王博士，您好！缩小batch_size是在main函数这里修改吗？我在这里改为1，还是OOM
parser.add_argument('--batch_size', type=int, default=1, help='number of samples in one batch')

tensorflow.python.framework.errors_impl.ResourceExhaustedError: OOM when allocating tensor with shape[312030,10,50] and type float on /job:localhost/replica:0/task:0/device:CPU:0 by allocator cpu [[node kcnn/embedding_lookup_1 (defined at D:\Jupyter_folder\DKN\src\dkn.py:94) ]] Hint: If you want to see a list of allocated tensors when OOM happens, add report_tensor_allocations_upon_oom to RunOptions for current allocation info.

这是我别的错误
Traceback (most recent call last): File "D:/Jupyter_folder/DKN/src/main.py", line 41, in <module> train(args, train_data, test_data) File "D:\Jupyter_folder\DKN\src\train.py", line 36, in train train_auc = model.eval(sess, get_feed_dict(model, train_data, 0, train_data.size)) File "D:\Jupyter_folder\DKN\src\dkn.py", line 162, in eval labels, scores = sess.run([self.labels, self.scores], feed_dict)

Dr. Wang, thank you so much for your wonderful work.
When I run last step : python main.py, an error occored:

2019-07-15 17:38:14.500279: W T:\src\github\tensorflow\tensorflow\core\framework\op_kernel.cc:1318] OP_REQUIRES failed at conv_ops.cc:673 : Resource exhausted: OOM when allocating tensor with shape[442410,128,9,1] and type float on /job:localhost/replica:0/task:0/device:GPU:0 by allocator GPU_0_bfc

My gpu information is :
2019-07-15 17:37:54.354919: I T:\src\github\tensorflow\tensorflow\core\common_runtime\gpu\gpu_device.cc:1356] Found device 0 with properties:
name: GeForce GTX 1080 Ti major: 6 minor: 1 memoryClockRate(GHz): 1.683
pciBusID: 0000:05:00.0
totalMemory: 11.00GiB freeMemory: 9.11GiB
2019-07-15 17:37:54.538331: I T:\src\github\tensorflow\tensorflow\core\common_runtime\gpu\gpu_device.cc:1356] Found device 1 with properties:
name: GeForce GTX 1080 Ti major: 6 minor: 1 memoryClockRate(GHz): 1.683
pciBusID: 0000:09:00.0
totalMemory: 11.00GiB freeMemory: 9.11GiB

other information before the error occur is :
2019-07-15 17:38:14.475725: I T:\src\github\tensorflow\tensorflow\core\common_runtime\bfc_allocator.cc:674] 1 Chunks of size 2265139200 totalling 2.11GiB
2019-07-15 17:38:14.480071: I T:\src\github\tensorflow\tensorflow\core\common_runtime\bfc_allocator.cc:678] Sum Total of in-use chunks: 5.72GiB
2019-07-15 17:38:14.484854: I T:\src\github\tensorflow\tensorflow\core\common_runtime\bfc_allocator.cc:680] Stats:
Limit: 9244818801
InUse: 6142976256
MaxInUse: 6369439488
NumAllocs: 16017
MaxAllocSize: 2265139200

I wonder how can I solve the problem,thank you

I am new to Deep learning, I am getting error OSError: TransH_entity2vec_50.vec not found.
prior to this step, everything run smoothly,
any help appreciated

Hi,
Thanks for publishing this wonderful project and outstanding paper related to recommendation systems.

As I am currently learning your DKN code after reading your paper, some confusion crossed my mind especially at the part ./data/kg/kg.txt.
Well, although you have described that head and tail are entity ids and relation is relation id, I still don't know how the relation ids are generated as entity ids are generated via previous functions.

I would be grateful if you could explain the process of the generation of relation ids. Look forward to your response.

Big thanks!

Hi, I tried to run the code you provided and compared the results mentioned on your paper.
Results i got are not the same which are mentioned in the paper?

anything wrong I am doing?
below is a snapshot of the result generated where as in your paper you claimed AUC of 65.9+/- 1.2

王老师好，代码中，在训练的时候，是把训练集中的30条划分成history clicked news ，请问为什么这么分，是实验做出来的超参数么
1.为什么不在训练的时候除了其中一条作为候选集其他的都当成历史记录，然后重复这个过程直到训练集中的所有数据都当过候选集呢
2.训练的时候是将用户所有标签为1的记录当成他的历史记录，还是标签为0的也会当成历史记录
3.候选集是怎么生成的？
4.有没有比较过不同条数的历史记录对模型的影响
5.测试集使用的历史记录和训练集一样么，还是使用用户的训练集作为测试时的历史记录
谢谢老师

Dear authors,

I found that the code implementation is slightly different from the structure presented in the paper (Figure 3). In the paper, both the output of the attention and the final output are produced by a two-layer MLP. However, in the code dkn.py, they are implemented as:

attention_weights = tf.reduce_sum(clicked_embeddings * news_embeddings_expanded, axis=-1)

and

self.scores_unnormalized = tf.reduce_sum(user_embeddings * news_embeddings, axis=1)

They are only a inner product between two vectors and there are no any parameters to learn the concat. My question is:

1. Why can this work? Can only updating embeddings without MLP weights get well trained?
2. Why is the code implementation different from the paper?

王老师您好，我在评估模型时将通过sigmoid转换之后得到的预测概率值scores输出，发现他们全部大于0.5，请问模型划分正负例的阈值是多少呢？（应该不是0.5？）

例如某一个epoch得到的scores:
[0.5778407 0.61155593 0.57902104 0.6150518 0.61513776 0.605235
0.5789735 0.57939625 0.58373684 0.5776857 0.5788436 0.5802991
0.58303684 0.57765114 0.58011097 0.5782152 0.58273846 0.59276307
0.5777418 0.5781118 0.58053976 0.5770587 0.5848505 0.6206129
0.5796546 0.5936298 0.5800667 0.5841336 0.5985693 0.5795024
0.57965916 0.57965916 0.5940067 0.58897567 0.581066 0.5788839
0.59576434 0.5785835 0.5783613 0.5984433 0.5783426 0.5845572
0.5854817 0.5777896 0.5783829 0.58128166 0.5843682 0.57781625
0.58254665 0.5966302 0.592908 0.57927406 0.57772964 0.5805329
0.5793939 0.5789313 0.580673 0.62930167 0.57700074 0.5771642
0.5793654 0.5801946 0.5798775 0.57729614 0.5778348 0.5773217
0.5789409 0.5803839 0.58697975 0.57710063 0.5773818 0.57703525
0.57772624 0.58213145 0.58688086 0.58625937 0.5790521 0.583711
0.6132932 0.6129931 0.5814061 0.5788394 0.57913923 0.61161196
0.5792562 0.5790782 0.6079845 0.5920574 0.5793936 0.578782
0.584119 0.5926744 0.5794713 0.6049938 0.6050354 0.60463595
0.6047081 0.61172765 0.619973 0.60173035 0.6009864 0.58058125
0.60348684 0.5824682 0.57886064 0.5884625 0.5974402 0.578675
0.578743 0.57945085 0.5773963 0.57785344 0.5776528 0.57766426
0.577414 0.5773697 0.5899348 0.5813688 0.5976933 0.58020246
0.5853191 0.5883258 0.5884458 0.57786816 0.57844496 0.61082983
0.58969355 0.59313047 0.5914415 0.61171293 0.5790724 0.58127683
0.5841514 0.58132464 0.58002347 0.57958114 0.57698256 0.57736045
0.5829809 0.5780082 0.5776665 0.58005613 0.5967956 0.5783363
0.5775785 0.57833403 0.6010081 0.57743883 0.5820395 0.57780135
0.57779527 0.5832522 0.5774031 0.5813955 0.64714086 0.5897089
0.58454174 0.5777361 0.57789356 0.5920431 0.58439016 0.5799597
0.5790718 0.57887685 0.58060193 0.58858824 0.58267885 0.57806975
0.59110266 0.59110266 0.6118623 0.5787403 0.58097917 0.579126
0.58345675 0.5919414 0.5799597 0.57892144 0.6185592 0.5814157
0.5828195 0.5790548 0.57929236 0.5790717 0.5797003 0.58137774
0.58160037 0.5781356 0.5776759 0.57675505 0.5767635 0.5821508
0.5782573 0.57824135 0.58539546 0.577819 0.57790375 0.583094
0.6229433 0.5897612 0.5794023 0.5936962 0.59367275 0.59711105
0.57778823 0.5777652 0.6049461 0.5809651 0.57757777 0.5777028
0.58331954 0.57804024 0.5784547 0.5791342 0.58031636 0.6046847
0.60520047 0.5937593 0.58349776 0.58918655 0.585288 0.57771075
0.57864183 0.5796051 0.5799217 0.57687336 0.58341265 0.5785788
0.5789991 0.5795324 0.5821006 0.5966433 0.57749677 0.57777035
0.5910414 0.5773269 0.58567226 0.5843944 0.5774361 0.5958311
0.57861346 0.5779085 0.5836908 0.5775305 0.5810351 0.58814144
0.5789568 0.57872987 0.5776423 0.6053786 0.5790132 0.5928653
0.6052447 0.61543584 0.58458525 0.5783247 0.5830784 0.5811673
0.57899153 0.5809654 0.5809654 0.58994776 0.6330279 0.61973125
0.58105934 0.6396998 0.5890624 0.5808206 0.63922584 0.61993045
0.652699 0.58934414 0.6063602 0.61180705 0.5771982 0.5777954
0.5781886 0.5786163 0.5782559 0.5784811 0.5793637 0.5890808
0.5830033 0.58249104 0.60447097 0.57843167 0.57877594 0.57799184
0.59080225 0.57873803 0.5795959 0.5849787 0.5797996 0.579682
0.5805923 0.5797924 0.6207965 0.6172974 0.579682 0.5796384
0.5796296 0.5794636 0.58130836 0.61685055 0.581454 0.5855957
0.58034426 0.57981044 0.5810036 0.5794597 0.57951933 0.5829626
0.5786414 0.6041503 0.5790655 0.57868075 0.58829325 0.5786846
0.5817085 0.57936454 0.60755956 0.57873625 0.5784768 0.58369076
0.58381605 0.5836887 0.5858248 0.6171568 0.5904521 0.61717606
0.57917815 0.585089 0.5946311 0.57941866 0.5956347 0.5800357
0.57921237 0.57977897 0.57950985 0.59980136 0.5794211 0.5854362
0.57901037 0.5847996 0.5779387 0.57809323 0.5930451 0.58091354
0.5784779 0.578267 0.58326787 0.5783494 0.57879704 0.57847404
0.6102496 0.6100861 0.57847404 0.5784781 0.5782633 0.57895887
0.57854944 0.5831228 0.5891479 0.5931571 0.5818151 0.5823464
0.5790161 0.5795622 0.57984537 0.578698 0.5784987 0.58045685
0.57848155 0.5787312 0.57871133 0.5899911 0.5964079 0.5974127
0.6086674 0.59137857 0.57906425 0.58017033 0.62536037 0.5906543
0.57994765 0.5836277 0.6060708 0.60615116 0.5792676 0.5819194
0.5815447 0.57876724 0.57835925 0.57756114 0.58080244 0.5824134
0.579684 0.579288 0.59036326 0.5790007 0.57904476 0.58419985
0.5783132 0.57883286 0.6001782 0.57851803 0.57872695 0.5791408
0.5799836 0.57987076 0.5796064 0.5799593 0.57966936 0.59356964
0.58866215 0.5802649 0.58128166 0.58034664 0.5771053 0.5780635
0.57781297 0.58299184 0.5775686 0.5819763 0.57897043 0.5776775
0.5777392 0.5810676 0.57942575 0.5794476 0.5778632 0.59433913
0.5929802 0.6113254 0.58432055 0.5885272 0.5841351 0.57888293
0.58728164 0.57860404 0.5781007 0.57765305 0.5887117 0.584502
0.58222306 0.5774401 0.58222306 0.57812953 0.64316356 0.581735
0.58159864 0.57851267 0.5795053 0.5944862 0.5847298 0.57949877
0.58130103 0.58020645 0.5800007 0.5978961 0.5795877 0.5803144 ]

Hi , thanks for providing code.

Could you please provide the complete dataset? any link or possible to email??

在数据处理过程中，若设置浏览历史数目为30的话，最终得到一个数据结构，每行的格式是
用户id 用户浏览历史标题（一个30×10的二维数组）用户浏览历史实体（一个30×10的二维数组） 新闻标题 新闻实体 Label
有以下几个疑惑：
1.若某一个用户的记录不足30条，代码中似乎使用重复新闻来填充这个30，这样在训练过程中会不会增加误差呢？
2.实验的训练环节采用了将整个训练集分为很多batch，我发现在attention中，会将clicked_words的维度改变为（batch_size*max_click_history,max_title_length），这样做是不是错把一个用户的浏览历史给换成一个batch中的用户的浏览历史了呢？
@hwwang55

Thank you for sharing the code and papers. I've been doing sequence Tagging recently. Can you share the entity embeddings you've trained? thanks.

.

请问怎么对推荐结果进行验证，能不能贴一贴参考代码

def get_neighbors_for_entity(file):
reader = open(file, encoding='utf-8')
entity2neighbor_map = {}
for line in reader:
array = line.strip().split('\t')
if len(array) != 3: # to skip the first line in triple2id.txt
continue
head = int(array[0])
tail = int(array[1])
DKN/data/kg/kg_preprocess.py should the line 30 be"tail = int(array[2])"?

dkn.py line 141

self.l2_loss needs to be claimed as non-trainable, i.e.,
self.l2_loss = tf.Variable(tf.constant(0., dtype=tf.float32), trainable=False)

In the current version, the l2_loss gets negative results since it is also trained. But I found that other model parameters are updated correctly.

def get_local_word2entity(entities): # 格式：entity_id: entity
"""
Given the entities information in one line of the dataset, construct a map from word to entity index
E.g., given entities = 'id_1:Harry Potter;id_2:England', return a map = {'harry':index_of(id_1),
'potter':index_of(id_1), 'england': index_of(id_2)}
:param entities: entities information in one line of the dataset
:return: a local map from word to entity index
"""
local_map = {}

for entity_pair in entities.split(';'):
    entity_id = entity_pair[:entity_pair.index(':')]
    entity_name = entity_pair[entity_pair.index(':') + 1:]

    # remove non-character word and transform words to lower case
    entity_name = PATTERN1.sub(' ', entity_name)
    entity_name = PATTERN2.sub(' ', entity_name).lower()
    # constructing map: word -> entity_index
    for w in entity_name.split(' '):
        entity_index = entity2index[entity_id]  
        local_map[w] = entity_index  # 这里有问题，不同的实体如果有相同的词，就覆盖了之前的entity_index，不知道这里什么意思？

return local_map

问下raw_train.txt中的entity_id跟kg.txt中的entity_id和relation_id都是Microsoft Satori knowledge graph知识图谱中的吗？怎么下载这个知识图谱？

王老师您好：
我有两个疑问
1.请问代码中为什么只有auc值而没有论文中提到的f1值的计算呢？
2.测试集和训练集中的label标签是点击与未点击的区别吗？

Thanks for providing the code.
I have two questions:
1.where is the source of the knowledge graph(kg.txt)? is it like the paper said extract from the bing server?
2. I 've checked other issues and I am wondering whether this is the full version of the dataset?
If it's because of the privacy policy, I totally understand, I just want to make sure a little bit.

Finally, thanks again~
@hwwang55

When I compile transE.cpp, I got a warning
. But I anyway run transE.exe, and see such a result
. Count of epoch is greater than 5000. It's normal? Do i have any runtime error?

想问下执行这步操作g++ transE.cpp -o transE -pthread -O3 -march=native时没有反应，也没有报错，不知道是什么原因。首先用transE执行，没有生成结果文件，也没有报错
(image_text_match_tf_1.15) [user@tianhu-worker-5 Fast-TransX]$ cd transE
(image_text_match_tf_1.15) [user@tianhu-worker-5 transE]$ g++ transE.cpp -o transE -pthread -O3 -march=native
(image_text_match_tf_1.15) [user@tianhu-worker-5 transE]$

然后用transD，生成了一个trnaD文件，但是也没有生成TransD_relation2vec_50.vec文件，请问是什么原因呢？
(image_text_match_tf_1.15) [user@tianhu-worker-5 kg]$ cd Fast-TransX/transD/
(image_text_match_tf_1.15) [user@tianhu-worker-5 transD]$ g++ transD.cpp -o transD -pthread -O3 -march=native
(image_text_match_tf_1.15) [user@tianhu-worker-5 transD]$ ls

你好，TransE代码有一部分实在看不明白，故求教一下
https://user-images.githubusercontent.com/35772225/68124864-49f60200-ff4b-11e9-9a72-a79d77ca5cc6.png
上图中这部分初始化的四个数组rigTail、lefTail、rigHead和lefHead..到底是干什么用的?我在纸上做了下演草，还是看不出所以然..还望王博抽空解疑一下，谢谢

Dr. Wang, thank you so much for your wonderful work which combines KG and recommender system. It seems that the procedure of disambiguation based on entity linking is not implemented in this code. Does the 'kg.txt' represent KG that has been disambiguated?

any plan to convert DKN model to be tf 2.0 compatible?

Thanks for your so great project. When I running TransE.cpp(others),there is a warning that "no return statement in function returning non-void". I read the code and found the function transetrainMode() and train_transe() don't have return statement while their types are void*. Because of this warning, there is no output file(TransE_relation2vec_50.vec). How can I solve this problem?