One-Shot Unsupervised Cross Domain Translation
Authors: Sagie Benaim, Lior Wolf
NeurIPS 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our experiments indicate that the new method does as well, when trained on one sample x, as the existing domain transfer methods, when these enjoy a multitude of training samples from domain A. Our code is made publicly available at https://github.com/sagiebenaim/One Shot Translation. We perform a wide variety of experiments and demonstrate that OST outperforms the existing algorithms in the low-shot scenario. On most datasets the method also presents a comparable accuracy with a single training example to the accuracy obtained by the other methods for the entire set of domain A images. We conduct a number of quantitative evaluations, including style and content loss comparison as well as a classification accuracy test for target images. For the MNIST to SVHN translation and the reverse, we conduct an ablation study, showing the importance of every component of our approach. For this task, we further evaluate our approach, when more samples are presented, showing that OST is able to perform well on larger training sets. |
| Researcher Affiliation | Collaboration | 1The School of Computer Science , Tel Aviv University, Israel 2Facebook AI Research |
| Pseudocode | No | The paper describes the method and uses mathematical equations for losses, but it does not include any explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | Our code is made publicly available at https://github.com/sagiebenaim/One Shot Translation. |
| Open Datasets | Yes | MNIST [26] image to an Street View House Number (SVHN) [27] image. We consider the tasks of two-way translation from Images to Monet-style painting [2], Summer to Winter translation [2] and the reverse translations. We consider the translation of Google Maps to Aerial View photos [24], Facades to Images [30], Cityscapes to Labels [31] and the reverse translations. |
| Dataset Splits | No | The paper does not provide explicit training, validation, or test split percentages or counts. It refers to training on a single sample or the entire training set, but no detailed dataset split information is given for reproducibility. |
| Hardware Specification | No | The paper does not specify any hardware details such as GPU models, CPU types, or memory specifications used for running the experiments. It does not mention where the models were trained (e.g., on specific cloud instances or local machines with detailed specs). |
| Software Dependencies | No | The paper mentions using |
| Experiment Setup | Yes | Both the shared and unshared encoders (resp. decoders) consist of between 1 and 2 2-stride convolutions (resp. deconvolutions). The shared encoder consists of between 1 and 3 residual blocks after the convolutional layers. The shared decoder also consists of between 1 and 3 residual blocks before its deconvolutional layers. Batch normalization and Re LU activations are used between layers. Cycle GAN employs a number of additional techniques to stabilize training, which OST borrows. The first is the use of a Patch GAN discriminator [24], and the second is the use of least-square loss for the discriminator [23] instead of negative log-likelihood loss. For the MNIST to SVHN translation and the reverse translation, the Patch GAN discriminator is not used, and, for these experiments, where the input is in R3 32 32, the standard DCGAN [28] architecture is used. The losses used, with their corresponding alpha parameters (α1, α2, α3, α4, α5), are defined in equations (1)-(5) and (10)-(14). |