Self-Supervised Generative Adversarial Compression

Authors: Chong Yu, Jeff Pool

NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this paper, we show that a standard model compression technique, weight pruning and knowledge distillation, cannot be applied to GANs using existing methods. We then develop a self-supervised compression technique which uses the trained discriminator to supervise the training of a compressed generator. We show that this framework has compelling performance to high degrees of sparsity, can be easily applied to new tasks and models, and enables meaningful comparisons between different compression granularities.
Researcher Affiliation Industry Chong Yu NVIDIA chongy@nvidia.com Jeff Pool NVIDIA jpool@nvidia.com
Pseudocode No The paper describes the methods textually but does not include any explicit pseudocode or algorithm blocks.
Open Source Code No The paper links to baseline repositories (e.g., StarGAN, DCGAN, Pix2Pix/CycleGAN) but does not provide a direct link or explicit statement about the availability of the source code for the methodology described in this paper.
Open Datasets Yes It uses the Celeb Faces Attributes (Celeb A) [33] as the dataset.
Dataset Splits No The paper mentions 'validation datasets' and reports metrics for them (Table 3), but it does not specify the explicit split percentages or sample counts for a validation set.
Hardware Specification Yes We use PyTorch [37], implement the pruning and training schedules with Distiller [32], and train and generate results with V100 GPU [38] to match public baselines.
Software Dependencies No The paper mentions PyTorch and Distiller, but does not provide specific version numbers for these software components.
Experiment Setup Yes Following AGP [12], we gradually increase the sparsity from 5% at the beginning to our target of 50% halfway through the self-supervised training process, and we set the loss adjustment parameter λ to 0.5 in all experiments.