upload pytorch weights and fix detach bug

Author: lyuwenyu

rtdetr_pytorch/README.md

@@ -3,13 +3,13 @@
 - [x] Evaluation
 - [x] Export onnx
 - [x] Upload source code
-- [ ] Align with the [Paddle version](../rtdetr_paddle/)
-- [ ] Upload weight convert from paddle
+- [x] Upload weight convert from paddle, see [links](https://github.com/lyuwenyu/RT-DETR/issues/42)
+- [ ] Align training with the [paddle version](../rtdetr_paddle/)
 
 
 ## Quick start
 
-<details open>
+<details>
 <summary>Install</summary>
 
 ```bash
@@ -76,9 +76,9 @@ python tools/export_onnx.py -c configs/rtdetr/rtdetr_r18vd_6x_coco.yml -r path/t
 
 
 
-<details>
-<summary>Finetune</summary>
+<details open>
+<summary>Train custom data</summary>
 
-Set `remap_mscoco_category: False`. This var only works for ms-coco dataset.
+set `remap_mscoco_category: False`. This variable only works for ms-coco dataset.
 
 </details>

rtdetr_pytorch/src/zoo/rtdetr/rtdetr_decoder.py

@@ -507,11 +507,12 @@ def _get_decoder_input(self,
         else:
             target = output_memory.gather(dim=1, \
                 index=topk_ind.unsqueeze(-1).repeat(1, 1, output_memory.shape[-1]))
+            target = target.detach()
 
         if denoising_class is not None:
             target = torch.concat([denoising_class, target], 1)
 
-        return target.detach(), reference_points_unact.detach(), enc_topk_bboxes, enc_topk_logits
+        return target, reference_points_unact.detach(), enc_topk_bboxes, enc_topk_logits
 
 
     def forward(self, feats, targets=None):