yangxudong / deeplearning Goto Github PK

View Code? Open in Web Editor NEW

1.0K 35.0 379.0 65.51 MB

深度学习相关的模型训练、评估和预测相关代码

Python 96.13% Shell 3.87%

deeplearning's People

Contributors

Stargazers

Watchers

Forkers

yjwcode mufusu mysqlsc chaffeechen chenzk1993 asdlei99 xaxfan ferrero-zhang ichixiao fuyanzhe hhh920406 zhouyonglong chrisgao001 datianshi21 yipeng5 renlang97 lansatiankong youngjt yanmuerbanduweiliang jerrycatleung yxz9087 caibinbupt l732114684 yueyedeai yuyichen09 ningshiqi lihengwannafly say543 redsuncmx dlreseach rongbohui zhengjie123 xiaoqingwang lyj7110226 karlyang2013 litaohz baokunguo utlww hustsheng gccrpm xxqcheers recomddn boxianlai kt8506 andong0323 lxianwei003 yanmovfer gym0569 kyroad gzgd-rtrs-group kevinyang94486 dencechen yu3401 panchaichai cxvista q1161602026 changya1990 gavinzjchao 10183308 dreadlord1984 lzjtt2017 allensmile legendtianjin wuntoguo tonybb9089 qianrenjian bowen92 sr023 ncepuchris liyuanyuan0507 gentletorch lxcheng exitna lufenggui guangyuyan linsongkai001 gaiserchan damonhao zwcdp lovehoroscoper gearchen lwj314 zhanghonglishanzai xuhuihero pkfresher dbaymax lyatdawn tichocan daturah dazecat snowcranestart yanzhelee lilonghua1987 jerryzhong continue7777 liouxy wuliang211 franter666 rongchen89 wushicanasl

deeplearning's Issues

nce_biases 在预测部分如何处理？

您好，请问 nce_loss 中biases 在youtube match 模型中，使用user embedding近邻查找item embedding时，如何把 nce_biases 考虑进来呢，如果不考虑是不是就预测、训练不一致了呢？

ESMM模型预估cvr全为0

hi 请教个问题 ESMM模型预估的ctr看起来正常但是预估cvr全为0 会是什么原因呢？

DIN

请问din模型中的din.data 跟 din_raw.data 是怎么用的呢用户行为序列怎么解析的不太清楚麻烦能讲解一下吗

ESMM的输入

您好，请问一下ESMM模型构架的主任务cvr和辅助任务ctr的输入是一样的吗（所有的曝光样本）？也就是 cvr和ctr是否都是全量样本呢？还是cvr是点击样本，ctr是全量样本。

因为我看那个模型架构图，它是两个网络分别输入进去的，如果两个网络的输入是一样的，为何架构图不画成MMOE那种一个输入进去，共享底层再分成不同的任务，我这里有点疑惑。

请问youtube_match_model.py为何不用原论文的softmax的方式呢？

我看YouTubeNet原本的召回阶段使用的是softmax分类，是因为认为这个应该被理解为多分类问题，可以“同时”选择多个“下个”要播放的视频吗？
https://github.com/yangxudong/deeplearning/blob/master/youtube_match_model/youtube_match_model.py
loss = tf.nn.sigmoid_cross_entropy_with_logits(
labels=labels_one_hot,
logits=logits)

谢谢！

请问一下对于数据中，用户行为序列长度不等，您这边是如何处理的呢？我看到您的raw.data里面就有不等长的，但是在Input_fn里又设置了定长

queation

你好，word_cnn中，自定义模型model_fn的输入，features和labels是怎么传递的呢

练习一下git，请无视我

ESMM cvr预测值接近0

关于预测

你好，看到你用export_savemodel的方式导出模型，想问下predict文件该怎么写呢，是否还要包括feature column那些预处理？

DCN sample data

Hello。

could you please provide some sample data for DCN?
thank you.

大佬请问din模型中的din.data是怎么生成的？是根据raw.data生成的tfrecorder吗？在raw.data中好像没有click字段

ESMM导出模型export_savedmodel之后的数据输入格式

请教大佬，在esmm中export_savedmodel之后得到保存的模型文件，然后想在新的session里load saved_model进行预测，这时候如果用feed_dict方式传入的话
sess.run('CTCVR:0', feed_dict={'input_example_tensor:0': [X]})
这里的X应该是什么格式呢？尝试了直接传入字符串会报如下错误
InvalidArgumentError (see above for traceback): Could not parse example input, value: '{"behaviorBids":[24844,2674,29474,2855,30250,32057,35],"behaviorC1ids":[3,5,591,1230,2],"behaviorCids":[2187,164,713,47,73,78,61],"behaviorPids":[8567830,1098681,3861814,577623,2965309,4019581,8369742],"behaviorSids":[85039235,138228,36774732,118520,46571321,49310857,38369486],"bidWeights":[0.9999832,0.99967366,0.999643,0.9995961,0.999551,0.9890053,0.9889464],"brandPrefer":0,"c1idWeights":[0.9999832,0.99967366,0.999643,0.9995961,0.9890053],"cate2Prefer":0.11423163,"catePrefer":0.04157726,"cidWeights":[0.9999832,0.99967366,0.999643,0.9995961,0.999551,0.9890053,0.9889464],"click":0,"matchScore":0.3556864,"matchType":4,"pay":0,"phoneBrand":"apple","phoneOs":"ios","phoneResolution":"375*812","pidWeights":[0.9999832,0.99967366,0.999643,0.9995961,0.999551,0.9890053,0.9889464],"popScore":0.001,"position":23,"productId":8518322,"sellerPrefer":0,"share":0,"sidWeights":[0.9999832,0.99967366,0.999643,0.9995961,0.999551,0.9890053,0.9889464],"time":1544615913,"triggerNum":1,"triggerRank":10,"type":1,"userId":41233203,"userType":0}' [[node ParseExample/ParseExample ]]