A few-shot image classification problem aims to recognize previously unseen objects with a small amount of data. Many works have been offered to solve the problem, while a simple transfer learning method with the cosine similarity based cross-entropy loss is still powerful compared with other methods. To improve the performance, we propose a novel Non-Probabilistic Cosine similarity (NPC) loss for few-shot classification that can replace the cross-entropy loss with the cosine similarity. A key difference of NPC loss is that it uses values of inputs instead of their probabilities. By simply changing the loss function, our model avoids overfitting on a training set and performs well on few-shot tasks. Experimental results show that the model with NPC loss clearly outperforms those with other loss functions and also achieves excellent performance compared with state-of-the-art algorithms on Mini-Imagenet and CUB-200-2011 datasets.