In this paper, we conduct experiments using six settings with Adam optimizer [18]. For the contrastive coefficientλ (see Algorithm 1), the value is fixed at 1.0 for a fair comparison with [19, 8]. In all experiments, we use the temperaturet = 1.0. We stop training GANs with SNDCGAN, SNResGAN, and BigGAN architectures after 200k, 100k, and 80k generator updates, respectively. Experimental setup used for Table 3 in the main paper: FID values on CIFAR10 dataset are reported using the setting (E) with the batch size of 64.

International [46] Y. Liu, labels: Learning, 2019.

By pretraining with a mix of both framelevel (spatial) and video-level (temporal) objectives, our model not only learns to match images to temporally corresponding words, but also to contextualize what is happening globally over time.