seeing really big improvements with per token learned value residual …

…mixing values

setup.py

@@ -3,7 +3,7 @@
 setup(
   name = 'x-transformers',
   packages = find_packages(exclude=['examples']),
-  version = '1.42.12',
+  version = '1.42.15',
   license='MIT',
   description = 'X-Transformers - Pytorch',
   author = 'Phil Wang',

tests/test_x_transformers.py

@@ -331,7 +331,10 @@ def test_reinject_input():
 
     model(x) # (1, 1024, 20000)
 
-def test_value_residual():
+@pytest.mark.parametrize('learned_value_residual_mix', (False, True))
+def test_value_residual(
+    learned_value_residual_mix: bool
+):
 
     model = TransformerWrapper(
         num_tokens = 20000,
@@ -341,6 +344,7 @@ def test_value_residual():
             depth = 6,
             heads = 8,
             add_value_residual = True,
+            learned_value_residual_mix = learned_value_residual_mix
         )
     )
 

x_transformers/x_transformers.py

@@ -1072,6 +1072,7 @@ def __init__(
         logit_softclamp_value = 50.,
         neutreno_value_residual = False, # Nguyen et al. https://arxiv.org/abs/2312.00751
         neutreno_alpha = 0.4,
+        learned_value_residual_mix = False,
         onnxable = False,
         attend_sdp_kwargs: dict = dict(
             enable_flash = True,
@@ -1231,6 +1232,14 @@ def __init__(
             self.mem_k = nn.Parameter(torch.randn(kv_heads, num_mem_kv, dim_head))
             self.mem_v = nn.Parameter(torch.randn(kv_heads, num_mem_kv, dim_head))
 
+        # maybe learned value residual mixer per token
+
+        self.to_value_residual_mix = nn.Sequential(
+            nn.Linear(dim, 1),
+            nn.Sigmoid(),
+            Rearrange('b n 1 -> b 1 n 1')
+         ) if learned_value_residual_mix else always(0.5)
+
         # attention on attention
 
         self.attn_on_attn = on_attn
@@ -1303,7 +1312,8 @@ def forward(
                 diff_values = repeat(diff_values, 'b h n d -> b (r h) n d', r = h // kv_h)
             else:
                 # https://arxiv.org/abs/2410.17897v1
-                v = 0.5 * (v + value_residual)
+                value_residual_mix = self.to_value_residual_mix(q_input)
+                v = v * value_residual_mix + value_residual * (1. - value_residual_mix)
 
         # take care of caching
 
@@ -1541,8 +1551,9 @@ def __init__(
         use_layerscale = False,
         layerscale_init_value = 0.,
         unet_skips = False,
-        reinject_input = False,         # seen first in DEQ paper https://arxiv.org/abs/1909.01377, but later used in a number of papers trying to achieve depthwise generalization https://arxiv.org/abs/2410.03020v1
-        add_value_residual = False,     # resformer from Zhou et al - https://arxiv.org/abs/2410.17897v1
+        reinject_input = False,              # seen first in DEQ paper https://arxiv.org/abs/1909.01377, but later used in a number of papers trying to achieve depthwise generalization https://arxiv.org/abs/2410.03020v1
+        add_value_residual = False,          # resformer from Zhou et al - https://arxiv.org/abs/2410.17897v1
+        learned_value_residual_mix = False,  # seeing big improvements when the value residual mix value is learned per token - credit goes to @faresobeid for taking the first step with learned scalar mix, then @Blinkdl for taking it a step further with data dependent. here we will use per token learned
         rel_pos_kwargs: dict = dict(),
         **kwargs
     ):
@@ -1786,6 +1797,10 @@ def __init__(
 
         self.add_value_residual = add_value_residual
 
+        is_first_self_attn = True
+        is_first_cross_attn = True
+        learned_value_residual_mix &= add_value_residual
+
         # iterate and construct layers
 
         for ind, (layer_type, layer_shift_tokens) in enumerate(zip(self.layer_types, shift_tokens)):
@@ -1801,9 +1816,13 @@ def __init__(
             # attention, cross attention, feedforward
 
             if layer_type == 'a':
-                layer = Attention(dim, heads = heads, causal = causal, **attn_kwargs)
+                self_attn_learned_value_residual = learned_value_residual_mix and not is_first_self_attn
+                layer = Attention(dim, heads = heads, causal = causal, learned_value_residual_mix = self_attn_learned_value_residual, **attn_kwargs)
+                is_first_self_attn = False
             elif layer_type == 'c':
-                layer = Attention(dim, heads = heads, **{**attn_kwargs, **cross_attn_kwargs})
+                cross_attn_learned_value_residual = learned_value_residual_mix and not is_first_cross_attn
+                layer = Attention(dim, heads = heads, learned_value_residual_mix = cross_attn_learned_value_residual, **{**attn_kwargs, **cross_attn_kwargs})
+                is_first_cross_attn = False
             elif layer_type == 'f':
                 layer = FeedForward(dim, **ff_kwargs)
                 layer = layer if not macaron else Scale(0.5, layer)