fix gram matrix in style transfer

chapter_computer-vision/neural-style.md

@@ -123,7 +123,7 @@ def content_loss(y_hat, y):
 def gram(x):
     c, n = x.shape[1], x.size // x.shape[1]
     y = x.reshape((c, n))
-    return nd.dot(y, y.T) / n
+    return nd.dot(y, y.T) / (c * n)
 ```
 
 和对应的损失函数，这里假设样式图像的样式特征协方差已经预先计算好了。
@@ -147,7 +147,7 @@ def tv_loss(y_hat):
 
 ```{.python .input  n=12}
 style_channels = [net[l].weight.shape[0] for l in style_layers]
-style_weights = [1e4 / c**2 for c in style_channels]
+style_weights = [1e4] * len(style_channels)
 content_weights, tv_weight = [1], 10
 ```