Detectron2-ResNeSt
See details in [Detectron2-ResNeSt repo](https://github.com/zhanghang1989/detectron2-ResNeSt)
Object Detection on MS-COCO validation set
<table class="tg">
<tr>
<th class="tg-0pky">Method</th>
<th class="tg-0pky">Backbone</th>
<th class="tg-0pky">mAP%</th>
</tr>
<tr>
<td rowspan="4" class="tg-0pky">Faster R-CNN</td>
<td class="tg-0pky">ResNet-50</td>
<td class="tg-0pky">39.25</td>
</tr>
<tr>
<td class="tg-0lax">ResNet-101</td>
<td class="tg-0lax">41.37</td>
</tr>
<tr>
<td class="tg-0lax">ResNeSt-50 (<span style="color:red">ours</span>)</td>
<td class="tg-0lax"><b>42.33</b></td>
</tr>
<tr>
<td class="tg-0lax">ResNeSt-101 (<span style="color:red">ours</span>)</td>
<td class="tg-0lax"><b>44.72</b></td>
</tr>
<tr>
<td rowspan="5" class="tg-0lax">Cascade R-CNN</td>
<td class="tg-0lax">ResNet-50</td>
<td class="tg-0lax">42.52</td>
</tr>
<tr>
<td class="tg-0lax">ResNet-101</td>
<td class="tg-0lax">44.03</td>
</tr>
<tr>
<td class="tg-0lax">ResNeSt-50 (<span style="color:red">ours</span>)</td>
<td class="tg-0lax"><b>45.41</b></td>
</tr>
<tr>
<td class="tg-0lax">ResNeSt-101 (<span style="color:red">ours</span>)</td>
<td class="tg-0lax"><b>47.50</b></td>
</tr>
<tr>
<td class="tg-0lax">ResNeSt-200 (<span style="color:red">ours</span>)</td>
<td class="tg-0lax"><b>49.03</b></td>
</tr>
</table>
Instance Segmentation
<table class="tg">
<tr>
<th class="tg-0pky">Method</th>
<th class="tg-0pky">Backbone</th>
<th class="tg-0pky">bbox</th>
<th class="tg-0lax">mask</th>
</tr>
<tr>
<td rowspan="4" class="tg-0pky">Mask R-CNN</td>
<td class="tg-0pky">ResNet-50</td>
<td class="tg-0pky">39.97</td>
<td class="tg-0lax">36.05</td>
</tr>
<tr>
<td class="tg-0lax">ResNet-101</td>
<td class="tg-0lax">41.78</td>
<td class="tg-0lax">37.51</td>
</tr>
<tr>
<td class="tg-0lax">ResNeSt-50 (<span style="color:red">ours</span>)</td>
<td class="tg-0lax"><b>42.81</b></td>
<td class="tg-0lax"><b>38.14</td>
</tr>
<tr>
<td class="tg-0lax">ResNeSt-101 (<span style="color:red">ours</span>)</td>
<td class="tg-0lax"><b>45.75</b></td>
<td class="tg-0lax"><b>40.65</b></td>
</tr>
<tr>
<td rowspan="4" class="tg-0lax">Cascade R-CNN</td>
<td class="tg-0lax">ResNet-50</td>
<td class="tg-0lax">43.06</td>
<td class="tg-0lax">37.19</td>
</tr>
<tr>
<td class="tg-0lax">ResNet-101</td>
<td class="tg-0lax">44.79</td>
<td class="tg-0lax">38.52</td>
</tr>
<tr>
<td class="tg-0lax">ResNeSt-50 (<span style="color:red">ours</span>)</td>
<td class="tg-0lax"><b>46.19</b></td>
<td class="tg-0lax"><b>39.55</b></td>
</tr>
<tr>
<td class="tg-0lax">ResNeSt-101 (<span style="color:red">ours</span>)</td>
<td class="tg-0lax"><b>48.30</b></td>
<td class="tg-0lax"><b>41.56</b></td>
</tr>
</table>
Semantic Segmentation with New SoTA on ADE20K
Semantic Segmentation
- PyTorch models and training: Please visit [PyTorch Encoding Toolkit](https://hangzhang.org/PyTorch-Encoding/model_zoo/segmentation.html).
- Training with Gluon: Please visit [GluonCV Toolkit](https://gluon-cv.mxnet.io/model_zoo/segmentation.html#ade20k-dataset).
Results on ADE20K
<table class="tg">
<tr>
<th class="tg-cly1">Method</th>
<th class="tg-cly1">Backbone</th>
<th class="tg-cly1">pixAcc%</th>
<th class="tg-cly1">mIoU%</th>
</tr>
<tr>
<td rowspan="5" class="tg-cly1">Deeplab-V3<br></td>
<td class="tg-cly1">ResNet-50</td>
<td class="tg-cly1">80.39</td>
<td class="tg-cly1">42.1</td>
</tr>
<tr>
<td class="tg-cly1">ResNet-101</td>
<td class="tg-cly1">81.11</b></td>
<td class="tg-cly1">44.14</b></td>
</tr>
<tr>
<td class="tg-cly1">ResNeSt-50 (<span style="color:red">ours</span>)</td>
<td class="tg-cly1"><b>81.17</b></td>
<td class="tg-cly1"><b>45.12</b></td>
</tr>
<tr>
<td class="tg-0lax">ResNeSt-101 (<span style="color:red">ours</span>)</td>
<td class="tg-0lax"><b>82.07</td>
<td class="tg-0lax"><b>46.91</b></td>
</tr>
<tr>
<td class="tg-0lax">ResNeSt-269 (<span style="color:red">ours</span>)</td>
<td class="tg-0lax"><b>82.62</td>
<td class="tg-0lax"><b>47.60</b></td>
</tr>
</table>
Ablation Study Models
| | setting | P | GFLOPs | PyTorch | Gluon |
|-----------------|---------|-------|--------|---------|-------|
| ResNeSt-50-fast | 1s1x64d | 26.3M | 4.34 | 80.33 | 80.35 |
| ResNeSt-50-fast | 2s1x64d | 27.5M | 4.34 | 80.53 | 80.65 |
| ResNeSt-50-fast | 4s1x64d | 31.9M | 4.35 | 80.76 | 80.90 |
| ResNeSt-50-fast | 1s2x40d | 25.9M | 4.38 | 80.59 | 80.72 |
| ResNeSt-50-fast | 2s2x40d | 26.9M | 4.38 | 80.61 | 80.84 |
| ResNeSt-50-fast | 4s2x40d | 30.4M | 4.41 | 81.14 | 81.17 |
| ResNeSt-50-fast | 1s4x24d | 25.7M | 4.42 | 80.99 | 80.97 |
ImageNet Training with MXNet Gluon
Install MXNet with Horovod
bash
assuming you have CUDA 10.0 on your machine
pip install mxnet-cu100
HOROVOD_GPU_ALLREDUCE=NCCL pip install -v --no-cache-dir horovod
pip install --no-cache mpi4py
Prepare ImageNet recordio data format
- Unfortunately ,this is required for training using MXNet Gluon. Please follow the [GluonCV tutorial](https://gluon-cv.mxnet.io/build/examples_datasets/recordio.html) to prepare the data.
- Copy the data into ramdisk (optional):
cd ~/
sudo mkdir -p /media/ramdisk
sudo mount -t tmpfs -o size=200G tmpfs /media/ramdisk
cp -r /home/ubuntu/data/ILSVRC2012/ /media/ramdisk
Training command
Using ResNeSt-50 as the target model:
bash
horovodrun -np 64 --hostfile hosts python train.py \
--rec-train /media/ramdisk/ILSVRC2012/train.rec \
--rec-val /media/ramdisk/ILSVRC2012/val.rec \
--model resnest50 --lr 0.05 --num-epochs 270 --batch-size 128 \
--use-rec --dtype float32 --warmup-epochs 5 --last-gamma --no-wd \
--label-smoothing --mixup --save-dir params_ resnest50 \
--log-interval 50 --eval-frequency 5 --auto_aug --input-size 224
Verify pretrained model
bash
python verify.py --model resnest50 --crop-size 224 --resume params_ resnest50/imagenet-resnest50-269.params