jackyhawk

should it be "int align = kAlign * aux_size;" => "int align = kAlign;" for ftrl in ffm_score.cc line 57

2

should it be "int align = kAlign * aux_size;" => "int align = kAlign;" for **ftrl in ffm_score.cc line 57** as for "for (index_t d = 0; d < align0;...

param_state_dict 里边的 "item_emb.weight" 是用索引（0,1,2，……）对所有的 特征进行编码（编号）的，对吗？

4

PaddleRec\tools\utils 里边的 save_load.py这个文件中， param_state_dict 里边的 "item_emb.weight" 是用索引对所有的 特征进行编码（编号）的，对吗？ 所有特征的顺序-->对于特征的名称，这个在哪能取到？多谢 def load_model(model_path, net, prefix='rec'): logger.info("start load model from {}".format(model_path)) model_prefix = os.path.join(model_path, prefix) **param_state_dict = paddle.load(model_prefix + ".pdparams")** net.set_dict(param_state_dict)

as for the recall ensfm , can you help to give some information about the format of the ensfm training dataset?

6

as for the recall ensfm , can you help to give some information about the format of the ensfm training dataset? for the two part of the data,what does each...

对于tensornet，是否支持一个slot下多个feature，需要对该slot下的多个feature 求sum/mean pooling？

1

对于tensornet，是否支持一个slot下多个feature，需要对该slot下的多个feature 求sum/mean pooling？ 比如一个itemId，可能有两个tag，tag_科比，tag_库里，需要在tag这个slot下的多个tag，做下pooling，作为这个slot（field）的输入 一个用户的tag兴趣，可能有 tag_湖人，tag_科比，tag_勇士，tag_库里，需要把需要在tag兴趣这个slot下的多个tag兴趣，做下pooling，作为这个slot（field）的输入 还是说，直接用TensorFlow自带的即可？ combiner_embed = tf.nn.embedding_lookup_sparse(self.embedding,sp_ids=idx, sp_weights=val, combiner='sum')

can you help to give some information about the format of the training dataset? for the two part of the data,what does each stand for? the first part of each...

as for 2 version ffm(ftrl), which is better or maybe one is not right implemented

1

https://github.com/superclocks/ffm/blob/master/libffm-ftrl-1.13/ffm.cpp ffm_long align0 = (ffm_long)model.k*3; **vs https://github.com/CNevd/libffm-ftrl/blob/master/ffm.cpp** _ffm_long align0 = (ffm_long)model.k*2;_ or maybe one is not right implemented? thanks very much

data set is too big (which is too big to be held in one machine's mem), and I should break it to small daily set

2

Thanks for the excellent code. and I met one question, my data set is too big (which can not be held in one machine's mem), and I should break it...