add stochastic depth to resnet

pytorch_tools/models/resnet.py

@@ -70,6 +70,9 @@ class ResNet(nn.Module):
             Flag to overwrite forward pass to return 5 tensors with different resolutions. Defaults to False.
         drop_rate (float):
             Dropout probability before classifier, for training. Defaults to 0.0.
+        drop_connect_rate (float):
+            Drop rate for StochasticDepth. Randomly removes samples each block. Used as regularization during training. 
+            keep prob will be linearly decreased from 1 to 1 - drop_connect_rate each block. Ref: https://arxiv.org/abs/1603.09382
         global_pool (str):
             Global pooling type. One of 'avg', 'max', 'avgmax', 'catavgmax'. Defaults to 'avg'.
         init_bn0 (bool):
@@ -95,6 +98,7 @@ def __init__(
         antialias=False,
         encoder=False,
         drop_rate=0.0,
+        drop_connect_rate=0.0,
         global_pool="avg",
         init_bn0=True,
     ):
@@ -108,6 +112,9 @@ def __init__(
         self.block = block
         self.expansion = block.expansion
         self.norm_act = norm_act
+        self.block_idx = 0
+        self.num_blocks = sum(layers)
+        self.drop_connect_rate = drop_connect_rate
         super(ResNet, self).__init__()
 
         if deep_stem:
@@ -185,6 +192,7 @@ def _make_layer(
                 norm_layer=norm_layer,
                 norm_act=norm_act,
                 antialias=antialias,
+                keep_prob=self.keep_prob,
             )
         ]
 
@@ -201,6 +209,7 @@ def _make_layer(
                     norm_layer=norm_layer,
                     norm_act=norm_act,
                     antialias=antialias,
+                    keep_prob=self.keep_prob,
                 )
             )
         return nn.Sequential(*layers)
@@ -266,6 +275,11 @@ def load_state_dict(self, state_dict, **kwargs):
                 state_dict[k.replace("layer0.", "")] = state_dict.pop(k)
         super().load_state_dict(state_dict, **kwargs)
 
+    @property
+    def keep_prob(self):
+        keep_prob = 1 - self.drop_connect_rate * self.block_idx / self.num_blocks
+        self.block_idx += 1
+        return keep_prob
 
 # fmt: off
 CFGS = {

pytorch_tools/models/tresnet.py

@@ -51,6 +51,8 @@ class TResNet(ResNet):
             Flag to overwrite forward pass to return 5 tensors with different resolutions. Defaults to False.
         drop_rate (float):
             Dropout probability before classifier, for training. Defaults to 0.0. to 'avg'.
+        drop_connect_rate (float):
+            Drop rate for StochasticDepth. Randomly removes samples each block. Used as regularization during training. Ref: https://arxiv.org/abs/1603.09382
     """
 
     def __init__(
@@ -65,6 +67,7 @@ def __init__(
         norm_act="leaky_relu",
         encoder=False,
         drop_rate=0.0,
+        drop_connect_rate=0.0,
     ):
         nn.Module.__init__(self) 
         stem_width = int(64 * width_factor)
@@ -74,6 +77,9 @@ def __init__(
         self.groups = 1 # not really used but needed inside _make_layer
         self.base_width = 64 # used inside _make_layer
         self.norm_act = norm_act
+        self.block_idx = 0
+        self.num_blocks = sum(layers)
+        self.drop_connect_rate = drop_connect_rate
 
         # in the paper they use conv1x1 but in code conv3x3 (which seems better)
         self.conv1 = nn.Sequential(SpaceToDepth(), conv3x3(in_channels * 16, stem_width))

pytorch_tools/modules/residual.py

@@ -151,6 +151,7 @@ def __init__(
         norm_layer=ABN,
         norm_act="relu",
         antialias=False,
+        keep_prob=1,
     ):
         super(BasicBlock, self).__init__()
         antialias = antialias and stride == 2
@@ -167,6 +168,7 @@ def __init__(
         self.downsample = downsample
         self.blurpool = BlurPool(channels=planes) if antialias else nn.Identity()
         self.antialias = antialias
+        self.drop_connect = DropConnect(keep_prob) if keep_prob < 1 else nn.Identity()
 
     def forward(self, x):
         residual = x
@@ -180,11 +182,11 @@ def forward(self, x):
         if self.antialias:
             out = self.blurpool(out)
         out = self.conv2(out)
-        # avoid 2 inplace ops by chaining into one long op. Neede for inplaceabn
+        # avoid 2 inplace ops by chaining into one long op. Needed for inplaceabn
         if self.se_module is not None:
-            out = self.se_module(self.bn2(out)) + residual
+            out = self.drop_connect(self.se_module(self.bn2(out))) + residual
         else:
-            out = self.bn2(out) + residual
+            out = self.drop_connect(self.bn2(out)) + residual
         return self.final_act(out)
 
 
@@ -204,6 +206,7 @@ def __init__(
         norm_layer=ABN,
         norm_act="relu",
         antialias=False,
+        keep_prob=1,  # for drop connect
     ):
         super(Bottleneck, self).__init__()
         antialias = antialias and stride == 2
@@ -222,6 +225,7 @@ def __init__(
         self.downsample = downsample
         self.blurpool = BlurPool(channels=width) if antialias else nn.Identity()
         self.antialias = antialias
+        self.drop_connect = DropConnect(keep_prob) if keep_prob < 1 else nn.Identity()
 
     def forward(self, x):
         residual = x
@@ -241,9 +245,9 @@ def forward(self, x):
         out = self.conv3(out)
         # avoid 2 inplace ops by chaining into one long op
         if self.se_module is not None:
-            out = self.se_module(self.bn3(out)) + residual
+            out = self.drop_connect(self.se_module(self.bn3(out))) + residual
         else:
-            out = self.bn3(out) + residual
+            out = self.drop_connect(self.bn3(out)) + residual
         return self.final_act(out)
 
 # TResnet models use slightly modified versions of BasicBlock and Bottleneck
@@ -292,5 +296,5 @@ def forward(self, x):
 
         out = self.conv3(out)
         # avoid 2 inplace ops by chaining into one long op
-        out = self.bn3(out) + residual
+        out = self.drop_connect(self.bn3(out)) + residual
         return self.final_act(out)

pytorch_tools/segmentation_models/unet.py

@@ -38,7 +38,7 @@ def __init__(
         self.layer3 = UnetDecoderBlock(in_channels[2], out_channels[2], **bn_params)
         self.layer4 = UnetDecoderBlock(in_channels[3], out_channels[3], **bn_params)
         self.layer5 = UnetDecoderBlock(in_channels[4], out_channels[4], **bn_params)
-        self.dropout = nn.Dropout2d(drop_rate, inplace=True)
+        self.dropout = nn.Dropout2d(drop_rate, inplace=False) # inplace=True raises a backprop error
         self.final_conv = conv1x1(out_channels[4], final_channels)
 
         initialize(self)

tests/models/test_models.py

@@ -73,6 +73,11 @@ def test_dilation(arch, output_stride):
     W, H = INP.shape[-2:]
     assert res.shape[-2:] == (W // output_stride, H // output_stride)
 
+@pytest.mark.parametrize("arch", TEST_MODEL_NAMES)
+def test_drop_connect(arch):
+    m = models.__dict__[arch](drop_connect_rate=0.2)
+    _test_forward(m)
+
 NUM_PARAMS = {
     "tresnetm": 31389032,
     "tresnetl": 55989256,

tests/models/test_weights.py

@@ -1,8 +1,9 @@
 ## test that imagenet pretrained weights are valid and able to classify correctly the cat and dog
 
+import torch
+import pytest
 import numpy as np
 from PIL import Image
-import pytest
 
 from pytorch_tools.utils.preprocessing import get_preprocessing_fn
 from pytorch_tools.utils.visualization import tensor_from_rgb_image
@@ -53,3 +54,25 @@ def test_imagenet_pretrain(arch):
         im = im.view(1, *im.shape).float()
         pred_cls = m(im).argmax()
         assert pred_cls == im_cls
+
+# test that output mean for fixed input is the same
+MODEL_NAMES2 = [
+    "resnet34",
+    "se_resnet50",
+    "efficientnet_b0",
+]
+
+MODEL_MEAN = {
+    "resnet34": 7.6799e-06,
+    "se_resnet50": -2.6095e-06,
+    "efficientnet_b0": 0.0070,
+}
+
+@pytest.mark.parametrize("arch", MODEL_NAMES2)
+def test_output_mean(arch):
+    m = models.__dict__[arch](pretrained="imagenet")
+    m.eval()
+    inp = torch.ones(1, 3, 256, 256)
+    with torch.no_grad():
+        out = m(inp).mean().numpy()
+    assert np.allclose(out, MODEL_MEAN[arch], rtol=1e-4, atol=1e-4)