Instance Selection for GANs
Authors: Terrance DeVries, Michal Drozdzal, Graham W. Taylor
NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this work we propose a novel approach to improve sample quality: altering the training dataset via instance selection before model training has taken place. By refining the empirical data distribution before training, we redirect model capacity towards high-density regions, which ultimately improves sample fidelity, lowers model capacity requirements, and significantly reduces training time. ... In our experiments, we evaluate a variety of image embeddings and scoring functions... We demonstrate the model capacity savings of instance selection by achieving state-of-the-art performance (in terms of FID) on 64 × 64 resolution Image Net generation using a Self-Attention GAN with 1/2 the amount of trainable parameters of the current best model. We demonstrate training time savings by training a 128 × 128 resolution Big GAN on Image Net in 1/4 the time of the baseline, while also achieving superior performance across all image fidelity metrics. |
| Researcher Affiliation | Collaboration | Terrance De Vries University of Guelph Vector Institute Michal Drozdzal Facebook AI Research Graham W. Taylor University of Guelph Vector Institute |
| Pseudocode | No | The paper provides mathematical equations describing the scoring functions and instance selection process, but does not present a structured pseudocode or algorithm block. |
| Open Source Code | Yes | Code is available at https://github.com/uoguelph-mlrg/instance_selection_for_gans. |
| Open Datasets | Yes | we use the Image Net2 dataset [7]... ImageNet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248 255. Ieee, 2009. |
| Dataset Splits | Yes | When calculating FID we follow Brock et al. [2] in using all images in the training set to estimate the reference distribution, and sampling 50 k images to make up the generated distribution. For P&R and D&C we use an Inceptionv3 embedding. N and M are set to 10 k samples for both the reference and generated distributions, and K is set equal to 5 as recommended by Naeem et al. [19]. ... Each model is trained on a different 50% subset of Image Net, as chosen by instance selection... |
| Hardware Specification | Yes | Both models are trained on 8 NVIDIA V100 GPUs with 16GB of RAM... train a Big GAN on Image Net at 256 × 256 resolution using 4 V100s with 32GB of RAM each. |
| Software Dependencies | No | The paper mentions PyTorch and TensorFlow (for specific models/metrics) but does not provide version numbers for these or other software dependencies critical for replication. |
| Experiment Setup | Yes | We use the default settings for SAGAN, except that we use a batch size of 128 instead of 256, apply the self-attention module at 32 × 32 resolution instead of 64 × 64, and reduce the number of channels in each layer by half... All models are trained for 200k iterations. ... we extend training until 500k iterations... Our baseline model uses the default hyperparameters from Big GAN [2], with the exception that we reduce the channel multiplier from 96 to 64 (i.e. half of the capacity) and only use a single discriminator update instead of two for faster training. ... we reduce the batch size from Big GAN’s default of 2048 to 256 for the instance selection model. ... we use two discriminator update steps per generator update for this experiment. |