Gastric cancer is one of the significant lethal cancers in China. Most patients are diagnosed at an advanced stage, and if gastric cancer can be detected at an early stage through large-scale screening, patient survival can be considerably improved. In China, there are two obstacles toward the large-scale screening of early gastric cancer. One is that endoscopy is overly invasive, resulting in low patient acceptance, and the other is that the number of endoscopists is too small compared with China's large population. A capsule endoscopic robot can alleviate the first obstacle, and the second obstacle is expected to be solved using artificial intelligence. We transferred the state-of-the-art Big Transfer (BiT) to a small dataset of early gastric cancer endoscopic images and built an early gastric cancer classification model based on white-light endoscopic images. We customized the BiT hyperparameter rules in transfer learning based on local situations. The batch size was selected according to the GPU memory limit, and based on the batch size, the linear scale rules were used to adjust the optimizer's initial learning rate dynamically. The total number of training images for the small dataset was set at 256000, on which other hyperparameters of the transfer learning were set. This study experimented with multiple models having the same structure of ResNet-v2 but different depths and widths. The best model has a depth of 101 and a width three times the original one. It has an accuracy of 97.14%, an F1 score of 94.77%, a sensitivity of 90.67%, and a specificity of 99.73% on the test set. Furthermore, the results show that the effect of batch size on the model training is statistically insignificant. This paper transferred a large model to a small dataset of endoscopic images with the BiT customization. This will promote the use of large-scale models in the field of endoscopic image analysis, which can help realize a large-scale screening of early gastric cancer.