InstaGAN: Instance-aware Image-to-Image Translation

Authors: Sangwoo Mo, Minsu Cho, Jinwoo Shin

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

Reproducibility Variable Result LLM Response
Research Type Experimental Our comparative evaluation demonstrates the effectiveness of the proposed method on different image datasets, in particular, in the aforementioned challenging cases. Code and results are available in https://github.com/sangwoomo/instagan.
Researcher Affiliation Collaboration Sangwoo Mo , Minsu Cho , Jinwoo Shin , Korea Advanced Institute of Science and Technology (KAIST), Daejeon, Korea Pohang University of Science and Technology (POSTECH), Pohang, Korea AItrics, Seoul, Korea
Pseudocode No The paper describes the sequential mini-batch training technique but does not provide it in a structured pseudocode or algorithm block.
Open Source Code Yes Code and results are available in https://github.com/sangwoomo/instagan.
Open Datasets Yes We sample two classes from various datasets, including clothing co-parsing (CCP) (Yang et al., 2014), multi-human parsing (MHP) (Zhao et al., 2018), and MS COCO (Lin et al., 2014) datasets, and use them as the two domains for translation.
Dataset Splits No The paper details training parameters (learning rate, epochs, batch size) and test evaluation, but does not explicitly describe a validation set split (e.g., percentages, counts, or explicit mention of a validation phase for hyperparameter tuning) or cross-validation setup.
Hardware Specification No The paper mentions 'limited GPU memory' and '4 GPUs in parallel' but does not provide specific GPU models, CPU models, or detailed hardware specifications used for experiments.
Software Dependencies No The paper references various architectural components and optimizers (e.g., Adam, LSGAN, Res Net, Patch GAN) and their original papers, but does not provide specific software library names with version numbers (e.g., PyTorch 1.x, TensorFlow 2.x) that were used in their implementation.
Experiment Setup Yes For all the experiments, we simply set λcyc = 10, λidt = 10, and λctx = 10 for our loss (7). We used Adam (Kingma & Ba, 2014) optimizer with batch size 4, training with 4 GPUs in parallel. All networks were trained from scratch, with learning rate of 0.0002 for G and 0.0001 for D, and β1 = 0.5, β2 = 0.999 for the optimizer. Similar to Cycle GAN (Zhu et al., 2017), we kept learning rate for first 100 epochs and linearly decayed to zero for next 100 epochs for multi-human parsing (MHP) (Zhao et al., 2018) and COCO (Lin et al., 2014) dataset, and kept learning rate for first 400 epochs and linearly decayed for next 200 epochs for clothing co-parsing (CCP) (Yang et al., 2014) dataset