|Number of watchers on Github||412|
|Number of open issues||8|
|Average time to close an issue||1 day|
|Open pull requests||0+|
|Closed pull requests||0+|
|Last commit||about 4 years ago|
|Repo Created||almost 5 years ago|
|Repo Last Updated||over 1 year ago|
|Organization / Author||stdcoutzyx|
|Do you use DeepID_FaceClassify? Leave a review!|
|View open issues (8)|
|View DeepID_FaceClassify activity|
|View on github|
|Fresh, new opensource launches 🚀🚀🚀|
Trendy new open source projects in your inbox! View examples
Implementation of DeepID using theano.
You have to install theano and related libs. There are enough information in the theano document. So i will assume all of the readers have installed theano correctly.
The structure of my code look like that:
src/ conv_net deepid_class.py deepid_generate.py layers.py load_data.py sample_optimization.py data_prepare vectorize_img.py youtube_data_split.py youtube_img_crop.py
Just as the names of the folders imply, there are two modules without reference to each other in my code. The
data_prepare module is used to prepare data. And the
conv_net module is the implemention of DeepID.
There are two parts which are important and neccessary for the amazing performance of DeepID, namely the net structure of the Convolutional Neural Network and the data.
I had asked the author for the data, got nothing but a polite reply. So in my experiment, some other data are used instead.
Take the youtube face data as an example. There are three levels of folders, which is showed below:
youtube_data/ people_folderA video_folderA img1.jpg img2.jpg imgN.jpg video_folderB people_folderB
The first thing need to be done is to seperate the data into train set and validate set. The way i choose train set and validate set is as below:
At last, i get 7975 imgs as the validation set and 31900 imgs as the train set. Obviously, you will know that there are 1595 classes(persons) totally.
Note: the file prefixed with
youtube is specifically for the youtube data because of the folder structure and the img property. So if you want to deal with some other dataset, please read the code of
*_data_split.py and re-implement them. I believe the code is readable and easy to understand for the readers.
Used to get the face out of the img. Face in youtube data has been aligned into the center of the img. So this programme aims to increase the ratio of the face in the whole img and resize the img into (47,55), which is the input size for the DeepID.
Usage: python youtube_img_crop.py aligned_db_folder new_folder
Used to split data into two set, One is for train and one is for valid.
Usage: python youtube_data_split.py src_folder test_set_file train_set_file
The format of test_set_file and train_set_file is like below. There are two parts in one line, the first is path of the img, the second is label of the img.
youtube_47_55/Alan_Ball/2/aligned_detect_2.405.jpg,0 youtube_47_55/Alan_Ball/2/aligned_detect_2.844.jpg,0 youtube_47_55/Xiang_Liu/5/aligned_detect_5.1352.jpg,1 youtube_47_55/Xiang_Liu/1/aligned_detect_1.482.jpg,1
Used to vectorize the imgs. To make the thousands of imgs into a two-d array, whose size is (m,n). m is the number of samples, n is the 47553.
To avoid occurance of super big file,
vectorize_img.py automatically seperate data into batches with 1000 samples in each batch.
Usage: python vectorize_img.py test_set_file train_set_file test_vector_folder train_vector_folder
Now it's the exciting time.
In the conv_net module, there are five programme files.
Usage: python deepid_class.py vec_valid vec_train params_file
Note: there are so many parameters need to be adjusted for DeepID, so i did not show them directly in the command line for the simple use of my code. If you want to change the epoch num, learning rate, batch size and so on, please change them in the last line of the file.
You can extract the hidden layer whose dimension is 160 with command below:
Usage: python deepid_generate.py dataset_folder params_file result_folder
After running the
deepid_class.py, you will get the output of the programme like that. The first part is the train error and valid error of each epoch, The second part is the summarization of the
epoch, train error, valid error.
epoch 15, train_score 0.000444, valid_score 0.066000 epoch 16, minibatch_index 62/63, error 0.000000 epoch 16, train_score 0.000413, valid_score 0.065733 epoch 17, minibatch_index 62/63, error 0.000000 epoch 17, train_score 0.000508, valid_score 0.065333 epoch 18, minibatch_index 62/63, error 0.000000 epoch 18, train_score 0.000413, valid_score 0.070267 epoch 19, minibatch_index 62/63, error 0.000000 epoch 19, train_score 0.000413, valid_score 0.064533 0 0.974349206349 0.962933333333 1 0.890095238095 0.897466666667 2 0.70126984127 0.666666666667 3 0.392031746032 0.520133333333 4 0.187619047619 0.360666666667 5 0.20526984127 0.22 6 0.054380952381 0.171066666667 7 0.0154920634921 0.128 8 0.00650793650794 0.100133333333 9 0.00377777777778 0.0909333333333 10 0.00292063492063 0.086 11 0.0015873015873 0.0792 12 0.00133333333333 0.0754666666667 13 0.00111111111111 0.0714666666667 14 0.000761904761905 0.068 15 0.000444444444444 0.066 16 0.000412698412698 0.0657333333333 17 0.000507936507937 0.0653333333333 18 0.000412698412698 0.0702666666667 19 0.000412698412698 0.0645333333333
You can also put the second part of the output into a figure with matplotlib.
deepid_generate.py, you will get output like below:
loading data of vec_test/0.pkl building the model ... generating ... writing data to deepid_test/0.pkl loading data of vec_test/3.pkl building the model ... generating ... writing data to deepid_test/3.pkl loading data of vec_test/1.pkl building the model ... generating ... writing data to deepid_test/1.pkl loading data of vec_test/7.pkl building the model ... generating ... writing data to deepid_test/7.pkl
The programme will extract on each sub file of the vectorized data.
After extracting the hidden layer, we can do some other things to prove the effiency of the deepid feature. For example, in the domain of feature retrieval, you can use my another github project to test on the data generated in this project, here is the link.
For comparison, i have done two experiments on the youtube face data for face retrieval.
vectorized_img.py, and do face retrieval exp on that
Note: In both experiments, i use the cosine distance to measure the similarity of two vectors.
Results of face retrieval are below:
Precision means if there is a photo who has the same people with the query image in the top-N results, it's correct. But AP will calculate how many photos who has the same people with the query image in the top-N results.
From the results, we can know the DeepID feature is superior to the pca method with the equal dimension.
. Sun Y, Wang X, Tang X. Deep learning face representation from predicting 10,000 classes[C]//Computer Vision and Pattern Recognition (CVPR), 2014 IEEE Conference on. IEEE, 2014: 1891-1898.