[Background] Drinking behavior after diuretics administration involves complex neuronal processes, in which participating hormonal signal is not well delineated. [purpose] To gain insight into frequency and timing of drinking behavior after diuretics, we employed the advantage and power of deep learning. [Method] Male mice with or without intraperitoneal injection of tolvaptan (20 mg/kg) were placed in metabolic cages and their behavior was monitored by video recording for six hours. Individual video frames were obtained using ffmpeg program. Mice were detected as an object and labelled as “drinking” or “away from water”. Labeled images were used to develop learning models of deep neural network (DNN) in Tensorflow and pyhton program. The trained model was evaluated its accuracy and further used to detect drinking behaviors of new mouse group in conda virtual environment. [Result] Our DNN model successfully detected a series of mouse behavior. We found a marked increase in the drinking behavior after diuretics treatment. [Discussion] Water-taking behaviors, which were composed of tens of thousands of images per mouse, can be automatically analyzed using trained DNN with high accuracy. Our results demonstrate the usefulness of deep learning in a pharmacological study of animal behavior.