Merge pull request #148 from rwightman/drop_block_improve

Improve dropblock impl, add fast variant, better AMP speed, inplace…

timm/models/layers/drop.py

@@ -22,44 +22,89 @@ import math
 
 
 def drop_block_2d(
-        x, drop_prob: float = 0.1, training: bool = False, block_size: int = 7,
-        gamma_scale: float = 1.0, drop_with_noise: bool = False):
+        x, drop_prob: float = 0.1, block_size: int = 7,  gamma_scale: float = 1.0,
+        with_noise: bool = False, inplace: bool = False, batchwise: bool = False):
     """ DropBlock. See https://arxiv.org/pdf/1810.12890.pdf
 
     DropBlock with an experimental gaussian noise option. This layer has been tested on a few training
     runs with success, but needs further validation and possibly optimization for lower runtime impact.
-
     """
-    if drop_prob == 0. or not training:
-        return x
-    _, _, height, width = x.shape
-    total_size = width * height
-    clipped_block_size = min(block_size, min(width, height))
+    B, C, H, W = x.shape
+    total_size = W * H
+    clipped_block_size = min(block_size, min(W, H))
     # seed_drop_rate, the gamma parameter
-    seed_drop_rate = gamma_scale * drop_prob * total_size / clipped_block_size ** 2 / (
-            (width - block_size + 1) *
-            (height - block_size + 1))
+    gamma = gamma_scale * drop_prob * total_size / clipped_block_size ** 2 / (
+        (W - block_size + 1) * (H - block_size + 1))
 
     # Forces the block to be inside the feature map.
-    w_i, h_i = torch.meshgrid(torch.arange(width).to(x.device), torch.arange(height).to(x.device))
-    valid_block = ((w_i >= clipped_block_size // 2) & (w_i < width - (clipped_block_size - 1) // 2)) & \
-                  ((h_i >= clipped_block_size // 2) & (h_i < height - (clipped_block_size - 1) // 2))
-    valid_block = torch.reshape(valid_block, (1, 1, height, width)).float()
-
-    uniform_noise = torch.rand_like(x, dtype=torch.float32)
-    block_mask = ((2 - seed_drop_rate - valid_block + uniform_noise) >= 1).float()
+    w_i, h_i = torch.meshgrid(torch.arange(W).to(x.device), torch.arange(H).to(x.device))
+    valid_block = ((w_i >= clipped_block_size // 2) & (w_i < W - (clipped_block_size - 1) // 2)) & \
+                  ((h_i >= clipped_block_size // 2) & (h_i < H - (clipped_block_size - 1) // 2))
+    valid_block = torch.reshape(valid_block, (1, 1, H, W)).to(dtype=x.dtype)
+
+    if batchwise:
+        # one mask for whole batch, quite a bit faster
+        uniform_noise = torch.rand((1, C, H, W), dtype=x.dtype, device=x.device)
+    else:
+        uniform_noise = torch.rand_like(x)
+    block_mask = ((2 - gamma - valid_block + uniform_noise) >= 1).to(dtype=x.dtype)
     block_mask = -F.max_pool2d(
         -block_mask,
-        kernel_size=clipped_block_size,  # block_size, ???
+        kernel_size=clipped_block_size,  # block_size,
         stride=1,
         padding=clipped_block_size // 2)
 
-    if drop_with_noise:
-        normal_noise = torch.randn_like(x)
-        x = x * block_mask + normal_noise * (1 - block_mask)
+    if with_noise:
+        normal_noise = torch.randn((1, C, H, W), dtype=x.dtype, device=x.device) if batchwise else torch.randn_like(x)
+        if inplace:
+            x.mul_(block_mask).add_(normal_noise * (1 - block_mask))
+        else:
+            x = x * block_mask + normal_noise * (1 - block_mask)
+    else:
+        normalize_scale = (block_mask.numel() / block_mask.to(dtype=torch.float32).sum().add(1e-7)).to(x.dtype)
+        if inplace:
+            x.mul_(block_mask * normalize_scale)
+        else:
+            x = x * block_mask * normalize_scale
+    return x
+
+
+def drop_block_fast_2d(
+        x: torch.Tensor, drop_prob: float = 0.1, block_size: int = 7,
+        gamma_scale: float = 1.0, with_noise: bool = False, inplace: bool = False, batchwise: bool = False):
+    """ DropBlock. See https://arxiv.org/pdf/1810.12890.pdf
+
+    DropBlock with an experimental gaussian noise option. Simplied from above without concern for valid
+    block mask at edges.
+    """
+    B, C, H, W = x.shape
+    total_size = W * H
+    clipped_block_size = min(block_size, min(W, H))
+    gamma = gamma_scale * drop_prob * total_size / clipped_block_size ** 2 / (
+            (W - block_size + 1) * (H - block_size + 1))
+
+    if batchwise:
+        # one mask for whole batch, quite a bit faster
+        block_mask = torch.rand((1, C, H, W), dtype=x.dtype, device=x.device) < gamma
+    else:
+        # mask per batch element
+        block_mask = torch.rand_like(x) < gamma
+    block_mask = F.max_pool2d(
+        block_mask.to(x.dtype), kernel_size=clipped_block_size, stride=1, padding=clipped_block_size // 2)
+
+    if with_noise:
+        normal_noise = torch.randn((1, C, H, W), dtype=x.dtype, device=x.device) if batchwise else torch.randn_like(x)
+        if inplace:
+            x.mul_(1. - block_mask).add_(normal_noise * block_mask)
+        else:
+            x = x * (1. - block_mask) + normal_noise * block_mask
     else:
-        normalize_scale = block_mask.numel() / (torch.sum(block_mask) + 1e-7)
-        x = x * block_mask * normalize_scale
+        block_mask = 1 - block_mask
+        normalize_scale = (block_mask.numel() / block_mask.to(dtype=torch.float32).sum().add(1e-7)).to(dtype=x.dtype)
+        if inplace:
+            x.mul_(block_mask * normalize_scale)
+        else:
+            x = x * block_mask * normalize_scale
     return x
 
 
@@ -70,15 +115,28 @@ class DropBlock2d(nn.Module):
                  drop_prob=0.1,
                  block_size=7,
                  gamma_scale=1.0,
-                 with_noise=False):
+                 with_noise=False,
+                 inplace=False,
+                 batchwise=False,
+                 fast=True):
         super(DropBlock2d, self).__init__()
         self.drop_prob = drop_prob
         self.gamma_scale = gamma_scale
         self.block_size = block_size
         self.with_noise = with_noise
+        self.inplace = inplace
+        self.batchwise = batchwise
+        self.fast = fast  # FIXME finish comparisons of fast vs not
 
     def forward(self, x):
-        return drop_block_2d(x, self.drop_prob, self.training, self.block_size, self.gamma_scale, self.with_noise)
+        if not self.training or not self.drop_prob:
+            return x
+        if self.fast:
+            return drop_block_fast_2d(
+                x, self.drop_prob, self.block_size, self.gamma_scale, self.with_noise, self.inplace, self.batchwise)
+        else:
+            return drop_block_2d(
+                x, self.drop_prob, self.block_size, self.gamma_scale, self.with_noise, self.inplace, self.batchwise)
 
 
 def drop_path(x, drop_prob: float = 0., training: bool = False):

timm/models/layers/split_attn.py

@@ -31,25 +31,24 @@ class RadixSoftmax(nn.Module):
 class SplitAttnConv2d(nn.Module):
     """Split-Attention Conv2d
     """
-    def __init__(self, in_channels, channels, kernel_size, stride=1, padding=0,
+    def __init__(self, in_channels, out_channels, kernel_size, stride=1, padding=0,
                  dilation=1, groups=1, bias=False, radix=2, reduction_factor=4,
                  act_layer=nn.ReLU, norm_layer=None, drop_block=None, **kwargs):
         super(SplitAttnConv2d, self).__init__()
         self.radix = radix
-        self.cardinality = groups
-        self.channels = channels
-        mid_chs = channels * radix
+        self.drop_block = drop_block
+        mid_chs = out_channels * radix
         attn_chs = max(in_channels * radix // reduction_factor, 32)
+
         self.conv = nn.Conv2d(
             in_channels, mid_chs, kernel_size, stride, padding, dilation,
             groups=groups * radix, bias=bias, **kwargs)
         self.bn0 = norm_layer(mid_chs) if norm_layer is not None else None
         self.act0 = act_layer(inplace=True)
-        self.fc1 = nn.Conv2d(channels, attn_chs, 1, groups=self.cardinality)
+        self.fc1 = nn.Conv2d(out_channels, attn_chs, 1, groups=groups)
         self.bn1 = norm_layer(attn_chs) if norm_layer is not None else None
         self.act1 = act_layer(inplace=True)
-        self.fc2 = nn.Conv2d(attn_chs, mid_chs, 1, groups=self.cardinality)
-        self.drop_block = drop_block
+        self.fc2 = nn.Conv2d(attn_chs, mid_chs, 1, groups=groups)
         self.rsoftmax = RadixSoftmax(radix, groups)
 
     def forward(self, x):
@@ -63,7 +62,7 @@ class SplitAttnConv2d(nn.Module):
         B, RC, H, W = x.shape
         if self.radix > 1:
             x = x.reshape((B, self.radix, RC // self.radix, H, W))
-            x_gap = torch.sum(x, dim=1)
+            x_gap = x.sum(dim=1)
         else:
             x_gap = x
         x_gap = F.adaptive_avg_pool2d(x_gap, 1)

timm/models/resnest.py

@@ -76,10 +76,10 @@ class ResNestBottleneck(nn.Module):
         else:
             avd_stride = 0
         self.radix = radix
+        self.drop_block = drop_block
 
         self.conv1 = nn.Conv2d(inplanes, group_width, kernel_size=1, bias=False)
         self.bn1 = norm_layer(group_width)
-        self.drop_block1 = drop_block if drop_block is not None else None
         self.act1 = act_layer(inplace=True)
         self.avd_first = nn.AvgPool2d(3, avd_stride, padding=1) if avd_stride > 0 and avd_first else None
 
@@ -88,20 +88,17 @@ class ResNestBottleneck(nn.Module):
                 group_width, group_width, kernel_size=3, stride=stride, padding=first_dilation,
                 dilation=first_dilation, groups=cardinality, radix=radix, norm_layer=norm_layer, drop_block=drop_block)
             self.bn2 = None  # FIXME revisit, here to satisfy current torchscript fussyness
-            self.drop_block2 = None
             self.act2 = None
         else:
             self.conv2 = nn.Conv2d(
                 group_width, group_width, kernel_size=3, stride=stride, padding=first_dilation,
                 dilation=first_dilation, groups=cardinality, bias=False)
             self.bn2 = norm_layer(group_width)
-            self.drop_block2 = drop_block if drop_block is not None else None
             self.act2 = act_layer(inplace=True)
         self.avd_last = nn.AvgPool2d(3, avd_stride, padding=1) if avd_stride > 0 and not avd_first else None
 
         self.conv3 = nn.Conv2d(group_width, planes * 4, kernel_size=1, bias=False)
         self.bn3 = norm_layer(planes*4)
-        self.drop_block3 = drop_block if drop_block is not None else None
         self.act3 = act_layer(inplace=True)
         self.downsample = downsample
 
@@ -113,8 +110,8 @@ class ResNestBottleneck(nn.Module):
 
         out = self.conv1(x)
         out = self.bn1(out)
-        if self.drop_block1 is not None:
-            out = self.drop_block1(out)
+        if self.drop_block is not None:
+            out = self.drop_block(out)
         out = self.act1(out)
 
         if self.avd_first is not None:
@@ -123,8 +120,8 @@ class ResNestBottleneck(nn.Module):
         out = self.conv2(out)
         if self.bn2 is not None:
             out = self.bn2(out)
-            if self.drop_block2 is not None:
-                out = self.drop_block2(out)
+            if self.drop_block is not None:
+                out = self.drop_block(out)
             out = self.act2(out)
 
         if self.avd_last is not None:
@@ -132,8 +129,8 @@ class ResNestBottleneck(nn.Module):
 
         out = self.conv3(out)
         out = self.bn3(out)
-        if self.drop_block3 is not None:
-            out = self.drop_block3(out)
+        if self.drop_block is not None:
+            out = self.drop_block(out)
 
         if self.downsample is not None:
             residual = self.downsample(x)