bojone / capsule Goto Github PK

Why in the last con2d layer the activation function is relu and not the squash function ?
shouldn't it be the squash function ?

Hi;
I could'nt understand why the following line is needed in test code?

cnn = Reshape((-1, 128))(cnn)

and what does 128 correspond here?

thanks in advance

From the paper "Dynamic Routing Between Capsules", I think the squash function should be
scale = K.sqrt(s_squared_norm)/ (K.sqrt(s_squared_norm) + s_squared_norm)
rather than the function written in you code
scale = K.sqrt(s_squared_norm)/ (0.5 + s_squared_norm)
When I change it to function in the paper, the accuracy get improved to 0.936. Could you try it?

Traceback (most recent call last):
File "capsule_test.py", line 46, in
capsule = Capsule(10, 16, 3, True)(cnn)
File "/home/.../anaconda3/lib/python3.6/site-packages/keras/engine/topology.py", line 603, in __call
__
output = self.call(inputs, **kwargs)
File "/data/.../study/Capsule/Capsule_Keras.py", line 60, in call
c = K.softmax(b)
File "/home/.../anaconda3/lib/python3.6/site-packages/keras/backend/theano_backend.py", line 1552, i
n softmax
return T.nnet.softmax(x)
File "/home/.../anaconda3/lib/python3.6/site-packages/theano/tensor/nnet/nnet.py", line 809, in soft
max
return softmax_op(c)
File "/home/.../anaconda3/lib/python3.6/site-packages/theano/gof/op.py", line 615, in call
node = self.make_node(*inputs, **kwargs)
File "/home/.../anaconda3/lib/python3.6/site-packages/theano/tensor/nnet/nnet.py", line 431, in make
_node
x.type)
ValueError: x must be 1-d or 2-d tensor of floats. Got TensorType(float32, 3D)

I try tensorflow backend , no error.

b = K.batch_dot(o, u_hat_vecs, [2, 3])
ValueError: Can not do batch_dot on inputs with shapes (None, 10, 10, 16) and (None, 10, None, 16) with axes=[2, 3]. x.shape[2] != y.shape[3] (10 != 16)

而且b = K.batch_dot(o, u_hat_vecs, [2, 3])
中b不需要和上一轮的b进行叠加吗

当embedding_dim=768或者是embedding=300，
也就是说如何接词向量model和bert，
参数如何设设置

您好，我想问下您使用了conv1d，和local_conv1d，但是这种操作还算是矩阵乘法么？我不太清楚conv1d的机制，网上好像也都讲的不太清楚。能详细说明下么，这个操作能代替W_ij*u_i？
还有，我不太理解您自己定义的softmax，为什么要减去一个最大值？再者看原文，softmax中求和应该是对j累加，也就是说对caps_num累加，看您的程序好像是对i累加了。
最后，Hiton貌似是不希望增加池化层的，但看你的程序好像加了个cnn = AveragePooling2D((2,2))(cnn)，是否不太合适？
最后谢谢您分享的代码，它在速度上的提升令人振奋！

The program is leaving me an error while running
For "cnn = Conv2D(64, (3, 3), activation='relu')(input_image)" in line 41
An error returns like
"ValueError: The channel dimension of the inputs should be defined. Found None."

Keras version == 2.1.5

Thanks for your contribution!~

大神您好。
阅读你关于capsule的三篇文章收益良多。
参考你的代码，换了一个数据集，迭代次数比你例子多，也会出现loss为nan的情况。按你的提升替换了squash中的K.epsilon()为0.001，并没有解决这个问题。

请问是什么地方会导致这个计算结果呢？

b = K.batch_dot(o, u_hat_vecs, [2, 3])
应该是
b += K.batch_dot(o, u_hat_vecs, [2, 3])

請問一下蘇大，由於您的capsule_test裏的前一層採用了：
Reshape((-1,128))應該是對應於primary capsule，
這是否是對應於128 capsule的dim而非數量？
因為我稍微看了一下原本Hinton的paper和下面這個實現
https://github.com/XifengGuo/CapsNet-Keras
似乎都把8-dim 當作primary capsule的特徵維度。
雖然說dim和num就算反過來也完全可以做運算，
只是想問一下跟原paper之間的關係是否是想選擇128當作capsule num？

P.S. Blog真的解釋得非常好，真的獲益良多，感謝

执行这一句的时候 greater = np.sort(Y_pred, axis=1)[:,-2] > 0.5 #判断预测结果是否大于0.5

遇到一个问题
numpy.core._internal.AxisError: axis 1 is out of bounds for array of dimension 1

不知道怎么解决了。谢谢。

你好,有个问题想请教以下:

#搭建CNN+Capsule分类模型
input_image = Input(shape=(None,None,1))
cnn = Conv2D(64, (3, 3), activation='relu')(input_image)
cnn = Conv2D(64, (3, 3), activation='relu')(cnn)
cnn = AveragePooling2D((2,2))(cnn)
cnn = Conv2D(128, (3, 3), activation='relu')(cnn)
cnn = Conv2D(128, (3, 3), activation='relu')(cnn)
cnn = Reshape((-1, 128))(cnn)
capsule = Capsule(10, 16, 3, True)(cnn)
output = Lambda(lambda x: K.sqrt(K.sum(K.square(x), 2)), output_shape=(10,))(capsule)

model = Model(inputs=input_image, outputs=output)
model.compile(loss=lambda y_true,y_pred: y_true*K.relu(0.9-y_pred)**2 + 0.25*(1-y_true)*K.relu(y_pred-0.1)**2,
              optimizer='adam',
metrics=['accuracy'])

在这个模型中,output = Lambda(lambda x: K.sqrt(K.sum(K.square(x), 2)), output_shape=(10,))(capsule) 这一行操作的输入capsule的的结构是[batch_size,num_capsule,dim_capsule]. output执行的含义是:

我的理解,这个操作是将capsule的每一个分量转换成可能性(概率),那么这个地方是否可以换乘求和或者求平均呢? 采用是否有其他的含义呢?

谢谢了.

非常感谢的作者的sharing.

我想用Capsule来尝试处理regression的问题，想法是最后吧vector 用neural转换成一个scalar,不知道这么做是否靠谱。