Official Pytorch code of Instance-Conditioned GAN by Arantxa Casanova, Marlene Careil, Jakob Verbeek, Michał Drożdżal, Adriana Romero-Soriano.
Generate images with IC-GAN in a Colab Notebook
We provide a Google Colab notebook to generate images with IC-GAN and its class-conditional counter part.
The figure below depicts two instances, unseen during training and downloaded from Creative Commons search, and the generated images with IC-GAN and class-conditional IC-GAN when conditioning on the class “castle”:
Additionally, and inspired by this Colab, we provide the funcionality in the same Colab notebook to guide generations with text captions, using the CLIP model. As an example, the following Figure shows three instance conditionings and a text caption (top), followed by the resulting generated images with IC-GAN (bottom), when optimizing the noise vector following CLIP’s gradient for 100 iterations.
A conda environment can be created from environment.yaml by entering the command: conda env create -f environment.yml, that contains the aforemention version of Pytorch and other required packages.
data_utils: A common folder to obtain and format the data needed to train and test IC-GAN, agnostic of the specific backbone.
inference: Scripts to test the models both qualitatively and quantitatively.
BigGAN_PyTorch: It provides the training, evaluation and sampling scripts for IC-GAN with a BigGAN backbone. The code base comes from Pytorch BigGAN repository, made available under the MIT License. It has been modified to add additional utilities and it enables IC-GAN training on top of it.
stylegan2_ada_pytorch: It provides the training, evaluation and sampling scripts for IC-GAN with a StyleGAN2 backbone. The code base comes from StyleGAN2 Pytorch, made available under the Nvidia Source Code License. It has been modified to add additional utilities and it enables IC-GAN training on top of it.
(Python script) Generate images with IC-GAN
Alternatively, we can generate images with IC-GAN models directly from a python script, by following the next steps:
model can be chosen from ["icgan", "cc_icgan"] to use the IC-GAN or the class-conditional IC-GAN model respectively.
backbone can be chosen from ["biggan", "stylegan2"].
res indicates the resolution at which the model has been trained. For ImageNet, choose one in [64, 128, 256], and for COCO-Stuff, one in [128, 256].
This script results in a .PNG file where several generated images are shown, given an instance feature (each row), and a sampled noise vector (each grid position).
Additional and optional parameters:
index: (None by default), is an integer from 0 to 999 that choses a specific instance feature vector out of the 1000 instances that have been selected with k-means on the ImageNet dataset and stored in pretrained_models_path/stored_instances.
swap_target: (None by default) is an integer from 0 to 999 indicating an ImageNet class label. This label will be used to condition the class-conditional IC-GAN, regardless of which instance features are being used.
which_dataset: (ImageNet by default) can be chosen from ["imagenet", "coco"] to indicate which dataset (training split) to sample the instances from.
trained_dataset: (ImageNet by default) can be chosen from ["imagenet", "coco"] to indicate the dataset in which the IC-GAN model has been trained on.
num_imgs_gen: (5 by default), it changes the number of noise vectors to sample per conditioning. Increasing this number results in a bigger .PNG file to save and load.
num_conditionings_gen: (5 by default), it changes the number of conditionings to sample. Increasing this number results in a bigger .PNG file to save and load.
z_var: (1.0 by default) controls the truncation factor for the generation.
Optionally, the script can be run with the following additional options --visualize_instance_images --dataset_path [dataset_path] to visualize the ground-truth images corresponding to the conditioning instance features, given a path to the dataset’s ground-truth images dataset_path. Ground-truth instances will be plotted as the leftmost image for each row.
Download SwAV feature extractor weights from here .
Replace the paths in data_utils/prepare_data.sh: out_path by the path where hdf5 files will be stored, path_imnet by the path where ImageNet dataset is downloaded, and path_swav by the path where SwAV weights are stored.
Execute ./data_utils/prepare_data.sh imagenet [resolution], where [resolution] can be an integer in {64,128,256}. This script will create several hdf5 files:
ILSVRC[resolution]_xy.hdf5 and ILSVRC[resolution]_val_xy.hdf5, where images and labels are stored for the training and validation set respectively.
ILSVRC[resolution]_feats_[feature_extractor]_resnet50.hdf5 that contains the instance features for each image.
ILSVRC[resolution]_feats_[feature_extractor]_resnet50_nn_k[k_nn].hdf5 that contains the list of [k_nn] neighbors for each of the instance features.
ImageNet-LT
Download ImageNet dataset from here . Following ImageNet-LT , the file ImageNet_LT_train.txt can be downloaded from this link and later stored in the folder ./BigGAN_PyTorch/imagenet_lt.
Replace the paths in data_utils/prepare_data.sh: out_path by the path where hdf5 files will be stored, path_imnet by the path where ImageNet dataset is downloaded, and path_classifier_lt by the path where the pre-trained ResNet50 weights are stored.
Execute ./data_utils/prepare_data.sh imagenet_lt [resolution], where [resolution] can be an integer in {64,128,256}. This script will create several hdf5 files:
ILSVRC[resolution]longtail_xy.hdf5, where images and labels are stored for the training and validation set respectively.
ILSVRC[resolution]longtail_feats_[feature_extractor]_resnet50.hdf5 that contains the instance features for each image.
ILSVRC[resolution]longtail_feats_[feature_extractor]_resnet50_nn_k[k_nn].hdf5 that contains the list of [k_nn] neighbors for each of the instance features.
Download SwAV feature extractor weights from here .
Replace the paths in data_utils/prepare_data.sh: out_path by the path where hdf5 files will be stored, path_imnet by the path where ImageNet dataset is downloaded, and path_swav by the path where SwAV weights are stored.
Execute ./data_utils/prepare_data.sh coco [resolution], where [resolution] can be an integer in {128,256}. This script will create several hdf5 files:
COCO[resolution]_xy.hdf5 and COCO[resolution]_val_test_xy.hdf5, where images and labels are stored for the training and evaluation set respectively.
COCO[resolution]_feats_[feature_extractor]_resnet50.hdf5 that contains the instance features for each image.
COCO[resolution]_feats_[feature_extractor]_resnet50_nn_k[k_nn].hdf5 that contains the list of [k_nn] neighbors for each of the instance features.
Other datasets
Download the corresponding dataset and store in a folder dataset_path.
Download SwAV feature extractor weights from here .
Replace the paths in data_utils/prepare_data.sh: out_path by the path where hdf5 files will be stored and path_swav by the path where SwAV weights are stored.
Execute ./data_utils/prepare_data.sh [dataset_name] [resolution] [dataset_path], where [dataset_name] will be the dataset name, [resolution] can be an integer, for example 128 or 256, and dataset_path contains the dataset images. This script will create several hdf5 files:
[dataset_name][resolution]_xy.hdf5, where images and labels are stored for the training set.
[dataset_name][resolution]_feats_[feature_extractor]_resnet50.hdf5 that contains the instance features for each image.
[dataset_name][resolution]_feats_[feature_extractor]_resnet50_nn_k[k_nn].hdf5 that contains the list of k_nn neighbors for each of the instance features.
How to subsample an instance feature dataset with k-means
To downsample the instance feature vector dataset, after we have prepared the data, we can use the k-means algorithm: python data_utils/store_kmeans_indexes.py --resolution [resolution] --which_dataset [dataset_name] --data_root [data_path]
Adding --gpu allows the faiss library to compute k-means leveraging GPUs, resulting in faster execution.
Adding the parameter --feature_extractor [feature_extractor] chooses which feature extractor to use, with feature_extractor in ['selfsupervised', 'classification'] , if we are using swAV as feature extactor or the ResNet pretrained on the classification task on ImageNet, respectively.
The number of k-means clusters can be set with --kmeans_subsampled [centers], where centers is an integer.
How to train the models
BigGAN or StyleGAN2 backbone
Training parameters are stored in JSON files in [backbone_folder]/config_files/[dataset]/*.json, where [backbone_folder] is either BigGAN_Pytorch or stylegan2_ada_pytorch and [dataset] can either be ImageNet, ImageNet-LT or COCO_Stuff.
data_root path where the data has been prepared and stored, following the previous section (Data preparation).
base_root path where to store the model weights and logs.
Note that one can create other JSON files to modify the training parameters.
Other backbones
To be able to run IC-GAN with other backbones, we provide some orientative steps:
Place the new backbone code in a new folder under ic_gan (ic_gan/new_backbone).
Modify the relevant piece of code in the GAN architecture to allow instance features as conditionings (for both generator and discriminator).
Create a trainer.py file with the training loop to train an IC-GAN with the new backbone. The data_utils folder provides the tools to prepare the dataset, load the data and conditioning sampling to train an IC-GAN. The IC-GAN with BigGAN backbone trainer.py file can be used as an inspiration.
How to test the models
To obtain the FID and IS metrics on ImageNet and ImageNet-LT:
For stratified FIDs in the ImageNet-LT dataset, the following parameters can be added --which_dataset 'imagenet_lt' --split 'val' --strat_name [stratified_split], where stratified_split can be in [few,low, many].
(Only needed once) Pre-compute reference moments with tensorflow code:
(Using this repository) FID can be computed using the pre-computed statistics obtained in 2) and the pre-computed ground-truth statistics obtain in 3). For example, to compute the FID with reference ImageNet validation set: python TTUR/fid.py [base_root]/[exp_name]/TF_pool_.npz [data_root]/imagenet_val_res[res]_tf_inception_moments_ground_truth.npz
Using this repository, compute FID on the two folders of ground-truth and generated images.
where:
dataset: option to select the dataset in `[‘imagenet’, ‘imagenet_lt’, ‘coco’]
exp_name: name of the experiment folder.
data_root: path where the data has been prepared and stored, following the previous section “Data preparation”.
base_root: path where to find the model (for example, where the pretrained models have been downloaded).
num_imgs: needs to be set to 50000 for ImageNet and ImageNet-LT (with validation set as reference) and set to 11500 for ImageNet-LT (with training set as reference). For COCO-Stuff, set to 75777, 2050, 675, 1375 if using the training, evaluation, evaluation seen or evaluation unseen set as reference.
ref_set: set to 'val' for ImageNet, ImageNet-LT (and COCO) to obtain metrics with the validation (evaluation) set as reference, or set to 'train' for ImageNet-LT or COCO to obtain metrics with the training set as reference.
kmeans_centers: set to 1000 for ImageNet and to -1 for ImageNet-LT.
backbone: model backbone architecture in ['biggan','stylegan2'].
res: integer indicating the resolution of the images (64,128,256).
gt_coco_images: folder to store the ground-truth JPEG images of that specific split.
filter_hd: only valid for ref_set=val. If -1, use the entire evaluation set; if 0, use only conditionings and their ground-truth images with seen class combinations during training (eval seen); if 1, use only conditionings and their ground-truth images with unseen class combinations during training (eval unseen).
Utilities for GAN backbones
We change and provide extra utilities to facilitate the training, for both BigGAN and StyleGAN2 base repositories.
BigGAN change log
The following changes were made:
BigGAN architecture:
In train_fns.py: option to either have the optimizers inside the generator and discriminator class, or directly in the G_D wrapper module. Additionally, added an option to augment both generated and real images with augmentations from DiffAugment.
In BigGAN.py: added a function get_condition_embeddings to handle the conditioning separately.
Small modifications to layers.py to adapt the batchnorm function calls to the pytorch 1.8 version.
Training utilities:
Added trainer.py file (replacing train.py):
Training now allows the usage of DDP for faster single-node and multi-node training.
Training is performed by epochs instead of by iterations.
Option to stop the training by using early stopping or when experiments diverge.
In utils.py:
Replaced MultiEpochSampler for CheckpointedSampler to allow experiments to be resumable when using epochs and fixing a bug where MultiEpochSampler would require a long time to fetch data permutations when the number of epochs increased.
ImageNet-LT: Added option to use different class distributions when sampling a class label for the generator.
ImageNet-LT: Added class balancing (uniform and temperature annealed).
In calculate_inception_moments.py: added option to obtain moments for ImageNet-LT dataset, as well as stratified moments for many, medium and few-shot classes (stratified FID computation).
In datasets.py, added option to load ImageNet-LT dataset.
Added ImageNet-LT.txt files with image indexes for training and validation split.
In utils.py:
Separate functions to obtain the data from hdf5 files (get_dataset_hdf5) or from directory (get_dataset_images), as well as a function to obtain only the data loader (get_dataloader).
Added the function sample_conditionings to handle possible different conditionings to train G with.
Experiment utilities:
Added JSON files to launch experiments with the proposed hyper-parameter configuration.
Script to launch experiments with either the submitit tool or locally in the same machine (run.py).
StyleGAN2 change log
Multi-node DistributedDataParallel training.
Added early stopping based on the training FID metric.
Automatic checkpointing when jobs are automatically rescheduled on a cluster.
Option to load dataset from hdf5 file.
Replaced the usage of Click python package by an `ArgumentParser`.
The majority of IC-GAN is licensed under CC-BY-NC, however portions of the project are available under separate license terms: BigGAN and PRDC are licensed under the MIT license; COCO-Stuff loader is licensed under Apache License 2.0; DiffAugment is licensed under BSD 2-Clause Simplified license; StyleGAN2 is licensed under a NVIDIA license, available here: https://github.com/NVlabs/stylegan2-ada-pytorch/blob/main/LICENSE.txt. In the Colab notebook, CLIP and pytorch-pretrained-BigGAN code is used, both licensed under the MIT license.
Disclaimers
THE DIFFAUGMENT SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS “AS IS” AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
THE CLIP 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.
THE PYTORCH-PRETRAINED-BIGGAN 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.
Cite the paper
If this repository, the paper or any of its content is useful for your research, please cite:
@misc{casanova2021instanceconditioned,
title={Instance-Conditioned GAN},
author={Arantxa Casanova and Marlène Careil and Jakob Verbeek and Michal Drozdzal and Adriana Romero-Soriano},
year={2021},
eprint={2109.05070},
archivePrefix={arXiv},
primaryClass={cs.CV}
}
John was the first writer to have joined pythonawesome.com. He has since then inculcated very effective writing and reviewing culture at pythonawesome which rivals have found impossible to imitate.
Previous Post
Moviepy-cli: Command line interface for MoviePy
Next Post
A production-ready pipeline for text mining and subject indexing
PIRDS - Prescription Image Recognition and Digitalizing System is a OCR make with Tensorflow that digitalises images of Prescription of Handwritten Texts by Doctors
