Unsupervised Discovery of Interpretable Directions in the GAN Latent Space
Authors: Andrey Voynov, Artem Babenko
ICML 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Here we evaluate our approach on several datasets in terms of both quantitative and qualitative results. In all experiments, we do not exploit any form of external supervision and operate in a completely unsupervised manner. ... We measure the model performance in terms of the mean average error (MAE)... |
| Researcher Affiliation | Collaboration | Andrey Voynov 1 Artem Babenko 1 2 1Yandex, Russia 2National Research University Higher School of Economics , Moscow, Russia. |
| Pseudocode | No | The paper describes the learning protocol and practical details using descriptive text and diagrams (Figure 3) but does not include any explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | The implementation of our method is available online1. 1http://github.com/anvoynov/ Gan Latent Discovery |
| Open Datasets | Yes | 1. MNIST (Le Cun, 1989), containing 32 × 32 images. ... 2. Anime Faces dataset (Jin et al., 2017), containing 64 × 64 images. ... 3. Celeb A-HQ dataset (Liu et al., 2015), containing 1024 × 1024 images. ... 4. Big GAN generator (Brock et al., 2019) trained on ILSVRC dataset (Deng et al., 2009), containing 128 × 128 images. |
| Dataset Splits | No | The paper mentions 'The dataset has separate train and test subsets' for ECSSD, but does not explicitly state validation dataset splits or mention cross-validation for any of the datasets used. |
| Hardware Specification | Yes | All the experiments were performed on the NVIDIA Tesla v100 card. |
| Software Dependencies | No | The paper mentions software components like 'Adam optimizer' and specific model architectures like 'U-net' and 'Res Net-18' but does not provide specific version numbers for these or any other software dependencies needed for replication. |
| Experiment Setup | Yes | We always train the models with a constant learning rate 0.0001. We perform 2 × 10^5 gradient steps for Prog GAN and 10^5 steps for others as the first has a significantly higher latent space dimension. We use a batch size of 128 for Spectral Norm GAN on the MNIST, and Anime Faces datasets, a batch size of 32 for Big GAN, and a batch size of 10 for Prog GAN. ... We train U on the pseudolabeled dataset with Adam optimizer and the per-pixel crossentropy loss with the temperature 10.0. We perform 15000 steps with the initial rate of 0.005 and decrease it by 0.2 every 4000 steps and a batch size equal to 128. |