A few more maybe_add_mask situations

Author: rwightman

timm/layers/attention_pool.py

@@ -4,6 +4,7 @@
 import torch.nn as nn
 import torch.nn.functional as F
 
+from .attention import maybe_add_mask
 from .config import use_fused_attn
 from .mlp import Mlp
 from .weight_init import trunc_normal_tf_
@@ -95,8 +96,7 @@ def forward(self, x, attn_mask: Optional[torch.Tensor] = None):
         else:
             q = q * self.scale
             attn = q @ k.transpose(-2, -1)
-            if attn_mask is not None:
-                attn = attn + attn_mask
+            attn = maybe_add_mask(attn, attn_mask)
             attn = attn.softmax(dim=-1)
             x = attn @ v
         x = x.transpose(1, 2).reshape(B, self.latent_len, C)

timm/models/vision_transformer.py

@@ -43,7 +43,7 @@
     OPENAI_CLIP_MEAN, OPENAI_CLIP_STD
 from timm.layers import Attention, PatchEmbed, Mlp, DropPath, AttentionPoolLatent, RmsNorm, PatchDropout, \
     SwiGLUPacked, SwiGLU, trunc_normal_, lecun_normal_, resample_patch_embed, resample_abs_pos_embed, use_fused_attn, \
-    get_act_layer, get_norm_layer, LayerType
+    get_act_layer, get_norm_layer, LayerType, maybe_add_mask
 from ._builder import build_model_with_cfg
 from ._features import feature_take_indices
 from ._manipulate import named_apply, checkpoint_seq, adapt_input_conv
@@ -256,8 +256,7 @@ def forward(self, x: torch.Tensor, attn_mask: Optional[torch.Tensor] = None) ->
         else:
             q = q * self.scale
             attn = q @ k.transpose(-2, -1)
-            if attn_mask is not None:
-                attn = attn + attn_mask
+            attn = maybe_add_mask(attn, attn_mask)
             attn = attn.softmax(dim=-1)
             attn = self.attn_drop(attn)
             x_attn = attn @ v

timm/models/vision_transformer_flex.py

@@ -823,7 +823,7 @@ def forward_features(
             attn_mask: Optional[torch.Tensor] = None,
     ) -> torch.Tensor:
 
-        if attn_mask is None and patch_valid is not None:
+        if attn_mask is None:
             attn_mask = create_attention_mask(
                 patch_valid,
                 num_prefix_tokens=self.num_prefix_tokens,