一個模型到底學到了什麼東西一直都是一個研究課題，也是大家在意的，包括表徵學習、小樣本學習、資料的偏差問題等議題，如果知道模型如何學習、最後學到什麼，都會對這些問題有幫助，有些人就在研究可解釋化（Explainable），今天介紹一個相關的技術：特徵圖視覺化（Feature Map Visualization）。

class FilterVisualizer():
    def __init__(self, size=56, upscaling_steps=12, upscaling_factor=1.2):
        self.size = size
        self.upscaling_steps = upscaling_steps
        self.upscaling_factor = upscaling_factor
        self.model = vgg16(pre=True).cuda().eval()
        set_trainable(self.model, False)

    def visualize(self, layer, filter, lr=0.1, opt_steps=20, blur=None):
        sz = self.size
        # generate random image and register hook
        img = np.uint8(np.random.uniform(150, 180, (sz, sz, 3)))/255  
        activations = SaveFeatures(list(self.model.children())[layer])
		
        # scale the image up upscaling_steps times
        for _ in range(self.upscaling_steps):
            train_tfms, val_tfms = tfms_from_model(vgg16, sz)
            # convert image to Variable that requires grad
            img_var = V(val_tfms(img)[None], requires_grad=True)  
            optimizer = torch.optim.Adam([img_var], lr=lr, weight_decay=1e-6)
            for n in range(opt_steps):  # optimize pixel values for opt_steps times
                optimizer.zero_grad()
                self.model(img_var)
                loss = -activations.features[0, filter].mean()
                loss.backward()
                optimizer.step()
            img = val_tfms.denorm(img_var.data.cpu().numpy()[0].transpose(1,2,0))
            self.output = img
            # calculate new image size then scale image up
            sz = int(self.upscaling_factor * sz)  
            img = cv2.resize(img, (sz, sz), interpolation = cv2.INTER_CUBIC)
            
            # blur image to reduce high frequency patterns
            if blur is not None:
                img = cv2.blur(img,(blur,blur))  
        self.save(layer, filter)
        activations.close()

    def save(self, layer, filter):
        plt.imsave("layer_"+str(layer)+"_filter_"+str(filter)+".jpg", np.clip(self.output, 0, 1))

參考資料

[1]：Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2012). Imagenet classification with deep convolutional neural networks. In Advances in neural information processing systems (pp. 1097-1105).

[2]：Girshick, R., Donahue, J., Darrell, T., & Malik, J. (2014). Rich feature hierarchies for accurate object detection and semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 580-587).

[3]：LeCun, Y., Boser, B. E., Denker, J. S., Henderson, D., Howard, R. E., Hubbard, W. E., & Jackel, L. D. (1990). Handwritten digit recognition with a back-propagation network. In Advances in neural information processing systems (pp. 396-404).

[4]：Matthew, D., & Fergus, R. (2014, September). Visualizing and understanding convolutional neural networks. In Proceedings of the 13th European Conference Computer Vision and Pattern Recognition, Zurich, Switzerland (pp. 6-12).

[5]：Yosinski, J., Clune, J., Nguyen, A., Fuchs, T., & Lipson, H. (2015). Understanding neural networks through deep visualization. arXiv preprint arXiv:1506.06579.

[6]：Zeiler, M. D., Taylor, G. W., & Fergus, R. (2011, November). Adaptive deconvolutional networks for mid and high level feature learning. In 2011 International Conference on Computer Vision(pp. 2018-2025). IEEE.