xx: change back downsampler1d conv to 2d (pseudo), make the whole nec…

…k pseudo 2D
makecent · Jun 29, 2023 · 90195fb · 90195fb
1 parent b846003
commit 90195fb
Show file tree

Hide file tree

Showing 3 changed files with 6 additions and 15 deletions.
diff --git a/configs/tadtr_my.py b/configs/tadtr_my.py
@@ -54,7 +54,6 @@
             kernel_size=1,
             out_channels=256,
             act_cfg=None,
-            conv_cfg=dict(type='Conv1d'),
             norm_cfg=dict(type='GN', num_groups=32),
             num_outs=4)
         # dict(type='FPN',

diff --git a/my_modules/detector/deformable_detr.py b/my_modules/detector/deformable_detr.py
@@ -32,13 +32,6 @@ def _init_layers(self) -> None:
         if not self.as_two_stage:
             self.reference_points_fc = Pseudo2DLinear(self.embed_dims, 1)
 
-    def extract_feat(self, batch_inputs: Tensor) -> Tuple[Tensor]:
-        x = self.backbone(batch_inputs)
-        if self.with_neck:
-            x = self.neck(x)
-        # insert the pseudo height dimension
-        x = tuple(i.unsqueeze(2) for i in x)
-        return x
 
     def forward_transformer(self,
                             img_feats: Tuple[Tensor],

diff --git a/my_modules/neck/temporal_downsampler.py b/my_modules/neck/temporal_downsampler.py
@@ -12,9 +12,9 @@ def __init__(self,
                  num_levels=4,
                  in_channels=2048,
                  out_channels=512,
-                 kernel_sizes=3,
-                 strides=2,
-                 paddings=1,
+                 kernel_sizes=(1, 3),
+                 strides=(1, 2),
+                 paddings=(0, 1),
                  out_indices=(0, 1, 2, 3),
                  mask=False,
                  ):
@@ -36,7 +36,7 @@ def __init__(self,
                                         kernel_sizes,
                                         strides,
                                         paddings,
-                                        conv_cfg=dict(type='Conv1d'),
+                                        conv_cfg=dict(type='Conv2d'),
                                         norm_cfg=dict(type='SyncBN'),
                                         act_cfg=dict(type='ReLU')))
             in_channels = out_channels
@@ -61,16 +61,15 @@ def __init__(self,
 
     def forward(self, x):
         # x: N, C, 1, T
-        x = x.squeeze(2)
 
         if self.mask:
-            B, C, T = x.size()
+            B, C, _, T = x.size()
             mask = x.new_zeros((B, 1, T))
             for feat_id, feat in enumerate(x):
                 # find valid feature length for each sample in the batch
                 # feat is of shape (C, T), we find the tails that are all zeros and take it as the padding
                 valid_feat_len = T - (feat == 0).all(dim=0).sum()
-                mask[feat_id, :, :valid_feat_len] = 1
+                mask[feat_id, :, :, :valid_feat_len] = 1
 
         outs = []
         if 0 in self.out_indices: