Merge pull request #15 from arnab39/dev

0.1.2

Author: sibasmarak

.github/pull_request_template.md

@@ -0,0 +1,20 @@
+This is a template for making a pull-request. You can remove the text and sections and write your own thing if you wish, just make sure you give enough information about how and why. If you have any issues or difficulties, don't hesitate to open an issue.
+
+
+# Description
+
+The aim is to add this feature ...
+
+# Proposed Changes
+
+I changed the `foo()` function so that ...
+
+
+# Checklist
+
+Here are some things to check before creating the pull request. If you encounter any issues, don't hesitate to ask for help :)
+
+- [ ] I have read the [contributor's guide](https://github.com/arnab39/equiadapt/blob/main/CONTRIBUTING.md).
+- [ ] The base branch of my pull request is the `dev` branch, not the `main` branch.
+- [ ] I ran the [code checks](https://github.com/arnab39/equiadapt/blob/main/CONTRIBUTING.md#implement-your-changes) on the files I added or modified and fixed the errors.
+- [ ] I updated the [changelog](https://github.com/arnab39/equiadapt/blob/main/CHANGELOG.md).

.github/workflows/ci.yml

@@ -58,14 +58,16 @@ jobs:
   test:
     needs: prepare
     strategy:
+      fail-fast: false
       matrix:
-        python:
-        - "3.7"  # oldest Python supported by PSF
-        - "3.10"  # newest Python that is stable
-        platform:
-        - ubuntu-latest
-        # - macos-latest
-        # - windows-latest
+        python: ["3.7", "3.8", "3.9", "3.10", "3.11", "3.12"]
+        platform: [ubuntu-latest, macos-latest, windows-latest]
+        exclude:  # Python < v3.8 does not support Apple Silicon ARM64.
+          - python: "3.7"
+            platform: macos-latest
+        include:  # So run those legacy versions on Intel CPUs.
+          - python: "3.7"
+            platform: macos-13
     runs-on: ${{ matrix.platform }}
     steps:
       - uses: actions/checkout@v3

.pre-commit-config.yaml

@@ -2,7 +2,7 @@ exclude: '^docs/conf.py'
 
 repos:
 - repo: https://github.com/pre-commit/pre-commit-hooks
-  rev: v4.5.0
+  rev: v4.6.0
   hooks:
   - id: trailing-whitespace
   - id: check-added-large-files
@@ -40,7 +40,7 @@ repos:
   - id: isort
 
 - repo: https://github.com/psf/black
-  rev: 24.2.0
+  rev: 24.4.2
   hooks:
   - id: black
     language_version: python3
@@ -66,7 +66,7 @@ repos:
 
 # Check for type errors with mypy:
 - repo: https://github.com/pre-commit/mirrors-mypy
-  rev: 'v1.9.0'
+  rev: 'v1.10.0'
   hooks:
   - id: mypy
     args: [--disallow-untyped-defs, --ignore-missing-imports]

CHANGELOG.md

@@ -5,15 +5,22 @@ All notable changes to this project will be documented in this file.
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
 and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
-## [Unreleased]
+## [0.1.2] - 2024-05-29
 
 ### Added
+- Added canonicalization with optimization approach.
+- Added evaluating transfer learning capabilities of canonicalizer.
+- Added pull request template.
+- Added test for discrete invert canonicalization.
 
 ### Fixed
+- Fixed segmentation evaluation for non-identity canonicalizers.
+- Fixed minor bugs in inverse canonicalization for discrete groups.
 
 ### Changed
-
-### Removed
+- Updated `README.md` with [Improved Canonicalization for Model Agnostic Equivariance](https://arxiv.org/abs/2405.14089) ([EquiVision](https://equivision.github.io/), CVPR 2024 workshop) paper details.
+- Updated `CONTRIBUTING.md` with more information on how to run the code checks.
+- Changed the OS used to test Python 3.7 on GitHub actions (macos-latest -> macos-13).
 
 ## [0.1.1] - 2024-03-15
 

CONTRIBUTING.md

@@ -155,17 +155,29 @@ This can easily be done via [Anaconda] or [Miniconda] and detailed [here](https:
    `git log --graph --decorate --pretty=oneline --abbrev-commit --all`
    to look for recurring communication patterns.
 
+#### Run code checks
 
-5. Please check that your changes don't break any unit tests with:
+Please make sure to see the validation messages from pre-commit and fix any
+eventual issues. This should automatically use [flake8]/[black] to check/fix
+the code style in a way that is compatible with the project.
 
-   ```
-   tox
-   ```
+To run pre-commit manually, you can use:
+
+```
+pre-commit run --all-files
+```
+
+Please also check that your changes don't break any unit tests with:
+
+```
+tox
+```
+
+(after having installed [tox] with `pip install tox` or `pipx`).
 
-   (after having installed [tox] with `pip install tox` or `pipx`).
+You can also use [tox] to run several other pre-configured tasks in the
+repository. Try `tox -av` to see a list of the available checks.
 
-   You can also use [tox] to run several other pre-configured tasks in the
-   repository. Try `tox -av` to see a list of the available checks.
 
 ### Submit your contribution
 

README.md

@@ -121,6 +121,8 @@ You can clone this repository and manually install it with:
 
 ## Setup Conda environment for examples
 
+The recommended way is to manually create an environment and install the dependencies from the `min_conda_env.yaml` file.
+
 To create a conda environment with the necessary packages:
 
 ```
@@ -168,7 +170,7 @@ You can also find [tutorials](https://github.com/arnab39/equiadapt/blob/main/tut
 
 # Related papers and Citations
 
-For more insights on this library refer to our original paper on the idea: [Equivariance with Learned Canonicalization Function (ICML 2023)](https://proceedings.mlr.press/v202/kaba23a.html) and how to extend it to make any existing large pre-trained model equivariant: [Equivariant Adaptation of Large Pretrained Models (NeurIPS 2023)](https://proceedings.neurips.cc/paper_files/paper/2023/hash/9d5856318032ef3630cb580f4e24f823-Abstract-Conference.html).
+For more insights on this library refer to our original paper on the idea: [Equivariance with Learned Canonicalization Function (ICML 2023)](https://proceedings.mlr.press/v202/kaba23a.html) and how to extend it to make any existing large pre-trained model equivariant: [Equivariant Adaptation of Large Pretrained Models (NeurIPS 2023)](https://proceedings.neurips.cc/paper_files/paper/2023/hash/9d5856318032ef3630cb580f4e24f823-Abstract-Conference.html). An improved approach for designing canonicalization network, which allows non-equivariant and expressive models as equivariant networks is presented in [Improved Canonicalization for Model Agnostic Equivariance (CVPR 2024: EquiVision Workshop)](https://arxiv.org/abs/2405.14089).
 
 
 If you find this library or the associated papers useful, please cite the following papers:
@@ -197,6 +199,17 @@ If you find this library or the associated papers useful, please cite the follow
 }
 ```
 
+```
+@inproceedings{
+  panigrahi2024improved,
+  title={Improved Canonicalization for Model Agnostic Equivariance},
+  author={Siba Smarak Panigrahi and Arnab Kumar Mondal},
+  booktitle={CVPR 2024 Workshop on Equivariant Vision: From Theory to Practice},
+  year={2024},
+  url={https://arxiv.org/abs/2405.14089}
+}
+```
+
 # Contact
 
 For questions related to this code, please raise an issue and you can mail us at:
@@ -206,7 +219,7 @@ For questions related to this code, please raise an issue and you can mail us at
 
 # Contributing
 
-You can check out the [contributor's guide](https://github.com/arnab39/equiadapt/blob/main/CHANGELOG.md).
+You can check out the [contributor's guide](https://github.com/arnab39/equiadapt/blob/main/CONTRIBUTING.md).
 
 This project uses `pre-commit`, you can install it before making any
 changes::

equiadapt/images/__init__.py

@@ -22,6 +22,8 @@
     RotationEquivariantConvLift,
     RotoReflectionEquivariantConv,
     RotoReflectionEquivariantConvLift,
+    WideResNet50Network,
+    WideResNet101Network,
     custom_equivariant_networks,
     custom_group_equivariant_layers,
     custom_nonequivariant_networks,
@@ -51,6 +53,8 @@
     "OptimizedGroupEquivariantImageCanonicalization",
     "OptimizedSteerableImageCanonicalization",
     "ResNet18Network",
+    "WideResNet50Network",
+    "WideResNet101Network",
     "RotationEquivariantConv",
     "RotationEquivariantConvLift",
     "RotoReflectionEquivariantConv",

equiadapt/images/canonicalization_networks/__init__.py

@@ -16,6 +16,8 @@
 from equiadapt.images.canonicalization_networks.custom_nonequivariant_networks import (
     ConvNetwork,
     ResNet18Network,
+    WideResNet50Network,
+    WideResNet101Network,
 )
 from equiadapt.images.canonicalization_networks.escnn_networks import (
     ESCNNEquivariantNetwork,
@@ -34,6 +36,8 @@
     "ESCNNWideBasic",
     "ESCNNWideBottleneck",
     "ResNet18Network",
+    "WideResNet101Network",
+    "WideResNet50Network",
     "RotationEquivariantConv",
     "RotationEquivariantConvLift",
     "RotoReflectionEquivariantConv",

equiadapt/images/canonicalization_networks/custom_nonequivariant_networks.py

@@ -110,7 +110,7 @@ def __init__(
             out_vector_size (int, optional): The size of the output vector of the network. Defaults to 128.
         """
         super().__init__()
-        self.resnet18 = torchvision.models.resnet18(weights=None)
+        self.resnet18 = torchvision.models.resnet18(weights="DEFAULT")
         self.resnet18.fc = nn.Sequential(
             nn.Linear(512, out_vector_size),
         )
@@ -128,3 +128,103 @@ def forward(self, x: torch.Tensor) -> torch.Tensor:
             torch.Tensor: The output of the network. It has the shape (batch_size, 1).
         """
         return self.resnet18(x)
+
+
+class WideResNet101Network(nn.Module):
+    """
+    This class represents a neural network based on the WideResNetNetwork architecture.
+
+    The network uses a pre-trained WideResNet model. The final fully connected layer of the WideResNet101 model is replaced with a new fully connected layer.
+
+    Attributes:
+        resnet18 (torchvision.models.ResNet): The ResNet-18 model.
+        out_vector_size (int): The size of the output vector of the network.
+    """
+
+    def __init__(
+        self,
+        in_shape: tuple,
+        out_channels: int,
+        kernel_size: int,
+        num_layers: int = 2,
+        out_vector_size: int = 128,
+    ):
+        """
+        Initializes the ResNet18Network instance.
+
+        Args:
+            in_shape (tuple): The shape of the input data. It should be a tuple of the form (in_channels, height, width).
+            out_channels (int): The number of output channels of the first convolutional layer.
+            kernel_size (int): The size of the kernel of the convolutional layers.
+            num_layers (int, optional): The number of convolutional layers. Defaults to 2.
+            out_vector_size (int, optional): The size of the output vector of the network. Defaults to 128.
+        """
+        super().__init__()
+        self.wideresnet = torchvision.models.wide_resnet101_2(weights="DEFAULT")
+        self.wideresnet.fc = nn.Sequential(
+            nn.Linear(2048, out_vector_size),
+        )
+
+        self.out_vector_size = out_vector_size
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """
+        Performs a forward pass through the network.
+
+        Args:
+            x (torch.Tensor): The input data. It should have the shape (batch_size, in_channels, height, width).
+
+        Returns:
+            torch.Tensor: The output of the network. It has the shape (batch_size, 1).
+        """
+        return self.wideresnet(x)
+
+
+class WideResNet50Network(nn.Module):
+    """
+    This class represents a neural network based on the WideResNetNetwork architecture.
+
+    The network uses a pre-trained WideResNet model. The final fully connected layer of the WideResNet50 model is replaced with a new fully connected layer.
+
+    Attributes:
+        resnet18 (torchvision.models.ResNet): The ResNet-18 model.
+        out_vector_size (int): The size of the output vector of the network.
+    """
+
+    def __init__(
+        self,
+        in_shape: tuple,
+        out_channels: int,
+        kernel_size: int,
+        num_layers: int = 2,
+        out_vector_size: int = 128,
+    ):
+        """
+        Initializes the ResNet18Network instance.
+
+        Args:
+            in_shape (tuple): The shape of the input data. It should be a tuple of the form (in_channels, height, width).
+            out_channels (int): The number of output channels of the first convolutional layer.
+            kernel_size (int): The size of the kernel of the convolutional layers.
+            num_layers (int, optional): The number of convolutional layers. Defaults to 2.
+            out_vector_size (int, optional): The size of the output vector of the network. Defaults to 128.
+        """
+        super().__init__()
+        self.wideresnet = torchvision.models.wide_resnet50_2(weights="DEFAULT")
+        self.wideresnet.fc = nn.Sequential(
+            nn.Linear(2048, out_vector_size),
+        )
+
+        self.out_vector_size = out_vector_size
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """
+        Performs a forward pass through the network.
+
+        Args:
+            x (torch.Tensor): The input data. It should have the shape (batch_size, in_channels, height, width).
+
+        Returns:
+            torch.Tensor: The output of the network. It has the shape (batch_size, 1).
+        """
+        return self.wideresnet(x)

examples/images/classification/configs/canonicalization/opt_group_equivariant.yaml

@@ -1,5 +1,5 @@
 canonicalization_type: opt_group_equivariant
-network_type: cnn # Options for canonization method 1) cnn 2) wideresnet
+network_type: cnn # Options for canonization method 1) cnn 2) non_equivariant_wrn_50 3) non_equivariant_wrn_101 4) non_equivariant_resnet18
 network_hyperparams:
   kernel_size: 7 # Kernel size for the canonization network
   out_channels: 16 # Number of output channels for the canonization network