From 3bc95a433223e261dcc9cba4c329eca644cbaded Mon Sep 17 00:00:00 2001
From: johnzja <johnzja@qq.com>
Date: Wed, 12 Aug 2020 15:58:48 +0800
Subject: [PATCH] Inverted Residual kept coherent with mmcl.

---
 .../fast_scnn_4x8_80k_lr0.12_cityscapes.py    |  1 -
 mmseg/ops/inverted_residual_module.py         | 85 +++++++++++--------
 2 files changed, 51 insertions(+), 35 deletions(-)

diff --git a/configs/fastscnn/fast_scnn_4x8_80k_lr0.12_cityscapes.py b/configs/fastscnn/fast_scnn_4x8_80k_lr0.12_cityscapes.py
index 0da288735..53fcfc420 100644
--- a/configs/fastscnn/fast_scnn_4x8_80k_lr0.12_cityscapes.py
+++ b/configs/fastscnn/fast_scnn_4x8_80k_lr0.12_cityscapes.py
@@ -2,7 +2,6 @@ _base_ = [
     '../_base_/models/fast_scnn.py', '../_base_/datasets/cityscapes.py',
     '../_base_/default_runtime.py', '../_base_/schedules/schedule_80k.py'
 ]
-cudnn_benchmark = True
 
 # Re-config the data sampler.
 data = dict(samples_per_gpu=8, workers_per_gpu=4)
diff --git a/mmseg/ops/inverted_residual_module.py b/mmseg/ops/inverted_residual_module.py
index 3be849982..53f6744e8 100644
--- a/mmseg/ops/inverted_residual_module.py
+++ b/mmseg/ops/inverted_residual_module.py
@@ -1,71 +1,88 @@
-from mmcv.cnn import ConvModule, build_norm_layer
+import torch.utils.checkpoint as cp
+from mmcv.cnn import ConvModule
 from torch import nn
 
 
 class InvertedResidual(nn.Module):
-    """Inverted residual module.
+    """InvertedResidual block for MobileNetV2.
 
     Args:
-        inp (int): input channels.
-        oup (int): output channels.
-        stride (int): downsampling factor.
-        expand_ratio (int): 1 or 2.
-        dilation (int): Dilated conv. Default: 1.
-        conv_cfg (dict | None): Config of conv layers. Default: None.
-        norm_cfg (dict | None): Config of norm layers. Default:
-            dict(type='BN').
-        act_cfg (dict): Config of activation layers. Default:
-            dict(type='ReLU6').
+        in_channels (int): The input channels of the InvertedResidual block.
+        out_channels (int): The output channels of the InvertedResidual block.
+        stride (int): Stride of the middle (first) 3x3 convolution.
+        expand_ratio (int): adjusts number of channels of the hidden layer
+            in InvertedResidual by this amount.
+        conv_cfg (dict): Config dict for convolution layer.
+            Default: None, which means using conv2d.
+        norm_cfg (dict): Config dict for normalization layer.
+            Default: dict(type='BN').
+        act_cfg (dict): Config dict for activation layer.
+            Default: dict(type='ReLU6').
+        with_cp (bool): Use checkpoint or not. Using checkpoint will save some
+            memory while slowing down the training speed. Default: False.
+    Returns:
+        Tensor: The output tensor
     """
 
     def __init__(self,
-                 inp,
-                 oup,
+                 in_channels,
+                 out_channels,
                  stride,
                  expand_ratio,
-                 dilation=1,
                  conv_cfg=None,
                  norm_cfg=dict(type='BN'),
-                 act_cfg=dict(type='ReLU6')):
+                 act_cfg=dict(type='ReLU6'),
+                 with_cp=False):
         super(InvertedResidual, self).__init__()
         self.stride = stride
-        assert stride in [1, 2]
-
-        hidden_dim = int(round(inp * expand_ratio))
-        self.use_res_connect = self.stride == 1 and inp == oup
+        assert stride in [1, 2], f'stride must in [1, 2]. ' \
+            f'But received {stride}.'
+        self.with_cp = with_cp
+        self.use_res_connect = self.stride == 1 and in_channels == out_channels
+        hidden_dim = int(round(in_channels * expand_ratio))
 
         layers = []
         if expand_ratio != 1:
-            # pw
             layers.append(
                 ConvModule(
-                    inp,
-                    hidden_dim,
+                    in_channels=in_channels,
+                    out_channels=hidden_dim,
                     kernel_size=1,
                     conv_cfg=conv_cfg,
                     norm_cfg=norm_cfg,
                     act_cfg=act_cfg))
         layers.extend([
-            # dw
             ConvModule(
-                hidden_dim,
-                hidden_dim,
+                in_channels=hidden_dim,
+                out_channels=hidden_dim,
                 kernel_size=3,
-                padding=dilation,
                 stride=stride,
-                dilation=dilation,
+                padding=1,
                 groups=hidden_dim,
                 conv_cfg=conv_cfg,
                 norm_cfg=norm_cfg,
                 act_cfg=act_cfg),
-            # pw-linear
-            nn.Conv2d(hidden_dim, oup, 1, 1, 0, bias=False),
-            build_norm_layer(norm_cfg, oup)[1],
+            ConvModule(
+                in_channels=hidden_dim,
+                out_channels=out_channels,
+                kernel_size=1,
+                conv_cfg=conv_cfg,
+                norm_cfg=norm_cfg,
+                act_cfg=None)
         ])
         self.conv = nn.Sequential(*layers)
 
     def forward(self, x):
-        if self.use_res_connect:
-            return x + self.conv(x)
+
+        def _inner_forward(x):
+            if self.use_res_connect:
+                return x + self.conv(x)
+            else:
+                return self.conv(x)
+
+        if self.with_cp and x.requires_grad:
+            out = cp.checkpoint(_inner_forward, x)
         else:
-            return self.conv(x)
+            out = _inner_forward(x)
+
+        return out