It's great to see a new chinese hands-on book for DL. But I noticed several tutorials are very similar https://github.com/mli/gluon-tutorials-zh. It is often recommended to acknowledge the original author. For example, people often put the following sentence at the beginning or end of the tutorial:

This tutorial is motivated (or adapted) from xxx

计算loss的函数：

def binary_loss(y_pred, y):
    logits = (y * y_pred.clamp(1e-12).log() + (1 - y) * (1 - y_pred).clamp(1e-12).log()).mean()
    return -logits

前半部分是不是应该使用平均值？即

logits = (y * y_pred.clamp(1e-12).log().mean() + (1 - y) * (1 - y_pred).clamp(1e-12).log()).mean()

你好， 我在看您博客上介绍fcn的那篇文章，有个地方不太懂，特来讨教： 最简单的 fcn 前面是一个去掉全连接层的预训练网络，然后将去掉的全连接变为 1x1 的卷积，输出和类别数目相同的通道数，比如 voc 数据集是 21 分类，那么输出的通道数就是 21，然后最后接一个转置卷积将结果变成输入的形状大小，最后在每个 pixel 上做一个分类问题，使用交叉熵作为损失函数就可以了。

通道数为什么也得是21，这个地方不太懂。。

fcn语义分割的代码。 python main.py train Traceback (most recent call last): File "main.py", line 21, in from data import VocSegDataset, img_transforms, COLORMAP

ImportError: cannot import name 'VocSegDataset'

Anser of the exercise is wrong, the second column of j is wrong due to wrong variable name reference in the code: "m.grad.data.zero_() # 归零之前求得的梯度", change "m" to "x" to fix it.

under MacOS X 10.13, conda, python 3.6 Matplotlib 2.1.2

1. in plot_decision_boundary: change "plt.scatter(x[:, 0], x[:, 1], c=y, cmap=plt.cm.Spectral)" to "plt.scatter(x[:, 0], x[:, 1], c=y.reshape(-1), cmap=plt.cm.Spectral)"
2. change "plt.scatter(x[:, 0], x[:, 1], c=y, s=40, cmap=plt.cm.Spectral)" to "plt.scatter(x[:, 0], x[:, 1], c=y.reshape(-1), s=40, cmap=plt.cm.Spectral)"

@L1aoXingyu

作者你好，非常感谢您的分享。这里我提出一点小小的意见。

权重衰减可以通过 torch.optim.lr_scheduler 的相关类如 ExponentialLR完成，从而使权重衰减和训练逻辑分离。作权重衰减相关介绍的时候可以这么来介绍（实际使用也是这么使用比较广泛）。

https://pytorch.org/docs/stable/optim.html#how-to-adjust-learning-rate

net.py文件中的这段： vgg = models.vgg19(pretrained=True) self.feature = nn.Sequential(*list(vgg.children())[:-1]) self.feature.add_module('global average', nn.AvgPool2d(9))

前两行代码的输出尺寸是7*7，而第三行平均池化的卷积核是9，运行后会出现这样的错误：Given input size: (512x7x7). Calculated output size: (512x0x0). Output size is too small

我的pytorch版本是1.6，是不是版本的问题？但按理说，Pytorch版本的迭代不会影响VGG这种经典网络的参数啊。这到底是怎么回事呢？

画出参数更新之前的结果

w0 = w[0].data[0] w1 = w[1].data[0] b0 = b.data[0]

plot_x = np.arange(0.2, 1, 0.01) plot_y = (-w0 * plot_x - b0) / w1 # 上面会导致乘法不能正常操作, 应该讲w0,w1,w2转换成numpy array 才能调通

In[10]:# 画出参数更新之前的结果 w0 = w[0].data[0] w1 = w[1].data[0] b0 = b.data[0]

plot_x = np.arange(0.2, 1, 0.01) plot_y = (-w0 * plot_x - b0) / w1 TypeError TypeError: mul() received an invalid combination of arguments - got (numpy.ndarray), but expected one of:

• (Tensor other) didn't match because some of the arguments have invalid types: (numpy.ndarray)
• (Number other) didn't match because some of the arguments have invalid types: (numpy.ndarray)