x-transformers 1.31.0__py3-none-any.whl → 1.31.2__py3-none-any.whl

x_transformers/x_transformers.py

@@ -573,12 +573,17 @@ class LayerNorm(Module):
         """
         super().__init__()
         self.unit_offset = unit_offset
+
+        self.ln = nn.LayerNorm(dim, elementwise_affine = False)
         self.gamma = nn.Parameter(torch.ones(dim))
+        nn.init.constant_(self.gamma, 1. - unit_offset)
+
         self.register_buffer('beta', torch.zeros(dim), persistent = False)
 
     def forward(self, x):
+        normed = self.ln(x)
         gamma = self.gamma + self.unit_offset
-        return F.layer_norm(x, x.shape[-1:], gamma, self.beta)
+        return normed * gamma
 
 class AdaptiveLayerNorm(Module):
     def __init__(
@@ -608,7 +613,9 @@ class ScaleNorm(Module):
         super().__init__()
         self.unit_offset = unit_offset
         self.scale = dim ** 0.5
-        self.g = nn.Parameter(torch.ones(1))
+
+        self.g = nn.Parameter(torch.zeros(1))
+        nn.init.constant_(self.g, 1. - unit_offset)
 
     def forward(self, x):
         return F.normalize(x, dim = -1) * self.scale * (self.g + self.unit_offset)
@@ -622,7 +629,9 @@ class RMSNorm(Module):
         super().__init__()
         self.unit_offset = unit_offset
         self.scale = dim ** 0.5
-        self.g = nn.Parameter(torch.ones(dim))
+
+        self.g = nn.Parameter(torch.zeros(dim))
+        nn.init.constant_(self.g, 1. - unit_offset)
 
     def forward(self, x):
         return F.normalize(x, dim = -1) * self.scale * (self.g + self.unit_offset)
@@ -724,10 +733,19 @@ class ShiftTokens(Module):
 # post branch operator
 
 class LayerScale(Module):
-    def __init__(self, fn: Module, dim, init_value = 0.):
+    def __init__(
+        self,
+        fn: Module,
+        dim,
+        init_value = 0.,
+        unit_offset = 0.
+    ):
         super().__init__()
+        self.unit_offset = unit_offset
+
         self.fn = fn
-        self.gamma = nn.Parameter(torch.ones(dim) * init_value)
+        self.gamma = nn.Parameter(torch.zeros(dim))
+        nn.init.constant_(self.gamma, init_value - unit_offset)
 
     def forward(self, x, **kwargs):
         out = self.fn(x, **kwargs)
@@ -739,7 +757,13 @@ class LayerScale(Module):
         return out * self.gamma, *rest
 
 class AdaptiveLayerScale(Module):
-    def __init__(self, fn: Module, dim, dim_condition = None, init_bias_value = -2.):
+    def __init__(
+        self,
+        fn: Module,
+        dim,
+        dim_condition = None,
+        init_bias_value = -2.
+    ):
         super().__init__()
         self.fn = fn
 
@@ -1210,7 +1234,7 @@ class AttentionLayers(Module):
         use_adaptive_layernorm = False,
         use_adaptive_rmsnorm = False,
         use_adaptive_layerscale = False, # paired with use_adaptive_layernorm for ada-ln-zero from DiT paper
-        norm_add_unit_offset = False,
+        norm_add_unit_offset = True,
         dim_condition = None,
         adaptive_condition_mlp = False,
         adaptive_condition_mlp_expansion = 4,
@@ -1346,7 +1370,7 @@ class AttentionLayers(Module):
         norm_fn = partial(norm_class, dim)
 
         if not norm_need_condition and norm_add_unit_offset:
-            # research Ohad Rubin shares in a blog post by adding an offset to gammas and betas, they can be subjected to weight decay safely
+            # researcher Ohad Rubin shares in a blog post by adding an offset to gammas and betas, they can be subjected to weight decay safely
             norm_fn = partial(norm_fn, unit_offset = 1.)
 
         self.norm_need_condition = norm_need_condition
@@ -1379,6 +1403,9 @@ class AttentionLayers(Module):
 
         self.post_branch_fn_needs_condition = post_branch_fn_needs_condition
 
+        if not post_branch_fn_needs_condition and norm_add_unit_offset:
+            post_branch_fn = partial(post_branch_fn, unit_offset = 1.)
+
         # setup mlp for conditioning
 
         self.need_condition = norm_need_condition or post_branch_fn_needs_condition

{x_transformers-1.31.0.dist-info → x_transformers-1.31.2.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: x-transformers
-Version: 1.31.0
+Version: 1.31.2
 Summary: X-Transformers - Pytorch
 Home-page: https://github.com/lucidrains/x-transformers
 Author: Phil Wang

{x_transformers-1.31.0.dist-info → x_transformers-1.31.2.dist-info}/RECORD

@@ -4,11 +4,11 @@ x_transformers/autoregressive_wrapper.py,sha256=uX8Mb0zLsQrZECt_9UGt35g7tC05Rk3n
 x_transformers/continuous.py,sha256=WO52n9lFAXv5-SGadi2cApGF8dkouN8QSTEOuC7erj8,6180
 x_transformers/dpo.py,sha256=LjvWgCkqTl-UuehrzQ8nkX5guLr4whYwsmm7SKSwdls,3450
 x_transformers/nonautoregressive_wrapper.py,sha256=ys_p8obc7lTeeodCqvkRKxOXQ1C9T3j5Jwr-JbVgnXk,10432
-x_transformers/x_transformers.py,sha256=Lbfp0VHMCMkZboYficEe8d-X7u55uOh4lQbvIz_db4c,75233
+x_transformers/x_transformers.py,sha256=aeFAk19BRxkLOS-pUgtDk3HaMS6_dsuJtz2o_KkUkOU,75805
 x_transformers/xl_autoregressive_wrapper.py,sha256=DCx4n0_c1tFai4nOqaWVnqx2p9eutsZsDMiMP1ckxNU,4117
 x_transformers/xval.py,sha256=QE1ltYZTR_eGgIHPP2BrMWVWVLqMW-OpDZh87BSmQEg,8563
-x_transformers-1.31.0.dist-info/LICENSE,sha256=As9u198X-U-vph5noInuUfqsAG2zX_oXPHDmdjwlPPY,1066
-x_transformers-1.31.0.dist-info/METADATA,sha256=0-TZQ6Y-FpgOjF3z3GIO2nPRX5__9lhe5N6Qct5U79Y,661
-x_transformers-1.31.0.dist-info/WHEEL,sha256=mguMlWGMX-VHnMpKOjjQidIo1ssRlCFu4a4mBpz1s2M,91
-x_transformers-1.31.0.dist-info/top_level.txt,sha256=hO6KGpFuGucRNEtRfme4A_rGcM53AKwGP7RVlRIxS5Q,15
-x_transformers-1.31.0.dist-info/RECORD,,
+x_transformers-1.31.2.dist-info/LICENSE,sha256=As9u198X-U-vph5noInuUfqsAG2zX_oXPHDmdjwlPPY,1066
+x_transformers-1.31.2.dist-info/METADATA,sha256=dxDZYzaZwwOCD-9I6ver-SSLlSG4diss-ovDo2H_9xE,661
+x_transformers-1.31.2.dist-info/WHEEL,sha256=mguMlWGMX-VHnMpKOjjQidIo1ssRlCFu4a4mBpz1s2M,91
+x_transformers-1.31.2.dist-info/top_level.txt,sha256=hO6KGpFuGucRNEtRfme4A_rGcM53AKwGP7RVlRIxS5Q,15
+x_transformers-1.31.2.dist-info/RECORD,,