[dev => main] Release 0.1.2 (#8)

PRs:
- #4 
- #6 
- #7 

Features:
- #3 

BugFix:
- #5

README.md

@@ -1,6 +1,6 @@
-<h3> <p align="center"> 🎉🎉 Our paper has been accepted at ECCV 2024! Stay tuned for more updates here !! 🎉🎉</h3>
-
 <div align="center">
+<h3> 🎉🎉 Our paper has been accepted at ECCV 2024! Stay tuned for more updates !! 🎉🎉 </h3>
+
 <h2><a href="https://arxiv.org/abs/2403.19588">DenseNets Reloaded: Paradigm Shift Beyond ResNets and ViTs</a></h2>
 
 [Donghyun Kim](https://scholar.google.co.kr/citations?hl=en&user=EBC8BMAAAAAJ)<sup>1*</sup>, [Byeongho Heo](https://sites.google.com/view/byeongho-heo/home)<sup>2</sup>, [Dongyoon Han](https://dongyoonhan.github.io/)<sup>2*</sup>
@@ -16,11 +16,10 @@
 <a href="https://huggingface.co/naver-ai" alt="Huggingface">
     <img src="https://img.shields.io/badge/huggingface-NAVERAILab-F58336" /></a>
 </p>
-
-</p>
 
 <p align="center">
-  <img src="./resources/images/rdnet_reloaded.gif" alt="Densenet Reloaded">
+  <img src="./resources/images/rdnet_reloaded.gif" alt="Densenet Reloaded" width="46.5%" height="100%">
+  <img src="./resources/images/densenet_becomes_rdnet.gif" alt="Densenet becomes RDNet" width="51%" height="100%">
 </p>
 
 We revitalize **Densely Connected Convolutional Networks (DenseNets)** and reveal their untapped potential to challenge the prevalent dominance of ResNet-style architectures. Our research indicates that DenseNets were previously underestimated, primarily due to conventional design choices and training methods that underexploited their full capabilities.
@@ -45,17 +44,27 @@ Our work aims to reignite interest in DenseNets by demonstrating their renewed r
 
 *We believe that various architectural designs that have been popular recently would be combined with dense connections successfully.*
 
+## Easy to use
+Install rdnet package with `pip install rdnet` !
+
+```python
+import timm
+import rdnet  # this will register the RDNet models to timm
+
+model = timm.create_model('rdnet_large', pretrained=True)
+```
+
+For detailed usage, please refer to the [huggingface model card](https://huggingface.co/naver-ai/rdnet_tiny.nv_in1k).
 
 ## Updates
+- **(2024.07.24)**: Pip installable pacakge added.
 - **(2024.04.19)**: Initial release of the repository.
 - **(2024.03.28)**: Paper is available on [arXiv](https://arxiv.org/abs/2403.19588).
 
 ## Coming Soon
 - [ ] More ImageNet-22k Pretrained Models.
 - [ ] More ImageNet-1k fine-tuned models.
 - [x] Cascade Mask R-CNN with RDNet.
-- [ ] DETR with RDNet.
-- [ ] Self-Supervised Learning with RDNet.
 - [ ] Transfer Learning with RDNet (with cifar10, cifar100, stanford cars, ...).
 
 ## RDNet for Image Classification

detection/backbone/rdnet.py

@@ -116,7 +116,10 @@ def __init__(
         self.gamma = nn.Parameter(ls_init_value * torch.ones(growth_rate)) if ls_init_value > 0 else None
         growth_rate = int(growth_rate)
         inter_chs = int(num_input_features * bottleneck_width_ratio / 8) * 8
-        self.drop_path = DropPath(drop_path_rate)
+
+        if self.drop_path_rate > 0:
+            self.drop_path = DropPath(drop_path_rate)
+
         self.layers = eval(block_type)(
             in_chs=num_input_features,
             inter_chs=inter_chs,
@@ -130,6 +133,9 @@ def forward(self, x):
 
         if self.gamma is not None:
             x = x.mul(self.gamma.reshape(1, -1, 1, 1))
+
+        if self.drop_path_rate > 0 and self.training:
+            x = self.drop_path(x)
         return x
 
 

pyproject.toml

@@ -0,0 +1,28 @@
+[build-system]
+requires = [
+    "setuptools>=42",
+    "wheel"
+]
+build-backend = "setuptools.build_meta"
+
+[project]
+name = "rdnet"
+authors = [
+    {name = "Donghyun Kim", email = "[email protected]"},
+]
+description = "RDNet (Densenets Reloaded: Paradigm shift beyond ResNets and ViTs)"
+readme = "README.md"
+requires-python = ">=3.8"
+keywords = ["backbone", "classification", "densenet", "convolutional-neural-networks", "revisit", "dense-connections", "rdnet"]
+license = {file = "LICENSE"}
+classifiers = [
+    "Programming Language :: Python :: 3",
+]
+dependencies = [
+    "torch>=1.9.1",
+    "timm~=0.9"
+]
+version = "0.1.2"
+
+[tool.setuptools.packages.find]
+include = ["rdnet", "rdnet.*"]

rdnet/__init__.py

@@ -0,0 +1 @@
+from .rdnet import RDNet

rdnet.py → rdnet/rdnet.py

@@ -116,7 +116,10 @@ def __init__(
         self.gamma = nn.Parameter(ls_init_value * torch.ones(growth_rate)) if ls_init_value > 0 else None
         growth_rate = int(growth_rate)
         inter_chs = int(num_input_features * bottleneck_width_ratio / 8) * 8
-        self.drop_path = DropPath(drop_path_rate)
+
+        if self.drop_path_rate > 0:
+            self.drop_path = DropPath(drop_path_rate)
+
         self.layers = eval(block_type)(
             in_chs=num_input_features,
             inter_chs=inter_chs,
@@ -130,6 +133,9 @@ def forward(self, x):
 
         if self.gamma is not None:
             x = x.mul(self.gamma.reshape(1, -1, 1, 1))
+
+        if self.drop_path_rate > 0 and self.training:
+            x = self.drop_path(x)
         return x
 
 

segmentation/backbone/rdnet.py

@@ -116,7 +116,10 @@ def __init__(
         self.gamma = nn.Parameter(ls_init_value * torch.ones(growth_rate)) if ls_init_value > 0 else None
         growth_rate = int(growth_rate)
         inter_chs = int(num_input_features * bottleneck_width_ratio / 8) * 8
-        self.drop_path = DropPath(drop_path_rate)
+
+        if self.drop_path_rate > 0:
+            self.drop_path = DropPath(drop_path_rate)
+
         self.layers = eval(block_type)(
             in_chs=num_input_features,
             inter_chs=inter_chs,
@@ -130,8 +133,10 @@ def forward(self, x):
 
         if self.gamma is not None:
             x = x.mul(self.gamma.reshape(1, -1, 1, 1))
-        return x
 
+        if self.drop_path_rate > 0 and self.training:
+            x = self.drop_path(x)
+        return x
 
 class DenseStage(nn.Sequential):
     def __init__(self, num_block, num_input_features, drop_path_rates, growth_rate, **kwargs):