update: models & weights for cspdarknet

README.md

@@ -40,6 +40,8 @@ Tianheng Cheng, <a href="https://xinggangw.info/">Xinggang Wang</a><sup><span>&#
 
 `This project is under active development, please stay tuned!` &#9749;
 
+* `[2022-10-31]`: We release the models & weights for the [`CSP-DarkNet53`](configs/sparse_inst_cspdarknet53_giam.yaml) backbone. Which is a strong baseline with highly-competitve inference speed and accuracy.
+
 * `[2022-10-19]`: We provide the implementation and inference code based on [MindSpore](https://www.mindspore.cn/), a nice and efficient Deep Learning framework. Thanks [Ruiqi Wang](https://github.com/RuiqiWang00) for this kind contribution!
 
 * `[2022-8-9]`: We provide the FLOPs counter [`get_flops.py`](./tools/get_flops.py) to obtain the FLOPs/Parameters of SparseInst. This update also includes some bugfixs.
@@ -85,12 +87,11 @@ All models are trained on MS-COCO *train2017*.
 | [SparseInst (G-IAM)](configs/sparse_inst_r50vd_dcn_giam_aug.yaml) | [R-50-vd-DCN](https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-weights/resnet50d_ra2-464e36ba.pth) | 608 | &#10003; | 37.4 | 37.9 | 40.0  | [model](https://drive.google.com/file/d/1clYPdCNrDNZLbmlAEJ7wjsrOLn1igOpT/view?usp=sharing)|
 | [SparseInst (G-IAM)](configs/sparse_inst_r50vd_dcn_giam_aug.yaml) | [R-50-vd-DCN](https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-weights/resnet50d_ra2-464e36ba.pth) | 640 | &#10003; | 37.7 | 38.1 | 39.3 |  [model](https://drive.google.com/file/d/1clYPdCNrDNZLbmlAEJ7wjsrOLn1igOpT/view?usp=sharing)| 
 
-<!-- #### SparseInst with other backbones
+#### SparseInst with other backbones
 
 | model | backbone | input | AP<sup>val</sup> |  AP  | FPS | weights |
 | :---- | :------ | :---: | :--------------: | :--: | :-: | :-----: |
-| SparseInst (G-IAM) | [VoVNet]() | 640 | - | - | - | [model]() |
-| SparseInst (G-IAM) | [CSPDarkNet]() | 640 | - | -| - | [model]() |  -->
+| SparseInst (G-IAM) | [CSPDarkNet](configs/sparse_inst_cspdarknet53_giam.yaml) | 640 | 35.1 | -| - | [model](https://drive.google.com/file/d/1rcUJWUbusM216Zbtmo_xB774jdjb3qSt/view?usp=sharing) |
 
 #### Larger models
 
@@ -238,11 +239,8 @@ python tools/train_net.py --config-file configs/sparse_inst_r50vd_dcn_giam_aug.y
 ### Custom Training of SparseInst
 
 1. We suggest you convert your custom datasets into the `COCO` format, which enables the usage of the default dataset mappers and loaders. You may find more details in the [official guide of detectron2](https://detectron2.readthedocs.io/en/latest/tutorials/datasets.html#register-a-coco-format-dataset).
-
 2. You need to check whether `NUM_CLASSES` and `NUM_MASKS` should be changed according to your scenarios or tasks.
-
 3. Change the configurations accordingly.
-
 4. After finishing the above procedures, you can easily train SparseInst by `train_net.py`.
 
 

configs/sparse_inst_cspdarknet53_giam.yaml

@@ -6,7 +6,9 @@ MODEL:
   SPARSE_INST:
     ENCODER:
       IN_FEATURES: ["csp2", "csp3", "csp4"]
+    DECODER:
+      NAME: "GroupIAMSoftDecoder"
   CSPNET:
-    NAME: "darknet53"
+    NAME: "cspdarknet53"
     OUT_FEATURES: ["csp2", "csp3", "csp4"]
 OUTPUT_DIR: "output/sparse_inst_cspdarknet53_giam"

sparseinst/backbones/cspnet.py

@@ -3,7 +3,8 @@
 import torch
 import torch.nn as nn
 
-from timm.models.layers import ConvBnAct, DropPath, AvgPool2dSame, create_attn
+from timm.models.layers import create_conv2d, create_act_layer
+from timm.models.layers import DropPath, AvgPool2dSame, create_attn
 
 
 from detectron2.layers import ShapeSpec, FrozenBatchNorm2d
@@ -84,6 +85,39 @@
     )
 )
 
+class ConvBnAct(nn.Module):
+    def __init__(self, in_channels, out_channels, kernel_size=1, stride=1, padding='', dilation=1, groups=1,
+                 bias=False, apply_act=True, norm_layer=nn.BatchNorm2d, act_layer=nn.ReLU, aa_layer=None,
+                 drop_block=None):
+        super(ConvBnAct, self).__init__()
+        use_aa = aa_layer is not None
+
+        self.conv = create_conv2d(
+            in_channels, out_channels, kernel_size, stride=1 if use_aa else stride,
+            padding=padding, dilation=dilation, groups=groups, bias=bias)
+
+        # NOTE for backwards compatibility with models that use separate norm and act layer definitions
+        self.bn = norm_layer(out_channels)
+        self.act = act_layer()
+        self.aa = aa_layer(
+            channels=out_channels) if stride == 2 and use_aa else None
+
+    @property
+    def in_channels(self):
+        return self.conv.in_channels
+
+    @property
+    def out_channels(self):
+        return self.conv.out_channels
+
+    def forward(self, x):
+        x = self.conv(x)
+        x = self.bn(x)
+        x = self.act(x)
+        if self.aa is not None:
+            x = self.aa(x)
+        return x
+
 
 def create_stem(
         in_chans=3, out_chs=32, kernel_size=3, stride=2, pool='',
@@ -394,8 +428,7 @@ def build_cspnet_backbone(cfg, input_shape=None):
     if norm_name == "FrozenBN":
         norm = FrozenBatchNorm2d
     elif norm_name == "SyncBN":
-        from detectron2.layers import NaiveSyncBatchNorm
-        norm = NaiveSyncBatchNorm
+        norm = nn.SyncBatchNorm
     else:
         norm = nn.BatchNorm2d