zengjichuan / tmn Goto Github PK

你好，可以完善一下readme嘛，代码一直有bug

你好，源代码中缺少记忆力机制这一部分，请问可以提供一下完整的嘛

你好,大神. 我想问下词向量的来源是？ (是自己训练的还是 有公开的词向量？)。谢谢了

1.通过np.arange(TOPIC_NUM)构造每个topic的索引输入，再重复BATCH次，通过embedding构建主题语义空间。
2.将主题语义空间映射到统一语义空间，至此，wt_emb的形状是[BATCH,TOPIC_NUM,DIM]。
3.将seq_input映射到词嵌入空间，再映射到统一语义空间，x的形状是[BATCH,SEQ_LEN,DIM]
4.通过dot，计算match得分，match的形状是[BATCH,SEQ_LEN,TOPIC_NUM]
5.将ntm生成的represent_mu与match得分add，topic_sum的形状是[BATCH,SEQ_LEN,TOPIC_NUM]
6.加入残差和全连接层，完成一个hop的计算。

跟论文中描述的：通过W生成S和Target矩阵，再计算P和R，有很大出入。

Hi,
I have trained this model use tmn_data.txt in data directory, but the result is 'val acc 0.8059, f1 0.8058', which is 0.851 in your paper. Is there some change about parameters or train set? i have do nothing when training.

The paper is "Topic Memory Networks for Short Text Classification"

This is the last output message when training:

850/850 [==============================] - 0s 277us/step
ntm estimated perplexity upper bound on validation set: 2149.354
No improvement in epoch 436
Epoch 437/800 training cls
106/106 [==============================] - 17s 159ms/step
cls train loss: 0.0082
No improvement in epoch 437 with val acc 0.8059, f1 0.8059
Epoch 438/800 training cls
106/106 [==============================] - 18s 168ms/step
cls train loss: 0.0087
No improvement in epoch 438 with val acc 0.8082, f1 0.8088
Epoch 439/800 training cls
106/106 [==============================] - 17s 161ms/step
cls train loss: 0.0077
No improvement in epoch 439 with val acc 0.7941, f1 0.7956
Epoch 440/800 training cls
106/106 [==============================] - 17s 159ms/step
cls train loss: 0.0090
No improvement in epoch 440 with val acc 0.8047, f1 0.8060
Epoch 441/800 training cls
106/106 [==============================] - 17s 163ms/step
cls train loss: 0.0077
No improvement in epoch 441 with val acc 0.8000, f1 0.7990
Epoch 442/800 training cls
106/106 [==============================] - 17s 161ms/step
cls train loss: 0.0076
No improvement in epoch 442 with val acc 0.8059, f1 0.8058

你好~可以提供TagMyNews完整的数据集嘛

can you please provide the complete information about the training and the testing dataset.

hello， good job！
could you share the complete dataset used in your paper(included training/dev/test set)