release

.gitignore

@@ -0,0 +1,134 @@
+# Byte-compiled / optimized / DLL files
+__pycache__/
+*.py[cod]
+*$py.class
+
+# C extensions
+*.so
+
+# Distribution / packaging
+.Python
+build/
+develop-eggs/
+dist/
+downloads/
+eggs/
+.eggs/
+lib/
+lib64/
+parts/
+sdist/
+var/
+wheels/
+pip-wheel-metadata/
+share/python-wheels/
+*.egg-info/
+.installed.cfg
+*.egg
+MANIFEST
+
+# PyInstaller
+#  Usually these files are written by a python script from a template
+#  before PyInstaller builds the exe, so as to inject date/other infos into it.
+*.manifest
+*.spec
+
+# Installer logs
+pip-log.txt
+pip-delete-this-directory.txt
+
+# Unit test / coverage reports
+htmlcov/
+.tox/
+.nox/
+.coverage
+.coverage.*
+.cache
+nosetests.xml
+coverage.xml
+*.cover
+*.py,cover
+.hypothesis/
+.pytest_cache/
+
+# Translations
+*.mo
+*.pot
+
+# Django stuff:
+*.log
+local_settings.py
+db.sqlite3
+db.sqlite3-journal
+
+# Flask stuff:
+instance/
+.webassets-cache
+
+# Scrapy stuff:
+.scrapy
+
+# Sphinx documentation
+docs/_build/
+
+# PyBuilder
+target/
+
+# Jupyter Notebook
+.ipynb_checkpoints
+
+# IPython
+profile_default/
+ipython_config.py
+
+# pyenv
+.python-version
+
+# pipenv
+#   According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control.
+#   However, in case of collaboration, if having platform-specific dependencies or dependencies
+#   having no cross-platform support, pipenv may install dependencies that don't work, or not
+#   install all needed dependencies.
+#Pipfile.lock
+
+# PEP 582; used by e.g. github.com/David-OConnor/pyflow
+__pypackages__/
+
+# Celery stuff
+celerybeat-schedule
+celerybeat.pid
+
+# SageMath parsed files
+*.sage.py
+
+# Environments
+.env
+.venv
+env/
+venv/
+ENV/
+env.bak/
+venv.bak/
+
+# Spyder project settings
+.spyderproject
+.spyproject
+
+# Rope project settings
+.ropeproject
+
+# mkdocs documentation
+/site
+
+# mypy
+.mypy_cache/
+.dmypy.json
+dmypy.json
+
+# Pyre type checker
+.pyre/
+
+*.pt
+!precalculated_basis/*.pt
+*.pth
+.DS_Store

LICENSE

@@ -0,0 +1,24 @@
+MIT License
+
+Copyright (c) 2020-2021 Ivan Sosnovik, Artem Moskalev
+Copyright (c) 2020 Ivan Sosnovik, Michał Szmaja
+Copyright (c) 2019 Erik J Bekkers
+Copyright (c) 2019 xternalz
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

README.md

@@ -0,0 +1,71 @@
+![](./src/logo.jpg)
+
+
+This is the official implementation of
+
+[DISCO: accurate Discrete Scale Convolutions](https://arxiv.org/abs/2106.02733)<br>
+**Ivan Sosnovik, Artem Moskalev, and Arnold Smeulders**<br>
+BMVC 2021 (**Oral**)
+
+and 
+
+[How to Transform Kernels for Scale-Convolutions](https://openreview.net/forum?id=rTpTF_-fOwm)<br>
+**Ivan Sosnovik, Artem Moskalev, and Arnold Smeulders**<br>
+ICCV VIPriors 2021
+
+
+*Scale is often seen as a given, disturbing factor in many vision tasks. When doing so it is one of the factors why we need more data during learning. In recent work scale equivariance was added to convolutional neural networks. We aim for accurate scale-equivariant convolutional neural networks (SE-CNNs). Current SE-CNNs rely on weight sharing and filter rescaling, the latter of which is accurate for integer scales only. To reach accurate scale equivariance, we derive general constraints under which scale-convolution remains equivariant to discrete rescaling. We find the exact solution for all cases where it exists, and compute the approximation for the rest.*
+
+## Scale-Equivariant CNNs
+![](./src/scheme.gif)
+
+DISCO is a new class of scale-equivariant convolutional neural networks (SE-CNNs).  To learn more about the backbone SE-CNNs we use, check [SESN](https://github.com/ISosnovik/sesn/). To use our models you first need to calculate the DISCO basis. You can do it by simply running 
+```bash
+python calculate_disco_basis.py \
+    --basis_size 5 \ # actual size of the filters
+    --basis_effective_size 3 \ # the size of the filter for the conventional CNN
+    --basis_scales 1.0 1.41 2.0 \ # the scales hyperparameter for convolutional layers
+    --basis_save_dir precalculated_basis \
+    --cuda \
+
+```
+We provide the bases we have used for our experiments in the `./precalculated_basis` folder.
+
+## Experiments
+To reproduce the experiments on Scale-MNIST you first need to generate the datasets. Follow the instructions provided in [SESN](https://github.com/ISosnovik/sesn/). Once the datasets are generated, you can simply run 
+```
+DATASET_SRC=path/to/datasets bash exp_mnist.sh
+```
+and for the STL-10 experiments, run
+```
+DATASET_SRC=path/to/datasets bash exp_stl.sh
+```
+The experiments on OTB-13 are coming soon
+
+
+
+## Acknowledgements
+The Robert Bosch GmbH is acknowledged for financial support.
+
+## BibTeX
+If you found this work useful in your research, please consider citing
+```
+@article{sosnovik2021disco,
+  title={DISCO: accurate Discrete Scale Convolutions},
+  author={Sosnovik, Ivan and Moskalev, Artem and Smeulders, Arnold},
+  journal={arXiv preprint arXiv:2106.02733},
+  year={2021}
+}
+```
+
+and 
+```
+@InProceedings{Sosnovik_2021_ICCV,
+    author    = {Sosnovik, Ivan and Moskalev, Artem and Smeulders, Arnold},
+    title     = {How To Transform Kernels for Scale-Convolutions},
+    booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops},
+    month     = {October},
+    year      = {2021},
+    pages     = {1092-1097}
+}
+```

calculate_disco_basis.py

@@ -0,0 +1,143 @@
+'''
+This file is a part of the official implementation of
+1) "DISCO: accurate Discrete Scale Convolutions"
+    by Ivan Sosnovik, Artem Moskalev, Arnold Smeulders, BMVC 2021
+    arxiv: https://arxiv.org/abs/2106.02733
+
+2) "How to Transform Kernels for Scale-Convolutions"
+    by Ivan Sosnovik, Artem Moskalev, Arnold Smeulders, ICCV VIPriors 2021
+    pdf: https://openaccess.thecvf.com/content/ICCV2021W/VIPriors/papers/Sosnovik_How_To_Transform_Kernels_for_Scale-Convolutions_ICCVW_2021_paper.pdf
+
+---------------------------------------------------------------------------
+
+MIT License. Copyright (c) 2021 Ivan Sosnovik, Artem Moskalev
+'''
+
+import os
+import time
+from argparse import ArgumentParser
+
+import torch
+import torch.optim as optim
+import torch.backends.cudnn as cudnn
+
+from models.basis import ApproximateProxyBasis
+from models.basis.disco import get_basis_filename
+from utils import loaders
+from utils.train_utils import train_equi_loss
+from utils.model_utils import get_num_parameters
+
+
+#########################################
+# arguments
+#########################################
+parser = ArgumentParser()
+parser.add_argument('--batch_size', type=int, default=64)
+parser.add_argument('--epochs', type=int, default=40)
+
+parser.add_argument('--lr', type=float, default=0.001)
+parser.add_argument('--lr_steps', type=int, nargs='+', default=[20, 30])
+parser.add_argument('--lr_gamma', type=float, default=0.1)
+
+parser.add_argument('--cuda', action='store_true', default=False)
+
+# basis hyperparameters
+parser.add_argument('--basis_size', type=int, default=7)
+parser.add_argument('--basis_effective_size', type=int, default=3)
+parser.add_argument('--basis_scales', type=float, nargs='+', default=[1.0])
+parser.add_argument('--basis_save_dir', type=str, required=True)
+
+
+args = parser.parse_args()
+
+print("Args:")
+for k, v in vars(args).items():
+    print("  {}={}".format(k, v))
+
+print(flush=True)
+
+
+#########################################
+# Data
+#########################################
+loader = loaders.random_loader(args.batch_size)
+
+
+print('Dataset:')
+print(loader.dataset)
+
+
+#########################################
+# Model
+#########################################
+basis = ApproximateProxyBasis(size=args.basis_size, scales=args.basis_scales,
+                              effective_size=args.basis_effective_size)
+
+print('\nBasis:')
+print(basis)
+print()
+
+use_cuda = args.cuda and torch.cuda.is_available()
+device = torch.device('cuda' if use_cuda else 'cpu')
+print('Device: {}'.format(device))
+
+if use_cuda:
+    cudnn.enabled = True
+    cudnn.benchmark = True
+    print('CUDNN is enabled. CUDNN benchmark is enabled')
+    basis.cuda()
+
+print(flush=True)
+
+#########################################
+# optimizer
+#########################################
+parameters = filter(lambda x: x.requires_grad, basis.parameters())
+optimizer = optim.Adam(parameters, lr=args.lr)
+print(optimizer)
+lr_scheduler = optim.lr_scheduler.MultiStepLR(optimizer, args.lr_steps, args.lr_gamma)
+
+
+#########################################
+# Paths
+#########################################
+
+
+save_basis_postfix = get_basis_filename(size=args.basis_size,
+                                        effective_size=args.basis_effective_size,
+                                        scales=args.basis_scales)
+save_basis_path = os.path.join(args.basis_save_dir, save_basis_postfix)
+print('Basis path: ', save_basis_path)
+print()
+
+if not os.path.isdir(args.basis_save_dir):
+    os.makedirs(args.basis_save_dir)
+
+#########################################
+# Training
+#########################################
+
+print('\nTraining\n' + '-' * 30)
+start_time = time.time()
+best_loss = float('inf')
+
+for epoch in range(args.epochs):
+    loss = train_equi_loss(basis, optimizer, loader, device)
+    print('Epoch {:3d}/{:3d}| Loss: {:.2e}'.format(epoch + 1, args.epochs, loss), flush=True)
+    if loss < best_loss:
+        best_loss = loss
+
+        with torch.no_grad():
+            torch.save(basis.get_basis().cpu(), save_basis_path)
+
+    lr_scheduler.step()
+
+print('-' * 30)
+print('Training is finished')
+print('Best Loss: {:.2e}'.format(best_loss), flush=True)
+end_time = time.time()
+elapsed_time = end_time - start_time
+time_per_epoch = elapsed_time / args.epochs
+
+print('Total Time Elapsed: {:.2f}'.format(elapsed_time))
+print('Time per Epoch: {:.2f}'.format(time_per_epoch))