chencodex / bilinear_cnn_dog_classifi Goto Github PK

在网络模型中，在BCNN的步骤中，做了如下的一个步骤：
self.phi_I = tf.divide(self.phi_I,784.0)
这里为什么要除784啊？

with tf.name_scope('preprocess') as scope:
mean = tf.constant([110.61220668946963, 127.15494646868024, 135.13440353131017]

这是整个数据集图片的均值吗

Hi.
Excellent work.
But could you tell me the accuracy that you model can achieve?
Best wishes.

您好，小白请教您一下，down_pic.py中的test.txt是训练数据集的地址吗？clone下来没有这个文件，可不可以发我一份，万分感谢。[email protected]

我用的时stanford dogs数据集，当对测试集数据生成hdf5格式时成功，但是对训练集数据生成hdf5格式时报错：

请问是什么原因

Hello!
I am repeating your source code, would like to ask you under the final test of the correct rate can be achieved? Because know the correct rate can give me a reference. Can you provide detailed running steps to help me better reproduce? Thank you for your help!
Thanks very much!

Hi, chencodeX

I think that the core codes of BCNN is the following codes from bcnn_finetuning.py in the end of class vgg16

        ''' Reshape conv5_3 from [batch_size, height, width, number_of_filters]
           to [batch_size, number_of_filters, height, width]'''
        self.conv5_3 = tf.transpose(self.conv5_3, perm=[0,3,1,2])


        ''' Reshape conv5_3 from [batch_size, number_of_filters, height*width] '''
        self.conv5_3 = tf.reshape(self.conv5_3,[-1,512,784])

        ''' A temporary variable which holds the transpose of conv5_3 '''
        conv5_3_T = tf.transpose(self.conv5_3, perm=[0,2,1])

        '''Matrix multiplication [batch_size,512,784] x [batch_size,784,512] '''
        self.phi_I = tf.matmul(self.conv5_3, conv5_3_T)

But I get confused when I try to understand it.

When I read the paper and other docs, I think that there must have 2 CNNs and they get the outer product when combine the outputs of them, then we can connect it to FC layers or others.

But in the BCNN codes, when they build models, it seems like there is only 1 CNN and it get outer product of it self.

I think 2 CNNs can extract different image features, but I don't understand the code of BCNN.

Anything wrong of my understanding? Could you give me some suggestions?

Thank you!

请问这个试过今年百度-西交的那个比赛么？准确率可以达到多少？

请问你用的哪个数据集，还有你这个最后的准确率大概有多少？

您好，请问bcnn_DD_woft.py和bcnn_finetuning.py这两个文件，前者只训练最后全连接层，后者finetune整个网络，请问这两个训练过程有什么关系吗，因为我看到后者的网络还是使用的最初始的vgg16_weights.npz