update rcil

Author: CharlesPikachu

csseg/configs/rcil/base_cfg.py

@@ -10,8 +10,8 @@
         'depth': 101,
         'outstride': 16,
         'out_indices': (0, 1, 2, 3),
-        'norm_cfg': {'type': 'InPlaceABNSync', 'activation': 'identity'},
-        'act_cfg': {'type': 'LeakyReLU', 'negative_slope': 0.01, 'inplace': False},
+        'norm_cfg': {'type': 'InPlaceABNSync', 'activation': 'leaky_relu', 'activation_param': 1.0},
+        'act_cfg': None,
         'pretrained': True,
     }, 
     'decoder_cfg': {
@@ -21,8 +21,8 @@
         'out_channels': 256,
         'dilations': (1, 6, 12, 18),
         'pooling_size': 32,
-        'norm_cfg': {'type': 'InPlaceABNSync', 'activation': 'identity'},
-        'act_cfg': {'type': 'LeakyReLU', 'negative_slope': 0.01, 'inplace': False},
+        'norm_cfg': {'type': 'InPlaceABNSync', 'activation': 'leaky_relu', 'activation_param': 1.0},
+        'act_cfg': None,
     },
     'losses_cfgs': {
         'segmentation_init': {

csseg/modules/models/decoders/rcilaspphead.py

@@ -47,14 +47,14 @@ def __init__(self, in_channels, feats_channels, out_channels, dilations, pooling
         self.global_branch = nn.Sequential(
             nn.Conv2d(in_channels, feats_channels, kernel_size=1, stride=1, padding=0, bias=False),
             BuildNormalization(placeholder=feats_channels, norm_cfg=norm_cfg),
-            BuildActivation(act_cfg=act_cfg),
+            nn.LeakyReLU(0.01),
             nn.Conv2d(feats_channels, feats_channels, kernel_size=1, stride=1, padding=0, bias=False),
         )
         # output project
         self.bottleneck_conv = nn.Conv2d(feats_channels * len(dilations), out_channels, kernel_size=1, stride=1, padding=0, bias=False)
         self.bottleneck_bn = nn.Sequential(
             BuildNormalization(placeholder=out_channels, norm_cfg=norm_cfg),
-            BuildActivation(act_cfg=act_cfg),
+            nn.LeakyReLU(0.01),
         )
         # initialize parameters
         assert norm_cfg['activation'] == 'identity'
@@ -77,9 +77,9 @@ def forward(self, x):
         input_size = x.shape
         # feed to parallel convolutions branch1 and branch2
         outputs_branch1 = torch.cat([conv(x) for conv in self.parallel_convs_branch1], dim=1)
-        outputs_branch1 = self.parallel_bn_branch1[0](outputs_branch1)
+        outputs_branch1 = self.parallel_bn_branch1(outputs_branch1)
         outputs_branch2 = torch.cat([conv(x) for conv in self.parallel_convs_branch2], dim=1)
-        outputs_branch2 = self.parallel_bn_branch2[0](outputs_branch2)
+        outputs_branch2 = self.parallel_bn_branch2(outputs_branch2)
         # merge
         r = torch.rand(1, outputs_branch1.shape[1], 1, 1, dtype=torch.float32)
         if not self.training: r[:, :, :, :] = 1.0
@@ -91,7 +91,7 @@ def forward(self, x):
         weight_branch2[(r < 0.66) & (r >= 0.33)] = 2.
         weight_branch2[r >= 0.66] = 1.
         outputs = outputs_branch1 * weight_branch1.type_as(outputs_branch1) * 0.5 + outputs_branch2 * weight_branch2.type_as(outputs_branch2) * 0.5
-        outputs = self.parallel_bn_branch1[1](outputs)
+        outputs = F.leaky_relu(outputs, negative_slope=0.01)
         outputs = self.bottleneck_conv(outputs)
         # feed to global branch
         global_feats = self.globalpooling(x)

csseg/modules/models/encoders/resnetplop.py

@@ -18,6 +18,7 @@ def __init__(self, inplanes, planes, stride=1, dilation=1, downsample=None, norm
         )
     '''forward'''
     def forward(self, x):
+        if isinstance(x, tuple): x = x[0]
         identity = x
         out = self.conv1(x)
         out = self.bn1(out)

csseg/modules/models/encoders/resnetrcil.py

@@ -7,6 +7,7 @@
 import re
 import torch
 import torch.nn as nn
+import torch.nn.functional as F
 from .bricks import BuildNormalization
 from .resnet import ResNet, BasicBlock, Bottleneck
 
@@ -23,10 +24,11 @@ def __init__(self, inplanes, planes, stride=1, dilation=1, downsample=None, norm
         self.bn2_branch2 = BuildNormalization(placeholder=planes, norm_cfg=shortcut_norm_cfg)
     '''forward'''
     def forward(self, x):
+        if isinstance(x, tuple): x = x[0]
         identity = x
         out = self.conv1(x)
         out = self.bn1(out)
-        out = self.relu(out)
+        out = F.leaky_relu(out, 0.01)
         out_branch1 = self.conv2(out)
         out_branch1 = self.bn2(out_branch1)
         out_branch2 = self.conv2_branch2(out)
@@ -41,7 +43,7 @@ def forward(self, x):
         weight_branch2[(r < 0.66) & (r >= 0.33)] = 2.
         weight_branch2[r >= 0.66] = 1.
         out = out_branch1 * weight_branch1.type_as(out_branch1) * 0.5 + out_branch2 * weight_branch2.type_as(out_branch2) * 0.5
-        out = self.relu(out)
+        out = F.leaky_relu(out, 0.01)
         if self.downsample is not None: identity = self.downsample(x)
         out = out + identity
         distillation = out
@@ -65,7 +67,7 @@ def forward(self, x):
         identity = x
         out = self.conv1(x)
         out = self.bn1(out)
-        out = self.relu(out)
+        out = F.leaky_relu(out, 0.01)
         out_branch1 = self.conv2(out)
         out_branch1 = self.bn2(out_branch1)
         out_branch2 = self.conv2_branch2(out)
@@ -80,7 +82,7 @@ def forward(self, x):
         weight_branch2[(r < 0.66) & (r >= 0.33)] = 2.
         weight_branch2[r >= 0.66] = 1.
         out = out_branch1 * weight_branch1.type_as(out_branch1) * 0.5 + out_branch2 * weight_branch2.type_as(out_branch2) * 0.5
-        out = self.relu(out)
+        out = F.leaky_relu(out, 0.01)
         out = self.conv3(out)
         out = self.bn3(out)
         if self.downsample is not None: identity = self.downsample(x)