Refractor: Call the base conv directly

Author: MILK-BIOS

sscma/models/backbones/MobileNetv2.py

@@ -72,7 +72,7 @@ def __init__(
 
         self.layers = []
         for idx, (t, c, n, s) in enumerate(inverted_residual_setting):
-            out_channels = _make_divisible(c * widen_factor, round_nearest) * (2 if rep else 1)
+            out_channels = _make_divisible(c * widen_factor, round_nearest)  # * (2 if rep else 1) why times 2?
             tmp_layers = []
             for i in range(n):
                 stride = s if i == 0 else 1

sscma/models/backbones/csp_darknet.py

@@ -5,7 +5,7 @@
 import torch
 from mmdet.utils import ConfigType, OptMultiConfig
 
-from sscma.models.base import ConvModule
+from sscma.models.base import ConvNormActivation
 from sscma.models.layers import CSPLayer, SPPFBottleneck
 from sscma.models.utils import make_divisible, make_round
 from sscma.registry import MODELS
@@ -61,14 +61,14 @@ def __init__(
         )
 
     def build_stem_layer(self):
-        return ConvModule(
+        return ConvNormActivation(
             self.input_channels,
             make_divisible(self.arch_settings[self.arch][0][0], self.widen_factor),
             kernel_size=6,
             stride=2,
             padding=2,
-            norm_cfg=self.norm_cfg,
-            act_cfg=self.act_cfg,
+            norm_layer=self.norm_cfg,
+            activation_layer=self.act_cfg,
         )
 
     def build_stage_layer(self, stage_idx: int, setting: list) -> list:
@@ -78,8 +78,14 @@ def build_stage_layer(self, stage_idx: int, setting: list) -> list:
         out_channels = make_divisible(out_channels, self.widen_factor)
         num_blocks = make_round(num_blocks)
         stage = []
-        conv_layer = ConvModule(
-            in_channels, out_channels, kernel_size=3, stride=2, padding=1, norm_cfg=self.norm_cfg, act_cfg=self.act_cfg
+        conv_layer = ConvNormActivation(
+            in_channels,
+            out_channels,
+            kernel_size=3,
+            stride=2,
+            padding=1,
+            norm_layer=self.norm_cfg,
+            activation_layer=self.act_cfg,
         )
         stage.append(conv_layer)
 

sscma/models/heads/yolov5_head.py

@@ -520,7 +520,7 @@ def loss_by_feat(
 
             # no gt bbox matches anchor
             if batch_targets_scaled.shape[0] == 0:
-                loss_box += bbox_preds[i].sum() * 0
+                loss_box += bbox_preds[i].sum() * 0  # why not + 0 directly
                 loss_cls += cls_scores[i].sum() * 0
                 loss_obj += self.loss_obj(objectnesses[i], target_obj) * self.obj_level_weights[i]
                 continue

sscma/models/layers/sppf.py

@@ -8,7 +8,7 @@
 from mmengine.model import BaseModule
 from torch import Tensor
 
-from ..base import ConvModule
+from ..base import ConvNormActivation
 
 
 class SPPFBottleneck(BaseModule):
@@ -28,8 +28,14 @@ def __init__(
 
         if use_conv_first:
             mid_channels = int(in_channels * mid_channels_scale)
-            self.conv1 = ConvModule(
-                in_channels, mid_channels, 1, stride=1, conv_cfg=conv_cfg, norm_cfg=norm_cfg, act_cfg=act_cfg
+            self.conv1 = ConvNormActivation(
+                in_channels,
+                mid_channels,
+                1,
+                stride=1,
+                conv_layer=conv_cfg,
+                norm_layer=norm_cfg,
+                activation_layer=act_cfg,
             )
         else:
             mid_channels = in_channels
@@ -44,8 +50,8 @@ def __init__(
             )
             conv2_in_channels = mid_channels * (len(kernel_sizes) + 1)
 
-        self.conv2 = ConvModule(
-            conv2_in_channels, out_channels, 1, conv_cfg=conv_cfg, norm_cfg=norm_cfg, act_cfg=act_cfg
+        self.conv2 = ConvNormActivation(
+            conv2_in_channels, out_channels, 1, conv_layer=conv_cfg, norm_layer=norm_cfg, activation_layer=act_cfg
         )
 
     def forward(self, x: Tensor) -> Tensor:

sscma/models/necks/fpn.py

@@ -15,7 +15,7 @@
 
 from sscma.registry import NECKS
 
-from ..base import ConvModule
+from ..base import ConvNormActivation
 from ..layers import CSPLayer
 from ..utils import make_divisible, make_round
 
@@ -416,12 +416,12 @@ def build_reduce_layer(self, idx: int) -> nn.Module:
             nn.Module: The reduce layer.
         """
         if idx == len(self.in_channels) - 1:
-            layer = ConvModule(
+            layer = ConvNormActivation(
                 make_divisible(self.in_channels[idx], self.widen_factor),
                 make_divisible(self.in_channels[idx - 1], self.widen_factor),
                 1,
-                norm_cfg=self.norm_cfg,
-                act_cfg=self.act_cfg,
+                norm_layer=self.norm_cfg,
+                activation_layer=self.act_cfg,
             )
         else:
             layer = nn.Identity()
@@ -461,12 +461,12 @@ def build_top_down_layer(self, idx: int):
                     norm_cfg=self.norm_cfg,
                     act_cfg=self.act_cfg,
                 ),
-                ConvModule(
+                ConvNormActivation(
                     make_divisible(self.in_channels[idx - 1], self.widen_factor),
                     make_divisible(self.in_channels[idx - 2], self.widen_factor),
                     kernel_size=1,
-                    norm_cfg=self.norm_cfg,
-                    act_cfg=self.act_cfg,
+                    norm_layer=self.norm_cfg,
+                    activation_layer=self.act_cfg,
                 ),
             )
 
@@ -479,14 +479,14 @@ def build_downsample_layer(self, idx: int) -> nn.Module:
         Returns:
             nn.Module: The downsample layer.
         """
-        return ConvModule(
+        return ConvNormActivation(
             make_divisible(self.in_channels[idx], self.widen_factor),
             make_divisible(self.in_channels[idx], self.widen_factor),
             kernel_size=3,
             stride=2,
             padding=1,
-            norm_cfg=self.norm_cfg,
-            act_cfg=self.act_cfg,
+            norm_layer=self.norm_cfg,
+            activation_layer=self.act_cfg,
         )
 
     def build_bottom_up_layer(self, idx: int) -> nn.Module:

tools/train.py

@@ -193,12 +193,12 @@ def main():
     args, cfg = build_config(args)
 
     if 'runner_type' not in cfg:
-        from mmengine.runner import Runner
+        from sscma.engine import Runner
 
         runner = Runner.from_cfg(cfg)
         runner.val_evaluator.dataset_meta = runner.val_dataloader.dataset.METAINFO
     else:
-        from mmengine.registry import RUNNERS
+        from sscma.registry import RUNNERS
 
         runner = RUNNERS.build(cfg)
         runner.val_evaluator.dataset_meta = runner.val_dataloader.dataset.METAINFO