Update: support DCNv4 in InternImage! (#277)

* Update: support DCNv4 in InternImage!

* fix mask softmax bug.

---------
Co-authored-by: Zhiqi Li <zhiqil@nvidia>

README.md

@@ -54,6 +54,7 @@ The official implementation of
 
 
 ## News
+- `Jan 22, 2024`: 🚀 Support [DCNv4](https://github.com/OpenGVLab/DCNv4) in InternImage!
 - `Mar 14, 2023`: 🚀 "INTERN-2.5" is released！
 - `Feb 28, 2023`: 🚀 InternImage is accepted to CVPR 2023!
 - `Nov 18, 2022`: 🚀 InternImage-XL merged into [BEVFormer v2](https://arxiv.org/abs/2211.10439) achieves state-of-the-art performance of `63.4 NDS` on nuScenes Camera Only.

detection/configs/coco/mask_rcnn_internimage_t_fpn_1x_coco_with_dcnv4.py

+# --------------------------------------------------------
+# InternImage
+# Copyright (c) 2022 OpenGVLab
+# Licensed under The MIT License [see LICENSE for details]
+# --------------------------------------------------------
+_base_ = [
+    '../_base_/models/mask_rcnn_r50_fpn.py',
+    '../_base_/datasets/coco_instance.py',
+    '../_base_/schedules/schedule_1x.py',
+    '../_base_/default_runtime.py'
+]
+pretrained = 'https://huggingface.co/OpenGVLab/InternImage/resolve/main/internimage_t_1k_224.pth'
+model = dict(
+    backbone=dict(
+        _delete_=True,
+        type='InternImage',
+        core_op='DCNv3',
+        channels=64,
+        depths=[4, 4, 18, 4],
+        groups=[4, 8, 16, 32],
+        mlp_ratio=4.,
+        drop_path_rate=0.2,
+        norm_layer='LN',
+        layer_scale=1.0,
+        offset_scale=1.0,
+        post_norm=False,
+        with_cp=False,
+        out_indices=(0, 1, 2, 3),
+        use_dcn_v4_op=True,
+        init_cfg=dict(type='Pretrained', checkpoint=pretrained)),
+    neck=dict(
+        type='FPN',
+        in_channels=[64, 128, 256, 512],
+        out_channels=256,
+        num_outs=5))
+# By default, models are trained on 8 GPUs with 2 images per GPU
+data = dict(samples_per_gpu=2)
+optimizer = dict(
+    _delete_=True, type='AdamW', lr=0.0001, weight_decay=0.05,
+    constructor='CustomLayerDecayOptimizerConstructor',
+    paramwise_cfg=dict(num_layers=30, layer_decay_rate=1.0,
+                       depths=[4, 4, 18, 4]))
+optimizer_config = dict(grad_clip=None)
+# fp16 = dict(loss_scale=dict(init_scale=512))
+evaluation = dict(save_best='auto')
+checkpoint_config = dict(
+    interval=1,
+    max_keep_ckpts=3,
+    save_last=True,
+)
\ No newline at end of file

detection/mmdet_custom/models/backbones/intern_image.py

@@ -15,7 +15,7 @@ from mmdet.utils import get_root_logger
 from mmdet.models.builder import BACKBONES
 import torch.nn.functional as F
 
-from ops_dcnv3 import modules as opsm
+from ops_dcnv3 import modules as dcnv3
 
 
 class to_channels_first(nn.Module):
@@ -365,7 +365,8 @@ class InternImageLayer(nn.Module):
                  with_cp=False,
                  dw_kernel_size=None, # for InternImage-H/G
                  res_post_norm=False, # for InternImage-H/G
-                 center_feature_scale=False): # for InternImage-H/G
+                 center_feature_scale=False,
+                 use_dcn_v4_op=False): # for InternImage-H/G
         super().__init__()
         self.channels = channels
         self.groups = groups
@@ -385,7 +386,8 @@ class InternImageLayer(nn.Module):
             act_layer=act_layer,
             norm_layer=norm_layer,
             dw_kernel_size=dw_kernel_size, # for InternImage-H/G
-            center_feature_scale=center_feature_scale) # for InternImage-H/G
+            center_feature_scale=center_feature_scale,
+            use_dcn_v4_op=use_dcn_v4_op) # for InternImage-H/G
         self.drop_path = DropPath(drop_path) if drop_path > 0. \
             else nn.Identity()
         self.norm2 = build_norm_layer(channels, 'LN')
@@ -469,7 +471,8 @@ class InternImageBlock(nn.Module):
                  dw_kernel_size=None, # for InternImage-H/G
                  post_norm_block_ids=None, # for InternImage-H/G
                  res_post_norm=False, # for InternImage-H/G
-                 center_feature_scale=False): # for InternImage-H/G
+                 center_feature_scale=False, # for InternImage-H/G
+                 use_dcn_v4_op=False):
         super().__init__()
         self.channels = channels
         self.depth = depth
@@ -493,7 +496,8 @@ class InternImageBlock(nn.Module):
                 with_cp=with_cp,
                 dw_kernel_size=dw_kernel_size, # for InternImage-H/G
                 res_post_norm=res_post_norm, # for InternImage-H/G
-                center_feature_scale=center_feature_scale # for InternImage-H/G
+                center_feature_scale=center_feature_scale, # for InternImage-H/G
+                use_dcn_v4_op=use_dcn_v4_op
             ) for i in range(depth)
         ])
         if not self.post_norm or center_feature_scale:
@@ -569,6 +573,7 @@ class InternImage(nn.Module):
                  level2_post_norm_block_ids=None,  # for InternImage-H/G
                  res_post_norm=False,  # for InternImage-H/G
                  center_feature_scale=False,  # for InternImage-H/G
+                 use_dcn_v4_op=False,
                  out_indices=(0, 1, 2, 3),
                  init_cfg=None,
                  **kwargs):
@@ -591,6 +596,7 @@ class InternImage(nn.Module):
         logger.info(f"level2_post_norm: {level2_post_norm}")
         logger.info(f"level2_post_norm_block_ids: {level2_post_norm_block_ids}")
         logger.info(f"res_post_norm: {res_post_norm}")
+        logger.info(f"use_dcn_v4_op: {use_dcn_v4_op}")
 
         in_chans = 3
         self.patch_embed = StemLayer(in_chans=in_chans,
@@ -611,7 +617,7 @@ class InternImage(nn.Module):
             post_norm_block_ids = level2_post_norm_block_ids if level2_post_norm and (
                 i == 2) else None # for InternImage-H/G
             level = InternImageBlock(
-                core_op=getattr(opsm, core_op),
+                core_op=getattr(dcnv3, core_op),
                 channels=int(channels * 2**i),
                 depth=depths[i],
                 groups=groups[i],
@@ -628,7 +634,8 @@ class InternImage(nn.Module):
                 dw_kernel_size=dw_kernel_size,  # for InternImage-H/G
                 post_norm_block_ids=post_norm_block_ids, # for InternImage-H/G
                 res_post_norm=res_post_norm, # for InternImage-H/G
-                center_feature_scale=center_feature_scale # for InternImage-H/G
+                center_feature_scale=center_feature_scale, # for InternImage-H/G
+                use_dcn_v4_op=use_dcn_v4_op,
             )
             self.levels.append(level)
 
@@ -687,7 +694,7 @@ class InternImage(nn.Module):
             nn.init.constant_(m.weight, 1.0)
 
     def _init_deform_weights(self, m):
-        if isinstance(m, getattr(opsm, self.core_op)):
+        if isinstance(m, getattr(dcnv3, self.core_op)):
             m._reset_parameters()
 
     def forward(self, x):

detection/ops_dcnv3/modules/dcnv3.py

@@ -14,7 +14,11 @@ from torch import nn
 import torch.nn.functional as F
 from torch.nn.init import xavier_uniform_, constant_
 from ..functions import DCNv3Function, dcnv3_core_pytorch
-
+try:
+    from DCNv4.functions import DCNv4Function
+except:
+    warnings.warn('Now, we support DCNv4 in InternImage.')
+import math
 
 class to_channels_first(nn.Module):
 
@@ -228,7 +232,9 @@ class DCNv3(nn.Module):
             offset_scale=1.0,
             act_layer='GELU',
             norm_layer='LN',
-            center_feature_scale=False):
+            center_feature_scale=False,
+            use_dcn_v4_op=False,
+            ):
         """
         DCNv3 Module
         :param channels
@@ -264,7 +270,9 @@ class DCNv3(nn.Module):
         self.group_channels = channels // group
         self.offset_scale = offset_scale
         self.center_feature_scale = center_feature_scale
-        
+         
+        self.use_dcn_v4_op = use_dcn_v4_op
+
         self.dw_conv = nn.Sequential(
             nn.Conv2d(
                 channels,
@@ -321,18 +329,44 @@ class DCNv3(nn.Module):
         x1 = self.dw_conv(x1)
         offset = self.offset(x1)
         mask = self.mask(x1).reshape(N, H, W, self.group, -1)
-        mask = F.softmax(mask, -1).reshape(N, H, W, -1).type(dtype)
-
-        x = DCNv3Function.apply(
-            x, offset, mask,
-            self.kernel_size, self.kernel_size,
-            self.stride, self.stride,
-            self.pad, self.pad,
-            self.dilation, self.dilation,
-            self.group, self.group_channels,
-            self.offset_scale,
-            256)
         
+        if not self.use_dcn_v4_op:
+            mask = F.softmax(mask, -1).reshape(N, H, W, -1).type(dtype)
+            x = DCNv3Function.apply(
+                x, offset, mask,
+                self.kernel_size, self.kernel_size,
+                self.stride, self.stride,
+                self.pad, self.pad,
+                self.dilation, self.dilation,
+                self.group, self.group_channels,
+                self.offset_scale,
+                256)
+        else:
+            # DCNv4 combines offset and weight mask into one tensor `offset_mask`.
+            # The following code is to align DCNv3 and DCNv4
+            offset = offset.view(N, H, W, self.group, -1)
+            mask = F.softmax(mask, -1)
+            mask = mask.view(N, H, W, self.group, -1)
+            offset_mask = torch.cat([offset, mask], -1).view(N, H, W, -1).contiguous()
+
+            # For efficiency, the last dimension of the offset_mask tensor in dcnv4 is a multiple of 8. 
+            K3 = offset_mask.size(-1)
+            K3_pad = int(math.ceil(K3/8)*8)
+            pad_dim = K3_pad - K3
+            offset_mask = torch.cat([offset_mask, offset_mask.new_zeros([*offset_mask.size()[:3], pad_dim])], -1)
+        
+            x = DCNv4Function.apply(
+                x, offset_mask,
+                self.kernel_size, self.kernel_size,
+                self.stride, self.stride,
+                self.pad, self.pad,
+                self.dilation, self.dilation,
+                self.group, self.group_channels,
+                self.offset_scale,
+                256,
+                False
+            )
+
         if self.center_feature_scale:
             center_feature_scale = self.center_feature_scale_module(
                 x1, self.center_feature_scale_proj_weight, self.center_feature_scale_proj_bias)

segmentation/mmseg_custom/models/backbones/intern_image.py

@@ -15,7 +15,7 @@ from mmseg.utils import get_root_logger
 from mmseg.models.builder import BACKBONES
 import torch.nn.functional as F
 
-from ops_dcnv3 import modules as opsm
+from ops_dcnv3 import modules as dcnv3
 
 
 class to_channels_first(nn.Module):
@@ -365,7 +365,8 @@ class InternImageLayer(nn.Module):
                  with_cp=False,
                  dw_kernel_size=None, # for InternImage-H/G
                  res_post_norm=False, # for InternImage-H/G
-                 center_feature_scale=False): # for InternImage-H/G
+                 center_feature_scale=False,
+                 use_dcn_v4_op=False): # for InternImage-H/G
         super().__init__()
         self.channels = channels
         self.groups = groups
@@ -385,7 +386,8 @@ class InternImageLayer(nn.Module):
             act_layer=act_layer,
             norm_layer=norm_layer,
             dw_kernel_size=dw_kernel_size, # for InternImage-H/G
-            center_feature_scale=center_feature_scale) # for InternImage-H/G
+            center_feature_scale=center_feature_scale,
+            use_dcn_v4_op=use_dcn_v4_op) # for InternImage-H/G
         self.drop_path = DropPath(drop_path) if drop_path > 0. \
             else nn.Identity()
         self.norm2 = build_norm_layer(channels, 'LN')
@@ -469,7 +471,8 @@ class InternImageBlock(nn.Module):
                  dw_kernel_size=None, # for InternImage-H/G
                  post_norm_block_ids=None, # for InternImage-H/G
                  res_post_norm=False, # for InternImage-H/G
-                 center_feature_scale=False): # for InternImage-H/G
+                 center_feature_scale=False, # for InternImage-H/G
+                 use_dcn_v4_op=False):
         super().__init__()
         self.channels = channels
         self.depth = depth
@@ -493,7 +496,8 @@ class InternImageBlock(nn.Module):
                 with_cp=with_cp,
                 dw_kernel_size=dw_kernel_size, # for InternImage-H/G
                 res_post_norm=res_post_norm, # for InternImage-H/G
-                center_feature_scale=center_feature_scale # for InternImage-H/G
+                center_feature_scale=center_feature_scale, # for InternImage-H/G
+                use_dcn_v4_op=use_dcn_v4_op
             ) for i in range(depth)
         ])
         if not self.post_norm or center_feature_scale:
@@ -569,6 +573,7 @@ class InternImage(nn.Module):
                  level2_post_norm_block_ids=None,  # for InternImage-H/G
                  res_post_norm=False,  # for InternImage-H/G
                  center_feature_scale=False,  # for InternImage-H/G
+                 use_dcn_v4_op=False,
                  out_indices=(0, 1, 2, 3),
                  init_cfg=None,
                  **kwargs):
@@ -591,6 +596,7 @@ class InternImage(nn.Module):
         logger.info(f"level2_post_norm: {level2_post_norm}")
         logger.info(f"level2_post_norm_block_ids: {level2_post_norm_block_ids}")
         logger.info(f"res_post_norm: {res_post_norm}")
+        logger.info(f"use_dcn_v4_op: {use_dcn_v4_op}")
 
         in_chans = 3
         self.patch_embed = StemLayer(in_chans=in_chans,
@@ -611,7 +617,7 @@ class InternImage(nn.Module):
             post_norm_block_ids = level2_post_norm_block_ids if level2_post_norm and (
                 i == 2) else None # for InternImage-H/G
             level = InternImageBlock(
-                core_op=getattr(opsm, core_op),
+                core_op=getattr(dcnv3, core_op),
                 channels=int(channels * 2**i),
                 depth=depths[i],
                 groups=groups[i],
@@ -628,7 +634,8 @@ class InternImage(nn.Module):
                 dw_kernel_size=dw_kernel_size,  # for InternImage-H/G
                 post_norm_block_ids=post_norm_block_ids, # for InternImage-H/G
                 res_post_norm=res_post_norm, # for InternImage-H/G
-                center_feature_scale=center_feature_scale # for InternImage-H/G
+                center_feature_scale=center_feature_scale, # for InternImage-H/G
+                use_dcn_v4_op=use_dcn_v4_op,
             )
             self.levels.append(level)
 
@@ -687,7 +694,7 @@ class InternImage(nn.Module):
             nn.init.constant_(m.weight, 1.0)
 
     def _init_deform_weights(self, m):
-        if isinstance(m, getattr(opsm, self.core_op)):
+        if isinstance(m, getattr(dcnv3, self.core_op)):
             m._reset_parameters()
 
     def forward(self, x):

segmentation/ops_dcnv3/modules/dcnv3.py

@@ -14,7 +14,11 @@ from torch import nn
 import torch.nn.functional as F
 from torch.nn.init import xavier_uniform_, constant_
 from ..functions import DCNv3Function, dcnv3_core_pytorch
-
+try:
+    from DCNv4.functions import DCNv4Function
+except:
+    warnings.warn('Now, we support DCNv4 in InternImage.')
+import math
 
 class to_channels_first(nn.Module):
 
@@ -228,7 +232,9 @@ class DCNv3(nn.Module):
             offset_scale=1.0,
             act_layer='GELU',
             norm_layer='LN',
-            center_feature_scale=False):
+            center_feature_scale=False,
+            use_dcn_v4_op=False,
+            ):
         """
         DCNv3 Module
         :param channels
@@ -264,7 +270,9 @@ class DCNv3(nn.Module):
         self.group_channels = channels // group
         self.offset_scale = offset_scale
         self.center_feature_scale = center_feature_scale
-        
+         
+        self.use_dcn_v4_op = use_dcn_v4_op
+
         self.dw_conv = nn.Sequential(
             nn.Conv2d(
                 channels,
@@ -321,18 +329,44 @@ class DCNv3(nn.Module):
         x1 = self.dw_conv(x1)
         offset = self.offset(x1)
         mask = self.mask(x1).reshape(N, H, W, self.group, -1)
-        mask = F.softmax(mask, -1).reshape(N, H, W, -1).type(dtype)
-
-        x = DCNv3Function.apply(
-            x, offset, mask,
-            self.kernel_size, self.kernel_size,
-            self.stride, self.stride,
-            self.pad, self.pad,
-            self.dilation, self.dilation,
-            self.group, self.group_channels,
-            self.offset_scale,
-            256)
         
+        if not self.use_dcn_v4_op:
+            mask = F.softmax(mask, -1).reshape(N, H, W, -1).type(dtype)
+            x = DCNv3Function.apply(
+                x, offset, mask,
+                self.kernel_size, self.kernel_size,
+                self.stride, self.stride,
+                self.pad, self.pad,
+                self.dilation, self.dilation,
+                self.group, self.group_channels,
+                self.offset_scale,
+                256)
+        else:
+            # DCNv4 combines offset and weight mask into one tensor `offset_mask`.
+            # The following code is to align DCNv3 and DCNv4
+            offset = offset.view(N, H, W, self.group, -1)
+            mask = F.softmax(mask, -1)
+            mask = mask.view(N, H, W, self.group, -1)
+            offset_mask = torch.cat([offset, mask], -1).view(N, H, W, -1).contiguous()
+
+            # For efficiency, the last dimension of the offset_mask tensor in dcnv4 is a multiple of 8. 
+            K3 = offset_mask.size(-1)
+            K3_pad = int(math.ceil(K3/8)*8)
+            pad_dim = K3_pad - K3
+            offset_mask = torch.cat([offset_mask, offset_mask.new_zeros([*offset_mask.size()[:3], pad_dim])], -1)
+        
+            x = DCNv4Function.apply(
+                x, offset_mask,
+                self.kernel_size, self.kernel_size,
+                self.stride, self.stride,
+                self.pad, self.pad,
+                self.dilation, self.dilation,
+                self.group, self.group_channels,
+                self.offset_scale,
+                256,
+                False
+            )
+
         if self.center_feature_scale:
             center_feature_scale = self.center_feature_scale_module(
                 x1, self.center_feature_scale_proj_weight, self.center_feature_scale_proj_bias)