Self-Attention Generative Adversarial Networks

Authors: Han Zhang, Ian Goodfellow, Dimitris Metaxas, Augustus Odena

ICML 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We have conducted extensive experiments on the Image Net dataset to validate the effectiveness of the proposed self-attention mechanism and stabilization techniques. SAGAN significantly outperforms prior work1, boosting the best reported Inception score from 36.8 to 52.52 and reducing Fr echet Inception distance from 27.62 to 18.65.
Researcher Affiliation Collaboration 1Department of Computer Science, Rutgers University 2Google Research, Brain Team.
Pseudocode No No pseudocode or algorithm blocks are present in the paper.
Open Source Code Yes Our code is available at https://github.com/ brain-research/self-attention-gan.
Open Datasets Yes To evaluate the proposed methods, we conducted extensive experiments on the LSVRC2012 (Image Net) dataset (Russakovsky et al., 2015).
Dataset Splits No To evaluate the proposed methods, we conducted extensive experiments on the LSVRC2012 (Image Net) dataset (Russakovsky et al., 2015). (The paper mentions using the ImageNet dataset but does not explicitly state the training, validation, or test splits used within the text.)
Hardware Specification No No specific hardware details (e.g., GPU models, CPU models, memory) are mentioned for running experiments.
Software Dependencies No For all models, we use the Adam optimizer (Kingma & Ba, 2015) with β1 = 0 and β2 = 0.9 for training. (No specific software versions like PyTorch, TensorFlow, or Python are mentioned.)
Experiment Setup Yes All the SAGAN models we train are designed to generate 128 128 images. By default, spectral normalization (Miyato et al., 2018) is used for the layers in both the generator and the discriminator. Similar to (Miyato & Koyama, 2018), SAGAN uses conditional batch normalization in the generator and projection in the discriminator. For all models, we use the Adam optimizer (Kingma & Ba, 2015) with β1 = 0 and β2 = 0.9 for training. By default, the learning rate for the discriminator is 0.0004 and the learning rate for the generator is 0.0001.