Update Deformable-DETR-Box-Refinement results (#102)

* refine README

* update model zoo

Co-authored-by: ntianhe ren <rentianhe@dgx061.scc.idea>

docs/source/tutorials/Model_Zoo.md

@@ -24,8 +24,8 @@ Here we provides our pretrained baselines in `detrex beta v0.1.0`. And more pret
 <td align="center">R50</td>
 <td align="center">IN1k</td>
 <td align="center">50</td>
-<td align="center">46.32</td>
-<td align="center"> <a href="https://github.com/IDEA-Research/detrex-storage/releases/download/v0.1.0/deformable_detr_with_box_refinement_50ep.pth"> model </a></td>
+<td align="center">46.99</td>
+<td align="center"> <a href="https://github.com/IDEA-Research/detrex-storage/releases/download/v0.1.1/deformable_detr_with_box_refinement_50ep_new.pth"> model </a></td>
 </tr>
  <tr><td align="left"> <a href="https://github.com/IDEA-Research/detrex/blob/main/projects/deformable_detr/configs/deformable_detr_r50_two_stage_50ep.py"> Deformable-DETR + Box Refinement + Two Stage </a> </td>
 <td align="center">R50</td>

projects/deformable_detr/README.md

@@ -26,8 +26,8 @@ Here we provide the pretrained `Deformable-DETR` weights based on detrex.
 <td align="center">R-50</td>
 <td align="center">IN1k</td>
 <td align="center">50</td>
-<td align="center">46.32</td>
-<td align="center"> <a href="https://github.com/IDEA-Research/detrex-storage/releases/download/v0.1.0/deformable_detr_with_box_refinement_50ep.pth">model</a></td>
+<td align="center">46.99</td>
+<td align="center"> <a href="https://github.com/IDEA-Research/detrex-storage/releases/download/v0.1.1/deformable_detr_with_box_refinement_50ep_new.pth">model</a></td>
 </tr>
 <!-- ROW: deformable_detr_r50_two_stage_50ep -->
  <tr><td align="left"><a href="configs/deformable_detr_r50_two_stage_50ep.py">Deformable-DETR-R50 + Box-Refinement + Two-Stage</a></td>

projects/deformable_detr/configs/deformable_detr_r50_with_box_refinement_50ep.py

@@ -1,8 +1,22 @@
 from .deformable_detr_r50_50ep import train, dataloader, optimizer, lr_multiplier, model
 
 # modify model config
+# set loss_class to 1.0 brings better results for deformable-detr-box-refinement under lr=1e-4
 model.with_box_refine = True
+model.criterion.weight_dict = {
+    "loss_class": 1.0,
+    "loss_bbox": 5.0,
+    "loss_giou": 2.0,
+}
+if model.aux_loss:
+    weight_dict = model.criterion.weight_dict
+    aux_weight_dict = {}
+    for i in range(model.transformer.decoder.num_layers - 1):
+        aux_weight_dict.update({k + f"_{i}": v for k, v in weight_dict.items()})
+    aux_weight_dict.update({k + "_enc": v for k, v in weight_dict.items()})
+    weight_dict.update(aux_weight_dict)
+    model.criterion.weight_dict = weight_dict
 
 # modify training config
 train.init_checkpoint = "detectron2://ImageNetPretrained/torchvision/R-50.pkl"
-train.output_dir = "./output/deformable_detr_with_box_refinement_50ep"
+train.output_dir = "./output/deformable_detr_with_box_refinement_50ep"