pytorch-deformable-convolution-v2
Don't feel pain to use Deformable Convolution
https://github.com/developer0hye/pytorch-deformable-convolution-v2
Science Score: 54.0%
This score indicates how likely this project is to be science-related based on various indicators:
-
✓CITATION.cff file
Found CITATION.cff file -
✓codemeta.json file
Found codemeta.json file -
✓.zenodo.json file
Found .zenodo.json file -
○DOI references
-
✓Academic publication links
Links to: arxiv.org -
○Academic email domains
-
○Institutional organization owner
-
○JOSS paper metadata
-
○Scientific vocabulary similarity
Low similarity (5.8%) to scientific vocabulary
Keywords
Repository
Don't feel pain to use Deformable Convolution
Basic Info
Statistics
- Stars: 341
- Watchers: 2
- Forks: 33
- Open Issues: 1
- Releases: 0
Topics
Metadata Files
README.md
PyTorch-Deformable-Convolution-v2
Don't feel pain to use Deformable Convolution v2(DCNv2)
If you are curious about how to visualize offset(red point), refer to offset_visualization.py
Usage
```python
from dcn import DeformableConv2d
class Model(nn.Module): ... self.conv = DeformableConv2d(inchannels=32, outchannels=32, kernel_size=3, stride=1, padding=1) ...
```
Experiment
You can simply reproduce the results of my experiment on Google Colab.
Refer to experiment.ipynb!
Task
Scaled-MNIST Handwritten Digit Classification
Model
Simple CNN Model including 5 conv layers
```python class MNISTClassifier(nn.Module): def init(self, deformable=False):
super(MNISTClassifier, self).__init__()
self.conv1 = nn.Conv2d(1, 32, kernel_size=3, stride=1, padding=1, bias=True)
self.conv2 = nn.Conv2d(32, 32, kernel_size=3, stride=1, padding=1, bias=True)
self.conv3 = nn.Conv2d(32, 32, kernel_size=3, stride=1, padding=1, bias=True)
conv = nn.Conv2d if deformable==False else DeformableConv2d
self.conv4 = conv(32, 32, kernel_size=3, stride=1, padding=1, bias=True)
self.conv5 = conv(32, 32, kernel_size=3, stride=1, padding=1, bias=True)
self.pool = nn.MaxPool2d(2)
self.gap = nn.AdaptiveAvgPool2d((1, 1))
self.fc = nn.Linear(32, 10)
def forward(self, x):
x = torch.relu(self.conv1(x))
x = self.pool(x) # [14, 14]
x = torch.relu(self.conv2(x))
x = self.pool(x) # [7, 7]
x = torch.relu(self.conv3(x))
x = torch.relu(self.conv4(x))
x = torch.relu(self.conv5(x))
x = self.gap(x)
x = x.flatten(start_dim=1)
x = self.fc(x)
return x
```
Training
- Optimizer: Adam
- Learning Rate: 1e-3
- Learning Rate Scheduler: StepLR(step_size=1, gamma=0.7)
- Batch Size: 64
- Epochs: 14
- Augmentation: NONE
Test
In the paper, authors mentioned that the network's ability to model geometric transformation with DCNv2 is considerably enhanced.
I verified it with scale augmentation.
All images in the test set of MNIST dataset are augmented by scale augmentation(x0.5, x0.6, ..., x1.4, x1.5).
Results
|Model|Top-1 Accuracy(%)| |---|---| |w/o DCNv2|90.03%| |w/ DCNv2|92.90%|
References
To Do Lists
- [ ] Support Onnx Conversion
Owner
- Name: Yonghye Kwon
- Login: developer0hye
- Kind: user
- Location: Seoul, Korea
- Company: MarkAny
- Website: https://www.linkedin.com/in/yonghye-kwon-91641a174/
- Repositories: 15
- Profile: https://github.com/developer0hye
practical
Citation (CITATION.CFF)
cff-version: 1.2.0 message: "If you use this software, please cite it as below." authors: - family-names: "Kwon" given-names: "Yonghye" orcid: "https://orcid.org/my-orcid?orcid=0000-0001-8308-9909" title: "PyTorch-Deformable-Convolution-v2" version: 1.0.0 date-released: 2022-04-20 url: "https://github.com/developer0hye/PyTorch-Deformable-Convolution-v2"
GitHub Events
Total
- Watch event: 22
- Fork event: 3
Last Year
- Watch event: 22
- Fork event: 3