Fix offset norm hw order bug in MSDeformAttn module.

README.md

@@ -19,6 +19,10 @@ This repository is an official implementation of the paper [Deformable DETR: Def
 
 This project is released under the [Apache 2.0 license](./LICENSE).
 
+## Changelog
+
+See [changelog.md](./docs/changelog.md) for detailed logs of major changes. 
+
 
 ## Citing Deformable DETR
 If you find Deformable DETR useful in your research, please consider citing:
@@ -40,11 +44,11 @@ If you find Deformable DETR useful in your research, please consider citing:
 | <sub><sub>DETR-DC5</sub></sub>      | <sub>500</sub> | <sub>43.3</sub> | <sub>22.5</sub> | <sub>47.3</sub> | <sub>61.1</sub> | <sub>41</sub> |<sub>187</sub>|<sub>7000</sub>|<sub>14.0</sub>|<sub>11.4</sub>|<sub>12.4</sub>| <sub>-</sub> |
 | <sub><sub>DETR-DC5</sub></sub>      | <sub>50</sub> | <sub>35.3</sub> | <sub>15.2</sub> | <sub>37.5</sub> | <sub>53.6</sub> | <sub>41</sub> |<sub>187</sub>|<sub>700</sub>|<sub>14.0</sub>|<sub>11.4</sub>|<sub>12.4</sub>| <sub>-</sub> |
 | <sub><sub>DETR-DC5+</sub></sub>     | <sub>50</sub> | <sub>36.2</sub> | <sub>16.3</sub> | <sub>39.2</sub> | <sub>53.9</sub> | <sub>41</sub> |<sub>187</sub>|<sub>700</sub>|<sub>14.0</sub>|<sub>11.4</sub>|<sub>12.4</sub>| <sub>-</sub> |
-| **<sub><sub>Deformable DETR<br>(single scale)</sub></sub>** | <sub>50</sub> | <sub>39.1</sub> | <sub>20.2</sub> | <sub>43.1</sub> | <sub>55.6</sub> | <sub>34</sub> |<sub>78</sub>|<sub>160</sub>|<sub>3.2</sub>|<sub>27.0</sub>|<sub>42.4</sub>| <sub>[config](./configs/r50_deformable_detr_single_scale.sh)<br/>[log](https://drive.google.com/file/d/1AiXs9dHVl1jRjS2s6rfS9c5wuaXkAZk_/view?usp=sharing)<br/>[model](https://drive.google.com/file/d/1635D3bcdC3mRQr-ET5l5yI8eNkuy1i_2/view?usp=sharing)</sub> |
-| **<sub><sub>Deformable DETR<br>(single scale, DC5)</sub></sub>** | <sub>50</sub> | <sub>41.0</sub> | <sub>23.1</sub> | <sub>45.2</sub> | <sub>55.9</sub> | <sub>34</sub> |<sub>128</sub>|<sub>215</sub>|<sub>4.3</sub>|<sub>22.1</sub>|<sub>29.4</sub>| <sub>[config](./configs/r50_deformable_detr_single_scale_dc5.sh)<br/>[log](https://drive.google.com/file/d/1rDFFXkWOVupuhP9IhP-hGwAAdlFshy1E/view?usp=sharing)<br/>[model](https://drive.google.com/file/d/1bRgqYfQpMdSqe6U56HnW9ytF9VynDTCY/view?usp=sharing)</sub> |
-| **<sub><sub>Deformable DETR</sub></sub>** | <sub>50</sub> | <sub>43.8</sub> | <sub>26.3</sub> | <sub>46.8</sub> | <sub>58.2</sub> | <sub>40</sub> |<sub>173</sub>|<sub>325</sub>|<sub>6.5</sub>|<sub>15.0</sub>|<sub>19.4</sub>|<sub>[config](./configs/r50_deformable_detr.sh)<br/>[log](https://drive.google.com/file/d/1VB4vWZrF2xcU8d8PfBgtzyS5q0HbQ3xB/view?usp=sharing)<br/>[model](https://drive.google.com/file/d/1vw1PssvHBhLQI8s5wRWDjl9yDMfqElbg/view?usp=sharing)</sub>                   |
-| **<sub><sub>+ iterative bounding box refinement</sub></sub>** | <sub>50</sub> | <sub>45.8</sub> | <sub>28.4</sub> | <sub>49.0</sub> | <sub>61.8</sub> | <sub>41</sub> |<sub>173</sub>|<sub>325</sub>|<sub>6.5</sub>|<sub>15.0</sub>|<sub>19.4</sub>|<sub>[config](./configs/r50_deformable_detr_plus_iterative_bbox_refinement.sh)<br/>[log](https://drive.google.com/file/d/1Bp3EVhsvdUeNu4ASlBkTgNWloIkL1oi8/view?usp=sharing)<br/>[model](https://drive.google.com/file/d/1GhgOyWdSC-rPkDI1Gz2aDaj97pC_dozl/view?usp=sharing)</sub> |
-| **<sub><sub>++ two-stage Deformable DETR</sub></sub>** | <sub>50</sub> | <sub>46.4</sub> | <sub>28.6</sub> | <sub>49.7</sub> | <sub>61.3</sub> | <sub>41</sub> |<sub>173</sub>|<sub>340</sub>|<sub>6.8</sub>|<sub>14.5</sub>|<sub>18.8</sub>|<sub>[config](./configs/r50_deformable_detr_plus_iterative_bbox_refinement_plus_plus_two_stage.sh)<br/>[log](https://drive.google.com/file/d/1VTfDrBxEfpwXoC1aNkN70DI9WRI_JS3z/view?usp=sharing) <br/>[model](https://drive.google.com/file/d/1Xfsg7-glLyJOtH-Q0qt6QVlz23J6t8xT/view?usp=sharing)</sub> |
+| **<sub><sub>Deformable DETR<br>(single scale)</sub></sub>** | <sub>50</sub> | <sub>39.4</sub> | <sub>20.6</sub> | <sub>43.0</sub> | <sub>55.5</sub> | <sub>34</sub> |<sub>78</sub>|<sub>160</sub>|<sub>3.2</sub>|<sub>27.0</sub>|<sub>42.4</sub>| <sub>[config](./configs/r50_deformable_detr_single_scale.sh)<br/>[log](https://drive.google.com/file/d/1n3ZnZ-UAqmTUR4AZoM4qQntIDn6qCZx4/view?usp=sharing)<br/>[model](https://drive.google.com/file/d/1WEjQ9_FgfI5sw5OZZ4ix-OKk-IJ_-SDU/view?usp=sharing)</sub> |
+| **<sub><sub>Deformable DETR<br>(single scale, DC5)</sub></sub>** | <sub>50</sub> | <sub>41.5</sub> | <sub>24.1</sub> | <sub>45.3</sub> | <sub>56.0</sub> | <sub>34</sub> |<sub>128</sub>|<sub>215</sub>|<sub>4.3</sub>|<sub>22.1</sub>|<sub>29.4</sub>| <sub>[config](./configs/r50_deformable_detr_single_scale_dc5.sh)<br/>[log](https://drive.google.com/file/d/1-UfTp2q4GIkJjsaMRIkQxa5k5vn8_n-B/view?usp=sharing)<br/>[model](https://drive.google.com/file/d/1m_TgMjzH7D44fbA-c_jiBZ-xf-odxGdk/view?usp=sharing)</sub> |
+| **<sub><sub>Deformable DETR</sub></sub>** | <sub>50</sub> | <sub>44.5</sub> | <sub>27.1</sub> | <sub>47.6</sub> | <sub>59.6</sub> | <sub>40</sub> |<sub>173</sub>|<sub>325</sub>|<sub>6.5</sub>|<sub>15.0</sub>|<sub>19.4</sub>|<sub>[config](./configs/r50_deformable_detr.sh)<br/>[log](https://drive.google.com/file/d/18YSLshFjc_erOLfFC-hHu4MX4iyz1Dqr/view?usp=sharing)<br/>[model](https://drive.google.com/file/d/1nDWZWHuRwtwGden77NLM9JoWe-YisJnA/view?usp=sharing)</sub>                   |
+| **<sub><sub>+ iterative bounding box refinement</sub></sub>** | <sub>50</sub> | <sub>46.2</sub> | <sub>28.3</sub> | <sub>49.2</sub> | <sub>61.5</sub> | <sub>41</sub> |<sub>173</sub>|<sub>325</sub>|<sub>6.5</sub>|<sub>15.0</sub>|<sub>19.4</sub>|<sub>[config](./configs/r50_deformable_detr_plus_iterative_bbox_refinement.sh)<br/>[log](https://drive.google.com/file/d/1DFNloITi1SFBWjYzvVEAI75ndwmGM1Uj/view?usp=sharing)<br/>[model](https://drive.google.com/file/d/1JYKyRYzUH7uo9eVfDaVCiaIGZb5YTCuI/view?usp=sharing)</sub> |
+| **<sub><sub>++ two-stage Deformable DETR</sub></sub>** | <sub>50</sub> | <sub>46.9</sub> | <sub>29.6</sub> | <sub>50.1</sub> | <sub>61.6</sub> | <sub>41</sub> |<sub>173</sub>|<sub>340</sub>|<sub>6.8</sub>|<sub>14.5</sub>|<sub>18.8</sub>|<sub>[config](./configs/r50_deformable_detr_plus_iterative_bbox_refinement_plus_plus_two_stage.sh)<br/>[log](https://drive.google.com/file/d/1ozi0wbv5-Sc5TbWt1jAuXco72vEfEtbY/view?usp=sharing) <br/>[model](https://drive.google.com/file/d/15I03A7hNTpwuLNdfuEmW9_taZMNVssEp/view?usp=sharing)</sub> |
 
 *Note:*
 

docs/changelog.md

@@ -0,0 +1,3 @@
+## Changelog
+
+**[2020.12.07]** Fix a bug of sampling offset normalization (see [this issue](https://github.com/fundamentalvision/Deformable-DETR/issues/6)) in the MSDeformAttn module. The final accuracy on COCO is slightly improved. Code and pre-trained models have been updated. This bug only occurs in this released version but not in the original implementation used in our paper.

models/ops/modules/ms_deform_attn.py

@@ -100,8 +100,9 @@ def forward(self, query, reference_points, input_flatten, input_spatial_shapes,
         attention_weights = F.softmax(attention_weights, -1).view(N, Len_q, self.n_heads, self.n_levels, self.n_points)
         # N, Len_q, n_heads, n_levels, n_points, 2
         if reference_points.shape[-1] == 2:
+            offset_normalizer = torch.stack([input_spatial_shapes[..., 1], input_spatial_shapes[..., 0]], -1)
             sampling_locations = reference_points[:, :, None, :, None, :] \
-                                 + sampling_offsets / input_spatial_shapes[None, None, None, :, None, :]
+                                 + sampling_offsets / offset_normalizer[None, None, None, :, None, :]
         elif reference_points.shape[-1] == 4:
             sampling_locations = reference_points[:, :, None, :, None, :2] \
                                  + sampling_offsets / self.n_points * reference_points[:, :, None, :, None, 2:] * 0.5