你好！请问这预训练模型的数据集是多大的呀，训练的一些参数可以分享一下吗？数据集是只有database下面的几张图吗？请问有补充的数据集不

Currently, imagenet weights are used. Is it possible to train the weights based on the indexing images? (using techniques such as unsupervised learning)

waitting ~~~~~!! 3Q

你使用的keras的vgg16相对比较慢， 如果要换成keras的resnet， code哪些部分要改变？谢谢

Indexing worked fine

Using TensorFlow backend.
--------------------------------------------------
         feature extraction starts
--------------------------------------------------
2019-02-18 01:14:32.787757: I tensorflow/core/platform/cpu_feature_guard.cc:141] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2 FMA
extracting feature from image No. 1 , 9 images in total
extracting feature from image No. 2 , 9 images in total
extracting feature from image No. 3 , 9 images in total
extracting feature from image No. 4 , 9 images in total
extracting feature from image No. 5 , 9 images in total
extracting feature from image No. 6 , 9 images in total
extracting feature from image No. 7 , 9 images in total
extracting feature from image No. 8 , 9 images in total
extracting feature from image No. 9 , 9 images in total
--------------------------------------------------
      writing feature extraction results ...
--------------------------------------------------

dgtlmoon@acer:/tmp/flask-keras-cnn-image-retrieval$ ls -al index.save 
-rw-rw-r-- 1 dgtlmoon dgtlmoon 20741 Feb 18 01:14 index.save

but...

$ python query_online.py -query query/541daeb93a3e4cd8a101b7d5e83a6779.jpg -index ./index.save -result result/
Using TensorFlow backend.
[[0.         0.27786013 0.         ... 0.00188846 0.02606245 0.        ]
 [0.05857334 0.06311509 0.03466835 ... 0.         0.         0.        ]
 [0.02201021 0.         0.05797203 ... 0.01631281 0.06619522 0.01133065]
 ...
 [0.03652793 0.         0.01706417 ... 0.         0.10876585 0.        ]
 [0.01404525 0.         0.01023624 ... 0.         0.01086351 0.03146496]
 [0.03362079 0.00451447 0.         ... 0.         0.05480169 0.01739029]]
Traceback (most recent call last):
  File "query_online.py", line 28, in <module>
    imgNames = h5f['dataset_2'][:]
  File "h5py/_objects.pyx", line 54, in h5py._objects.with_phil.wrapper
  File "h5py/_objects.pyx", line 55, in h5py._objects.with_phil.wrapper
  File "/usr/local/lib/python2.7/dist-packages/h5py/_hl/dataset.py", line 463, in __getitem__
    selection = sel2.select_read(fspace, args)
  File "/usr/local/lib/python2.7/dist-packages/h5py/_hl/selections2.py", line 103, in select_read
    return ScalarReadSelection(fspace, args)
  File "/usr/local/lib/python2.7/dist-packages/h5py/_hl/selections2.py", line 88, in __init__
    raise ValueError("Illegal slicing argument for scalar dataspace")
ValueError: Illegal slicing argument for scalar dataspace

feat[0]/LA.norm(feat[0])

在提取特征那里，您好像是获取了1000维的score作为特征并不是4096维的特征。希望您看下！

你好呀~我在执行
python query_online.py -query database/001_accordion_image_0001.jpg -index featureCNN.h5 -result database 的时候提示以下错误是什情况呢？
ValueError: Illegal slicing argument for scalar dataspace

看了您的博客 http://yongyuan.name/blog/layer-selection-and-finetune-for-cbir.html?nsukey=N9oNq4fRSP%2FTBBMuA%2FGA9hb3dPzeOLNkGKvoYUUMRnPtwQbQjRqPm70SyQcuhTtLZSI5jBFsmeDLohW7s5rd%2FeZXGF4e0TbMwCEqh7VxxeGRKkgbhyJOHIj0yaynoglNQ4Y%2BV2ifXmF41tp%2FECVmERWqM%2BU55apP6mGdUc%2FQpAt5HB%2FxE5uwKI1MnYhzhsl3huTMx5Wk%2FB5dJz8RsdEcGA%3D%3D 提到了rmac 和 crow pooling，如何使用在该项目里呢

作者你好，从你的方法当中学到很多。 现在有一个问题想请教一下。 在海量数据的情况下，h5py的读取速度和保存速度会变得越来越慢。 我自己做实验的情况来看，刚开始数据从提取特征到最后查找相似度结束大概用时10ms左右。5w张图片以后， 单h5文件的读取和矩阵计算已经占到200ms以上了。 你这边有没有什么好的经验，如何在海量数据下保证图片处理过程的实时性呢?

test.py: error: the following arguments are required: --database, --index

跑在gpu上占用了4G的显存 也太大了吧 这个正常吗
Created TensorFlow device (/job:localhost/replica:0/task:0/device:GPU:0 with 4278 MB memory)

When execute index.py, it will download model file from https://github.com/fchollet/deep-learning-models/releases/download/v0.1/vgg16_weights_tf_dim_ordering_tf_kernels_notop.h5, but the download speed is very slow and always failed.
then I download the model file from browser. But it can not load model from my downloaded file.
How to load model from my local file? Thanks!

有两个问题想请教下:
1.除了余弦相似度以外还有啥比较推荐的相似度计算方式吗
2. 预训练模型是在imagenet上训练的 它的特征为何可以用它其他新的数据集上 作为特征提取??

谢谢

如题

没有在代码里面找到关于flask的部分，是还没有添加进来吗，谢谢啦～

重复定义了，所以建立索引的时候处理图片的时间较长

why do this? thank you for your answer.
scores = np.dot(norm_feat, feats.T) rank_ID = np.argsort(scores)[::-1] rank_score = scores[rank_ID]

python3.6的情况下，例如tf、cuda等还有要求吗？

图像特征求出来之后，普遍的做法是遍历所有的特征求余弦距离，如果特征特别多效率是不是不够高，如何用哈希索引之类来帮助我们在保持精度的情况下更高效的进行检索呢？ @

Can you expose the code of flask?

大哥
flask 有和这个代码 配合的flask界面了吗

您好，感谢您的分享！ 因为我还是个图像检索方面的小白，所以想系统地学习一下，请问这个项目您是根据哪篇论文做出来的吗？ 或者是什么其他的理论资料呢？ 谢谢了！！期待您的回复

请问预训练模型来提取的特征 但是对于新的图片集这样的特征为何有用呀?
谢谢

你好，我在ubuntu下尝试配置这个代码，但是碰到了一些问题

alec@alec-CP65S://home/alec/cnncode/flask-retrival$ python extract_CNN_VGG_Keras.py
Using Theano backend.
Traceback (most recent call last):

File "extract_CNN_VGG_Keras.py", line 102, in out = model.predict(im)
File "/usr/local/lib/python2.7/dist-packages/keras/models.py", line 724, in predict
return self.model.predict(x, batch_size=batch_size, verbose=verbose)
File "/usr/local/lib/python2.7/dist-packages/keras/engine/training.py", line 1269, in predict`
...
...
File "/usr/local/lib/python2.7/dist-packages/theano/gof/cmodule.py", line 302, in dlimport
rval = import(module_name, {}, {}, [module_name])

ImportError: ('The following error happened while compiling the node', CorrMM{valid, (1, 1), (1, 1)}(IncSubtensor{InplaceSet;::, ::, int64:int64:, int64:int64:}.0, Subtensor{::, ::, ::int64, ::int64}.0), '\n', '/home/alec/.theano/compiledir_Linux-4.2--generic-x86_64-with-Ubuntu-14.04-trusty-x86_64-2.7.6-64/tmpQAGgtF/c4bfcd4af5f141b0dbcc5236ff225cfd.so: undefined symbol: openblas_get_num_threads', '[CorrMM{valid, (1, 1), (1, 1)}(<TensorType(float32, 4D)>, <TensorType(float32, 4D)>)]')

应该是out = model.predict(im)中传递的参数有问题？尝试了几种图像，感觉应该不是数据集的问题，但是不知道怎么调整了，想请教一下...

还有个疑问就是，这份代码关于图像检索的部分是不是就是extract_CNN_VGG_Keras.py和searchDemo.py两个文件，deep_dream.py是与此无关的是吗？

I can't find the flask web code，the code was deleted？

你好
我正在尝试你的作品。
请教一个问题，请问该方法总结来说是以下方法吗？
用vgg16模型提取query image的特征，再依次做相似度计算，取出相似度最高的图片。
请问以上方法适用于「相似服装」的查找吗。

市面上相似服装的查找的方法最近有triplet loss的方法。请问triplet loss方法和本方法都有哪些差异呢。
谢谢

有上万张图片需要提取特征，请问有没有能够加速提取速度的tips？

报错代码：h5f.create_dataset('dataset_2', data = names)
原因：create_dataset 只接受 ascii 的数据，如果你是Python3或者python使用utf-8的编码就会报错。
解决办法：
#h5f.create_dataset('dataset_2', data = names)
namess = []
for j in names:
namess.append(j.encode())
h5f.create_dataset('dataset_2', data = namess)
h5f.close()

Sort of mentioned in #14 , but more specifically I think the problem is that model.extract_feat(img_path) in index.py is always looking at the top/last layer which is the most 'abstract' in VGG16, so if the network can get a good confidence score on something like seeing a dog in the image, then other images will always be 'similar'.

(Because we are using the existing trained weights)

I'm not totally sure what todo , but feels to me like the network should pull features from a middle layer, so that it's comparing edges, lines, and other low level objects instead of high level abstract information? thoughts?

for example

• If I have a lot of photos of paintings, but I want to search for very similar paintings.. the photos with the border/frame around the painting will always be more 'similar' regardless of the contents of the painting

大佬好、请教下 你写的flask web界面代码方便分享一下吗，我只看到了soutu这个、好像跟你使用的这个特征不太一样、这个使用的界面代码方便分享一下吗

感谢整理分享的论文！
看了下您的实现，用VGG16提取特征，把两张图片的点积作为相似分数，后续有考虑优化吗？

现在提取的是512维，可以减小吗