Initial pass of an 'init_non_persistent_buffers' scheme, WIP.. needs more test and probably missed a few things

Author: rwightman

timm/layers/__init__.py

@@ -150,6 +150,7 @@
 from .trace_utils import _assert, _float_to_int
 from .typing import LayerType, PadType, disable_compiler
 from .weight_init import (
+    is_meta_device,
     trunc_normal_,
     trunc_normal_tf_,
     variance_scaling_,

timm/layers/blur_pool.py

@@ -46,22 +46,24 @@ def __init__(
         self.stride = stride
         self.pad_mode = pad_mode
         self.padding = [get_padding(filt_size, stride, dilation=1)] * 4
+        self.register_buffer(
+            'filt',
+            self._create_blur_filter(device=device, dtype=dtype),
+            persistent=False,
+        )
 
+    def _create_blur_filter(self, device=None, dtype=None) -> torch.Tensor:
+        """Create the blur filter tensor."""
         # (0.5 + 0.5 x)^N => coefficients = C(N,k) / 2^N,  k = 0..N
         coeffs = torch.tensor(
-            [comb(filt_size - 1, k) for k in range(filt_size)],
+            [comb(self.filt_size - 1, k) for k in range(self.filt_size)],
             device='cpu',
             dtype=torch.float32,
-        ) / (2 ** (filt_size - 1))  # normalise so coefficients sum to 1
+        ) / (2 ** (self.filt_size - 1))  # normalise so coefficients sum to 1
         blur_filter = (coeffs[:, None] * coeffs[None, :])[None, None, :, :]
-        if channels is not None:
+        if self.channels is not None:
             blur_filter = blur_filter.repeat(self.channels, 1, 1, 1)
-
-        self.register_buffer(
-            'filt',
-            blur_filter.to(device=device, dtype=dtype),
-            persistent=False,
-        )
+        return blur_filter.to(device=device, dtype=dtype)
 
     def forward(self, x: torch.Tensor) -> torch.Tensor:
         x = F.pad(x, self.padding, mode=self.pad_mode)
@@ -73,6 +75,14 @@ def forward(self, x: torch.Tensor) -> torch.Tensor:
             weight = self.filt
         return F.conv2d(x, weight, stride=self.stride, groups=channels)
 
+    def init_non_persistent_buffers(
+            self,
+            device: torch.device,
+            dtype: torch.dtype,
+    ) -> None:
+        """Initialize non-persistent buffers."""
+        self.filt.data = self._create_blur_filter(device=device, dtype=dtype)
+
 
 def _normalize_aa_layer(aa_layer: LayerType) -> Callable[..., nn.Module]:
     """Map string shorthands to callables (class or partial)."""

timm/layers/lambda_layer.py

@@ -156,3 +156,16 @@ def forward(self, x):
         out = (content_out + position_out).transpose(-1, -2).reshape(B, C, H, W)  # B, C (num_heads * V), H, W
         out = self.pool(out)
         return out
+
+    def init_non_persistent_buffers(
+            self,
+            device: Optional[torch.device] = None,
+            dtype: Optional[torch.dtype] = None,
+    ) -> None:
+        """Initialize non-persistent buffers."""
+        if self.rel_pos_indices is None:
+            return
+        device = device or self.qkv.weight.device
+        # Compute feat_size from pos_emb shape: rel_size = 2 * feat_size - 1
+        feat_size = ((self.pos_emb.shape[0] + 1) // 2, (self.pos_emb.shape[1] + 1) // 2)
+        self.rel_pos_indices.data = rel_pos_indices(feat_size, device=device)

timm/layers/pos_embed_rel.py

@@ -311,6 +311,18 @@ def get_bias(self) -> torch.Tensor:
     def forward(self, attn, shared_rel_pos: Optional[torch.Tensor] = None):
         return attn + self.get_bias()
 
+    def init_non_persistent_buffers(
+            self,
+            device: Optional[torch.device] = None,
+            dtype: Optional[torch.dtype] = None,
+    ) -> None:
+        """Initialize non-persistent buffers."""
+        device = device or self.relative_position_bias_table.device
+        prefix_tokens = 1 if self.bias_shape[0] > self.window_area else 0
+        self.relative_position_index.data = gen_relative_position_index(
+            self.window_size, class_token=prefix_tokens > 0, device=device
+        ).view(-1)
+
 
 def gen_relative_log_coords(
         win_size: Tuple[int, int],
@@ -367,6 +379,8 @@ def __init__(
         self.prefix_tokens = prefix_tokens
         self.num_heads = num_heads
         self.bias_shape = (self.window_area,) * 2 + (num_heads,)
+        self.mode = mode
+        self.pretrained_window_size = pretrained_window_size
         if mode == 'swin':
             self.bias_act = nn.Sigmoid()
             self.bias_gain = 16
@@ -415,6 +429,25 @@ def get_bias(self) -> torch.Tensor:
     def forward(self, attn, shared_rel_pos: Optional[torch.Tensor] = None):
         return attn + self.get_bias()
 
+    def init_non_persistent_buffers(
+            self,
+            device: Optional[torch.device] = None,
+            dtype: Optional[torch.dtype] = None,
+    ) -> None:
+        """Initialize non-persistent buffers."""
+        device = device or self.mlp.fc1.weight.device
+        dtype = dtype or self.mlp.fc1.weight.dtype
+        self.relative_position_index.data = gen_relative_position_index(
+            self.window_size, device=device
+        ).view(-1)
+        self.rel_coords_log.data = gen_relative_log_coords(
+            self.window_size,
+            self.pretrained_window_size,
+            mode=self.mode,
+            device=device,
+            dtype=dtype,
+        )
+
 
 def generate_lookup_tensor(
         length: int,
@@ -519,3 +552,14 @@ def get_bias(self) -> torch.Tensor:
 
     def forward(self, attn, shared_rel_pos: Optional[torch.Tensor] = None):
         return attn + self.get_bias()
+
+    def init_non_persistent_buffers(
+            self,
+            device: Optional[torch.device] = None,
+            dtype: Optional[torch.dtype] = None,
+    ) -> None:
+        """Initialize non-persistent buffers."""
+        device = device or self.relative_position_bias_table.device
+        dtype = dtype or self.relative_position_bias_table.dtype
+        self.height_lookup.data = generate_lookup_tensor(self.window_size[0], device=device, dtype=dtype)
+        self.width_lookup.data = generate_lookup_tensor(self.window_size[1], device=device, dtype=dtype)