add cosine sim for self attention as well, as a setting

dalle2_pytorch/dalle2_pytorch.py

@@ -547,34 +547,40 @@ class NoiseScheduler(nn.Module):
 # diffusion prior
 
 class LayerNorm(nn.Module):
-    def __init__(self, dim, eps = 1e-5, stable = False):
+    def __init__(self, dim, eps = 1e-5, fp16_eps = 1e-3, stable = False):
         super().__init__()
         self.eps = eps
+        self.fp16_eps = fp16_eps
         self.stable = stable
         self.g = nn.Parameter(torch.ones(dim))
 
     def forward(self, x):
+        eps = self.eps if x.dtype == torch.float32 else self.fp16_eps
+
         if self.stable:
             x = x / x.amax(dim = -1, keepdim = True).detach()
 
         var = torch.var(x, dim = -1, unbiased = False, keepdim = True)
         mean = torch.mean(x, dim = -1, keepdim = True)
-        return (x - mean) * (var + self.eps).rsqrt() * self.g
+        return (x - mean) * (var + eps).rsqrt() * self.g
 
 class ChanLayerNorm(nn.Module):
-    def __init__(self, dim, eps = 1e-5, stable = False):
+    def __init__(self, dim, eps = 1e-5, fp16_eps = 1e-3, stable = False):
         super().__init__()
         self.eps = eps
+        self.fp16_eps = fp16_eps
         self.stable = stable
         self.g = nn.Parameter(torch.ones(1, dim, 1, 1))
 
     def forward(self, x):
+        eps = self.eps if x.dtype == torch.float32 else self.fp16_eps
+
         if self.stable:
             x = x / x.amax(dim = 1, keepdim = True).detach()
 
         var = torch.var(x, dim = 1, unbiased = False, keepdim = True)
         mean = torch.mean(x, dim = 1, keepdim = True)
-        return (x - mean) * (var + self.eps).rsqrt() * self.g
+        return (x - mean) * (var + eps).rsqrt() * self.g
 
 class Residual(nn.Module):
     def __init__(self, fn):
@@ -695,11 +701,12 @@ class Attention(nn.Module):
         dropout = 0.,
         causal = False,
         rotary_emb = None,
-        pb_relax_alpha = 128
+        cosine_sim = True,
+        cosine_sim_scale = 16
     ):
         super().__init__()
-        self.pb_relax_alpha = pb_relax_alpha
-        self.scale = dim_head ** -0.5 * (pb_relax_alpha ** -1)
+        self.scale = cosine_sim_scale if cosine_sim else (dim_head ** -0.5)
+        self.cosine_sim = cosine_sim
 
         self.heads = heads
         inner_dim = dim_head * heads
@@ -739,6 +746,13 @@ class Attention(nn.Module):
         k = torch.cat((nk, k), dim = -2)
         v = torch.cat((nv, v), dim = -2)
 
+        # whether to use cosine sim
+
+        if self.cosine_sim:
+            q, k = map(l2norm, (q, k))
+
+        q, k = map(lambda t: t * math.sqrt(self.scale), (q, k))
+
         # calculate query / key similarities
 
         sim = einsum('b h i d, b j d -> b h i j', q, k)
@@ -764,9 +778,6 @@ class Attention(nn.Module):
 
         # attention
 
-        sim = sim - sim.amax(dim = -1, keepdim = True).detach()
-        sim = sim * self.pb_relax_alpha
-
         attn = sim.softmax(dim = -1)
         attn = self.dropout(attn)
 
@@ -1598,6 +1609,7 @@ class Unet(nn.Module):
         lowres_noise_cond = False,       # for conditioning on low resolution noising, based on Imagen
         sparse_attn = False,
         cosine_sim_cross_attn = False,
+        cosine_sim_self_attn = False,
         attend_at_middle = True,         # whether to have a layer of attention at the bottleneck (can turn off for higher resolution in cascading DDPM, before bringing in efficient attention)
         cond_on_text_encodings = False,
         max_text_len = 256,
@@ -1718,7 +1730,7 @@ class Unet(nn.Module):
 
         # attention related params
 
-        attn_kwargs = dict(heads = attn_heads, dim_head = attn_dim_head)
+        attn_kwargs = dict(heads = attn_heads, dim_head = attn_dim_head, cosine_sim = cosine_sim_self_attn)
 
         self_attn = cast_tuple(self_attn, num_stages)
 

dalle2_pytorch/version.py

@@ -1 +1 @@
-__version__ = '1.4.0'
+__version__ = '1.4.3'