch.abs(n)
+
+        max_exact = num_buckets // 2
+        is_small = n < max_exact
+
+        val_if_large = max_exact + (
+            torch.log(n.float() / max_exact) / math.log(max_distance / max_exact) * (num_buckets - max_exact)
+        ).long()
+
+        val_if_large = torch.min(
+            val_if_large,
+            torch.full_like(val_if_large, num_buckets - 1)
+        )
+
+        ret += torch.where(is_small, n, val_if_large)
+        return ret
+
+    @property
+    def device(self):
+        return next(self.parameters()).device
+
+    def forward(self, n):
+        pos = torch.arange(n, device = self.device).long()
+        rel_pos = rearrange(pos, 'j -> 1 j') - rearrange(pos, 'i -> i 1')
+
+        rp_bucket = self._relative_position_bucket(rel_pos, num_buckets = self.num_buckets, max_distance = self.max_distance)
+        values = self.relative_attention_bias(rp_bucket)
+
+        bias = rearrange(values, 'i j h -> h i j')
+        return bias * self.scale
+
 # conformer
 
 class Swish(nn.Module):
@@ -213,7 +271,8 @@ def forward(
         x,
         context = None,
         mask = None,
-        rotary_emb = None
+        rotary_emb = None,
+        attn_bias = None
     ):
         n, device, h, has_context = x.shape[-2], x.device, self.heads, exists(context)
         context = default(context, x)
@@ -225,7 +284,7 @@ def forward(
             q = apply_rotary_pos_emb(rotary_emb, q)
             k = apply_rotary_pos_emb(rotary_emb, k)
 
-        out = self.attend(q, k, v, mask = mask)
+        out = self.attend(q, k, v, mask = mask, attn_bias = attn_bias)
 
         out = rearrange(out, 'b h n d -> b n (h d)')
         return self.to_out(out)
@@ -313,10 +372,11 @@ def forward(
         self,
         x,
         mask = None,
-        rotary_emb = None
+        rotary_emb = None,
+        attn_bias = None
     ):
         x = self.ff1(x) + x
-        x = self.attn(x, mask = mask, rotary_emb = rotary_emb) + x
+        x = self.attn(x, mask = mask, rotary_emb = rotary_emb, attn_bias = attn_bias) + x
         x = self.conv(x) + x
         x = self.ff2(x) + x
         x = self.post_norm(x)
@@ -339,13 +399,18 @@ def __init__(
         ff_dropout = 0.,
         conv_dropout = 0.,
         conv_causal = False,
-        attn_flash = True
+        attn_flash = True,
+        t5_rel_pos_bias = False
     ):
         super().__init__()
+
+        assert not (t5_rel_pos_bias and attn_flash), 'flash attention is not compatible with learned bias'
+
         self.dim = dim
         self.layers = nn.ModuleList([])
 
-        self.rotary_emb = RotaryEmbedding(dim_head)
+        self.rotary_emb = RotaryEmbedding(dim_head) if not t5_rel_pos_bias else None
+        self.rel_pos_bias = T5RelativePositionBias(dim_head ** 0.5, heads = heads) if t5_rel_pos_bias else None
 
         for _ in range(depth):
             self.layers.append(ConformerBlock(
@@ -361,11 +426,13 @@ def __init__(
             ))
 
     def forward(self, x):
+        seq_len = x.shape[-2]
 
-        rotary_emb = self.rotary_emb(x.shape[-2])
+        rotary_emb = self.rotary_emb(seq_len) if exists(self.rotary_emb) else None
+        attn_bias = self.rel_pos_bias(seq_len) if exists(self.rel_pos_bias) else None
 
         for block in self.layers:
-            x = block(x, rotary_emb = rotary_emb)
+            x = block(x, rotary_emb = rotary_emb, attn_bias = attn_bias)
 
         return x