A fuel injection device of an internal combustion engine executes tentative learning repeatedly while engine cooling water temperature increases. An actual learning value is renewed with the newest instrumental error learning value so that accuracy of the actual learning value is improved gradually. Thus, a variation of characteristics due to an instrumental error of a supply pump is absorbed as early as possible. In addition, accuracy of a tentative learning value obtained through the tentative learning is quickly brought close to the accuracy of the actual learning value obtained through main learning. As a result, a learning period at vehicle factory shipment is shortened and productivity at a vehicle factory is improved.