Move levit style pos bias resize with other rel pos bias utils

Author: rwightman

timm/layers/__init__.py

@@ -36,9 +36,9 @@
 from .patch_dropout import PatchDropout
 from .patch_embed import PatchEmbed, PatchEmbedWithSize, resample_patch_embed
 from .pool2d_same import AvgPool2dSame, create_pool2d
-from .pos_embed import resample_abs_pos_embed, resample_abs_pos_embed_nhwc, resample_relative_position_bias_table
+from .pos_embed import resample_abs_pos_embed, resample_abs_pos_embed_nhwc
 from .pos_embed_rel import RelPosMlp, RelPosBias, RelPosBiasTf, gen_relative_position_index, gen_relative_log_coords, \
-    resize_rel_pos_bias_table, resize_rel_pos_bias_table_simple
+    resize_rel_pos_bias_table, resize_rel_pos_bias_table_simple, resize_rel_pos_bias_table_levit
 from .pos_embed_sincos import pixel_freq_bands, freq_bands, build_sincos2d_pos_embed, build_fourier_pos_embed, \
     build_rotary_pos_embed, apply_rot_embed, apply_rot_embed_cat, apply_rot_embed_list, apply_keep_indices_nlc, \
     FourierEmbed, RotaryEmbedding, RotaryEmbeddingCat

timm/layers/pos_embed.py

@@ -78,38 +78,3 @@ def resample_abs_pos_embed_nhwc(
         _logger.info(f'Resized position embedding: {posemb.shape[-3:-1]} to {new_size}.')
 
     return posemb
-
-
-def resample_relative_position_bias_table(
-        position_bias_table,
-        new_size,
-        interpolation: str = 'bicubic',
-        antialias: bool = True,
-        verbose: bool = False
-):
-    """
-    Resample relative position bias table suggested in LeVit
-    Adapted from: https://github.com/microsoft/Cream/blob/main/TinyViT/utils.py
-    """
-    L1, nH1 = position_bias_table.size()
-    L2, nH2 = new_size
-    assert nH1 == nH2
-    if L1 != L2:
-        orig_dtype = position_bias_table.dtype
-        position_bias_table = position_bias_table.float()
-        # bicubic interpolate relative_position_bias_table if not match
-        S1 = int(L1 ** 0.5)
-        S2 = int(L2 ** 0.5)
-        relative_position_bias_table_resized = F.interpolate(
-            position_bias_table.permute(1, 0).view(1, nH1, S1, S1),
-            size=(S2, S2),
-            mode=interpolation,
-            antialias=antialias)
-        relative_position_bias_table_resized = \
-            relative_position_bias_table_resized.view(nH2, L2).permute(1, 0)
-        relative_position_bias_table_resized.to(orig_dtype)
-        if not torch.jit.is_scripting() and verbose:
-            _logger.info(f'Resized position bias: {L1, nH1} to {L2, nH2}.')
-        return relative_position_bias_table_resized
-    else:
-        return position_bias_table

timm/layers/pos_embed_rel.py

@@ -121,6 +121,38 @@ def resize_rel_pos_bias_table_simple(
     return rel_pos_bias
 
 
+def resize_rel_pos_bias_table_levit(
+        position_bias_table,
+        new_size,
+        interpolation: str = 'bicubic',
+        antialias: bool = True,
+):
+    """
+    Resample relative position bias table suggested in LeVit
+    Adapted from: https://github.com/microsoft/Cream/blob/main/TinyViT/utils.py
+    """
+    L1, nH1 = position_bias_table.size()
+    L2, nH2 = new_size
+    assert nH1 == nH2
+    if L1 != L2:
+        orig_dtype = position_bias_table.dtype
+        position_bias_table = position_bias_table.float()
+        # bicubic interpolate relative_position_bias_table if not match
+        S1 = int(L1 ** 0.5)
+        S2 = int(L2 ** 0.5)
+        relative_position_bias_table_resized = F.interpolate(
+            position_bias_table.permute(1, 0).view(1, nH1, S1, S1),
+            size=(S2, S2),
+            mode=interpolation,
+            antialias=antialias)
+        relative_position_bias_table_resized = \
+            relative_position_bias_table_resized.view(nH2, L2).permute(1, 0)
+        relative_position_bias_table_resized.to(orig_dtype)
+        return relative_position_bias_table_resized
+    else:
+        return position_bias_table
+
+
 def resize_rel_pos_bias_table(
         rel_pos_bias,
         new_window_size: Tuple[int, int],

timm/models/tiny_vit.py

@@ -19,7 +19,7 @@
 
 from timm.data import IMAGENET_DEFAULT_MEAN, IMAGENET_DEFAULT_STD
 from timm.layers import LayerNorm2d, NormMlpClassifierHead, DropPath,\
-    to_2tuple, trunc_normal_, resample_relative_position_bias_table, use_fused_attn
+    trunc_normal_, resize_rel_pos_bias_table_levit, use_fused_attn
 from ._builder import build_model_with_cfg
 from ._manipulate import checkpoint_seq
 from ._registry import register_model, generate_default_cfgs
@@ -182,6 +182,7 @@ def __init__(
         self.d = int(attn_ratio * key_dim)
         self.dh = int(attn_ratio * key_dim) * num_heads
         self.attn_ratio = attn_ratio
+        self.resolution = resolution
         self.fused_attn = use_fused_attn()
 
         h = self.dh + nh_kd * 2
@@ -551,17 +552,18 @@ def checkpoint_filter_fn(state_dict, model):
     # TODO: temporary use for testing, need change after weight convert
     if 'model' in state_dict.keys():
         state_dict = state_dict['model']
-    targe_sd = model.state_dict()
-    target_keys = list(targe_sd.keys())
+    target_sd = model.state_dict()
+    target_keys = list(target_sd.keys())
     out_dict = {}
     i = 0
     for k, v in state_dict.items():
-        if not k.endswith('attention_bias_idxs'):
-            if 'attention_biases' in k:
-                # dynamic window size by resampling relative_position_bias_table
-                # TODO: whether move this func into model for dynamic input resolution? (high risk)
-                v = resample_relative_position_bias_table(v.T, targe_sd[target_keys[i]].shape[::-1]).T
-            out_dict[target_keys[i]] = v
+        if k.endswith('attention_bias_idxs'):
+            continue
+        tk = target_keys[i]
+        if 'attention_biases' in k:
+            # TODO: whether move this func into model for dynamic input resolution? (high risk)
+            v = resize_rel_pos_bias_table_levit(v.T, target_sd[tk].shape[::-1]).T
+        out_dict[tk] = v
         i += 1
     return out_dict