Skip to content

Commit

Permalink
update ViT-Adapter-L + HTC++
Browse files Browse the repository at this point in the history
  • Loading branch information
@czczup
czczup committed Jun 9, 2022
1 parent 9441aef commit 05705cf
Show file tree
Hide file tree
Showing 4 changed files with 431 additions and 14 deletions.
105 changes: 95 additions & 10 deletions detection/configs/htc++/README.md
View file
Original file line number Diff line number Diff line change
Expand Up @@ -27,19 +27,104 @@ detection

## Results and Models

The results on **COCO 2017val** are shown in the below table.
The results on COCO mini-val and test-dev are shown in the below table.

| Backbone | Pre-train | Lr schd | box AP | mask AP | #Param | Config | Download |
|:------------------:|:---------------------------------------------------------------------------------------------------------------------:|:-------:|:------:|:-------:|:------:|:---------------------------------------------------:|:----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------:|
| ViT-Adapter-L | [BEiT-L](https://conversationhub.blob.core.windows.net/beit-share-public/beit/beit_large_patch16_224_pt22k_ft22k.pth) | 3x | 57.9 | 50.2 | 401M | [config](./htc++_beit_adapter_large_fpn_3x_coco.py) | [model](https://github.com/czczup/ViT-Adapter/releases/download/v0.1.0/htc++_beit_adapter_large_fpn_3x_coco.pth.tar) \| [log](https://github.com/czczup/ViT-Adapter/releases/download/v0.1.0/20220512_030903.log) |
| ViT-Adapter-L (MS) | - | - | 59.8 | 51.7 | 401M | TODO | - |
<table>
<tr align=center>
<td rowspan="2" align=center><b>Backbone</b></td>
<td rowspan="2" align=center><b>Pre-train</b></td>
<td rowspan="2" align=center><b>Lr schd</b></td>
<td colspan="2" align=center><b>mini-val</b></td>
<td colspan="2" align=center><b>test-dev</b></td>
<td rowspan="2" align=center><b>#Param</b></td>
<td rowspan="2" align=center><b>Config</b></td>
<td rowspan="2" align=center><b>Download</b></td>
</tr>
<tr>
<td>box AP</td>
<td>mask AP</td>
<td>box AP</td>
<td>mask AP</td>
</tr>
<tr align=center>
<td>ViT-Adapter-L</td>
<td><a href="https://conversationhub.blob.core.windows.net/beit-share-public/beit/beit_large_patch16_224_pt22k_ft22k.pth">BEiT-L</a></td>
<td>3x</td>
<td>58.4</td>
<td>50.8</td>
<td>-</td>
<td>-</td>
<td>401M</td>
<td><a href="./htc++_beit_adapter_large_fpn_3x_coco.py">config</a> </td>
<td><a href="https://github.com/czczup/ViT-Adapter/releases/download/v0.1.0/htc++_beit_adapter_large_fpn_3x_coco.pth.tar">model</a> |
<a href="https://github.com/czczup/ViT-Adapter/releases/download/v0.1.0/20220512_030903.log">log</a></td>
</tr>
</tr>
<tr align=center>
<td>ViT-Adapter-L (MS)</td>
<td><a href="https://conversationhub.blob.core.windows.net/beit-share-public/beit/beit_large_patch16_224_pt22k_ft22k.pth">BEiT-L</a></td>
<td>3x</td>
<td>60.2</td>
<td>52.2</td>
<td>-</td>
<td>-</td>
<td>401M</td>
<td>-</td>
<td>-</td>
</tr>
</table>

- MS denotes multi-scale testing. Note that the ms config is only for testing.
- We use 16 A100 GPUs with 1 image/GPU for ViT-Adapter-L models.

The results on **COCO 2017test-dev** are shown in the below table.
## Old Results

| Backbone | Pre-train | Lr schd | box AP | mask AP | #Param | Config | Download |
|:------------------:|:---------------------------------------------------------------------------------------------------------------------:|:-------:|:------:|:-------:|:------:|:---------------------------------------------------:|:----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------:|
| ViT-Adapter-L | [BEiT-L](https://conversationhub.blob.core.windows.net/beit-share-public/beit/beit_large_patch16_224_pt22k_ft22k.pth) | 3x | 58.5 | 50.8 | 401M | [config](./htc++_beit_adapter_large_fpn_3x_coco.py) | [model](https://github.com/czczup/ViT-Adapter/releases/download/v0.1.0/htc++_beit_adapter_large_fpn_3x_coco.pth.tar) \| [log](https://github.com/czczup/ViT-Adapter/releases/download/v0.1.0/20220512_030903.log) |
| ViT-Adapter-L (MS) | - | - | 60.1 | 52.1 | 401M | TODO | - |
The results on COCO mini-val and test-dev are shown in the below table.

<table>
<tr align=center>
<td rowspan="2" align=center><b>Backbone</b></td>
<td rowspan="2" align=center><b>Pre-train</b></td>
<td rowspan="2" align=center><b>Lr schd</b></td>
<td colspan="2" align=center><b>mini-val</b></td>
<td colspan="2" align=center><b>test-dev</b></td>
<td rowspan="2" align=center><b>#Param</b></td>
<td rowspan="2" align=center><b>Config</b></td>
<td rowspan="2" align=center><b>Download</b></td>
</tr>
<tr>
<td>box AP</td>
<td>mask AP</td>
<td>box AP</td>
<td>mask AP</td>
</tr>
<tr align=center>
<td>ViT-Adapter-L</td>
<td><a href="https://conversationhub.blob.core.windows.net/beit-share-public/beit/beit_large_patch16_224_pt22k_ft22k.pth">BEiT-L</a></td>
<td>3x</td>
<td>57.9</td>
<td>50.2</td>
<td>58.5</td>
<td>50.8</td>
<td>401M</td>
<td><a href="./htc++_beit_adapter_large_fpn_3x_coco_old.py">config</a> </td>
<td><a href="https://github.com/czczup/ViT-Adapter/releases/download/v0.1.0/htc++_beit_adapter_large_fpn_3x_coco_old.pth.tar">model</a> |
<a href="https://github.com/czczup/ViT-Adapter/releases/download/v0.1.0/20220512_030903.log">log</a></td>
</tr>
</tr>
<tr align=center>
<td>ViT-Adapter-L (MS)</td>
<td><a href="https://conversationhub.blob.core.windows.net/beit-share-public/beit/beit_large_patch16_224_pt22k_ft22k.pth">BEiT-L</a></td>
<td>3x</td>
<td>59.8</td>
<td>51.7</td>
<td>60.1</td>
<td>52.1</td>
<td>401M</td>
<td>-</td>
<td>-</td>
</tr>
</table>

- MS denotes multi-scale testing. Note that the ms config is only for testing.
- We use 16 A100 GPUs with 1 image/GPU for ViT-Adapter-L models.
2 changes: 1 addition & 1 deletion detection/configs/htc++/htc++_beit_adapter_large_fpn_3x_coco.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -20,7 +20,7 @@
use_abs_pos_emb=False,
use_rel_pos_bias=True,
init_values=1e-6,
drop_path_rate=0.3, # maybe 0.4 is better
drop_path_rate=0.4,
conv_inplane=64,
n_points=4,
deform_num_heads=16,
Expand Down
Loading

0 comments on commit 05705cf

Please sign in to comment.