Résumé : Vietnam is such an attractive tourist destination with its stunning and pristine landscapes and its top-rated unique food and drink. Among thousands of Vietnamese dishes, foreigners and native people are interested in easy-to-eat tastes and easy-to-do recipes, along with reasonable prices, mouthwatering flavors, and popularity. Due to the diversity and almost all the dishes have significant similarities and the lack of quality Vietnamese food datasets, it is hard to implement an auto system to classify Vietnamese food, therefore, make people easier to discover Vietnamese food. This paper introduces a new Vietnamese food dataset named VinaFood21, which consists of 13,950 images corresponding to 21 dishes. We use 10,044 images for model training and 6,682 test images to classify each food in the VinaFood21 dataset and achieved an average accuracy of 74.81% when fine-tuning CNN EfficientNet-B0. (https://github.com/nguyenvd-uit/uit-together-dataset)