frankwork/fudan_mtl_reviews

TensorFlow implementation of the paper `Adversarial Multi-task Learning for Text Classification`

fudan_mtl_reviews's Issues

missing embed300.trim.npy

作者您好！请问怎么用 google news word2vec. 得到embed300.trim.npy呢

关于bach数目设定的问题

您好，我已经看这份代码很多遍了，今天突然发现一个问题，我主要想知道下面这个循环：

---

for batch in range(82):
train_fetch = [m_train.tensors, m_train.train_ops, merged_train]

---

其中，“82”是如何得到的呢？
谢谢，希望您能和我交流！

命令行传参数无法识别

使用 python src/main.py --build_data True --adv True 这类tensorflow的标准格式传参数，总是报参数的unknown错误，请问正确的命令行形式是怎样设置的？谢谢！

questions of code

Hi, i am confused bypad_id = -1 word_embed[pad_id] = np.zeros([word_dim])
Please tell me why let pad_id = -1? I think these two lines should be removed.

note: this piece of code is in util.py ---> trim_embeddings function

关于main.pyu中的train函数

你好，train函数中
for batch in range(82):
for i in range(n_task):

请问这个batch行是什么含义？acc_loss和all_acc是求的所有样本上的均值是么？
谢谢！

验证集缺失问题

作者您好，我在解压后发现多任务数据集中每个类的数据都缺失了验证集，原论文中是有验证集的。能否请您补充一下呢。万分感谢！

domain_loss的一些问题

您好，在我们的代码中，利用gradient reverse得到对抗损失，但是这个损失一直不变稳定在0.7左右，显然是对抗不起作用，而作者您的实验结果中domain_loss是从0.6下降到0.2，我想问下这是什么问题，对抗训练有什么技巧吗？谢谢

About data missing

Good Afternoon,Author!First, thanks for your contribution to implement the paper,which help me to learn more about the Adversarial Multi-task Learning.During learning the code, I have found that there have missed two documents named embed300.google.npy and google_words.lst in data/pretrained document.Due to missing them, I couldn't run this program,also couldn't learn the right result about it.If it's convient to you, I sincerely hope that you can share the two data on this to let more students learn about it.Thanks, I'm sorry to bother you in week day~

Performance issues in fudan_mtl_reviews/blob/master/src/inputs/util.py(P2)

Hello,I found a performance issue in the definition of read_tfrecord ,
fudan_mtl_reviews/blob/master/src/inputs/util.py,
dataset = dataset.map(parse_func) was called without num_parallel_calls.
I think it will increase the efficiency of your program if you add this.

Here is the documemtation of tensorflow to support this thing.

Looking forward to your reply. Btw, I am very glad to create a PR to fix it if you are too busy.

UnrecognizedFlagError!!

Traceback (most recent call last):
File "src/main.py", line 7, in
from inputs import util
File "/home/young/Downloads/fudan_mtl_reviews-master/src/inputs/util.py", line 50, in
def write_vocab(vocab, vocab_file=FLAGS.vocab_file):
File "/home/young/Downloads/ENTER/envs/deeplearning/lib/python3.5/site-packages/tensorflow/python/platform/flags.py", line 84, in getattr
wrapped(_sys.argv)
File "/home/young/Downloads/ENTER/envs/deeplearning/lib/python3.5/site-packages/absl/flags/_flagvalues.py", line 630, in call
name, value, suggestions=suggestions)
absl.flags._exceptions.UnrecognizedFlagError: Unknown command line flag 'build_data'

any idea how to fix it?

请问多任务学习中的测试集是怎么样工作的？

请问训练出来一个多任务学习文本分类模型后，测试集里面的数据需不需要先确定是那个任务的数据在进行测试；还是一个完全无标签的数据，通过多任务学习模型会先给它确定在那个任务，再进行分类？

FileNotFoundError: [Errno 2] No such file or directory: 'data/generated/vocab.mtl.txt'

请问data/generated/vocab.mtl.txt数据是所有task，分词后的word列表么？

Loss linear combination

Are the coefficients used while summing up the losses fixed? How did you choose them?
In the original paper, they would be alpha_t, where t is the particular task.

Thank you!

关于代码

请问有torch版的代码吗？我尝试将代码修改为torch版但是模型效果一直不好

关于adv+diff与adv的实验结果

你好，两个问题

1. 我重跑三个实验得到的error rate %结果是
   mtl 13.77 / 您的结果 13.75
   mtl +adv 12.73 /您的结果 12.79
   mtl + adv + diff 12.86 /您的结果 12.70
   可以看到加上diff的结果反而略差于不加的。根据您的代码我在mtl_model.py的三种实现：
   mtl : loss = loss_ce + FLAGS.l2_coefloss_l2
   mtl +adv : loss = loss_ce + 0.05 * loss_adv + FLAGS.l2_coef * (loss_l2 + loss_adv_l2)
   mtl + adv +diff : loss = loss_ce + 0.05loss_adv + FLAGS.l2_coef*(loss_l2+loss_adv_l2) + loss_diff
   请问只需要设置这三个目标函数即可么？
   2.由于实验数据没有validation，实际上train的时候得到的best acc也就是在test数据集上得到的acc，两者应该是完全一致的。但我用自己处理的两份数据跑完train之后，重新载入ckpt 进行test, test与train的acc结果总是不一致，找不到原因，可否请您提示下原因？【载入的ckpt文件确实是我每次train后生成的best epoch的结果】
   非常感谢！

I wonder to know why when batch size=512(16 in code), loss is increasing ?

Line 872: Epoch 0 all_batch_num 510  loss 2.18 acc 0.61 0.8349 time 225.83
Line 1418: Epoch 1 all_batch_num 510  loss 2.70 acc 0.75 0.8872 time 224.98
Line 1964: Epoch 2 all_batch_num 510  loss 2.43 acc 0.83 0.9069 time 221.33
Line 2510: Epoch 3 all_batch_num 510  loss 2.96 acc 0.87 0.9104 time 221.77
Line 3056: Epoch 4 all_batch_num 510  loss 3.74 acc 0.88 0.8987 time 216.13
Line 3602: Epoch 5 all_batch_num 510  loss 4.38 acc 0.89 0.9288 time 210.82
Line 4148: Epoch 6 all_batch_num 510  loss 5.18 acc 0.90 0.9268 time 218.51
Line 4694: Epoch 7 all_batch_num 510  loss 6.28 acc 0.90 0.9312 time 210.94
Line 5240: Epoch 8 all_batch_num 510  loss 7.62 acc 0.91 0.9326 time 218.76
Line 5786: Epoch 9 all_batch_num 510  loss 8.98 acc 0.91 0.9245 time 221.23
Line 6332: Epoch 10 all_batch_num 510  loss 10.36 acc 0.91 0.9182 time 212.32
Line 6878: Epoch 11 all_batch_num 510  loss 11.74 acc 0.92 0.9292 time 211.89

typical patterns captured by task-specific layer

Hi, I want to extract the typical patterns captured by task-specific layer, How can I do it?
In the paper :Task-Movie's typical patterns are well-directed,pointless,cut,cheap,infantile
ans Task-Baby: cute, safety, mild, broken, simple
I want to get typical patterns like those,Thanks!

Adv Loss is not supported by the paper???

Hi, I want to know where the adv loss is different from the domain loss??
In another word, the adv loss in the paper "Adversarial Multi-task Learning for Text Classification" has not described clearly. So i want to know what the equation is??

关于single task的训练问题

您好，您的代码中有去除adv和diff的MTL的实现，
但根据代码：
#TODO NOTICE 非对抗时只使用private，即single task训练,并不是所有数据放在一起训练
if self.adv:
feature = tf.concat([conv_out, shared_out], axis=1)
else:
feature = conv_out
实际上MTL中没有用到share structure,请问这样模型与single task有区别么？我只输入一个数据集进行single task的结果与使用MTL的版本跑出来的结果还要差一些【initial lr均为0.01】请问大概是什么原因呢？谢谢！

Performance issue in /src/inputs/util.py (by P3)

Hello! I've found a performance issue in /src/inputs/util.py: dataset.batch(batch_size)(line 229) should be called before dataset.map(parse_func)(line 224), which could make your program more efficient.

Here is the tensorflow document to support it.

Besides, you need to check the function parse_func called in dataset = dataset.map(parse_func) whether to be affected or not to make the changed code work properly. For example, if parse_func needs data with shape (x, y, z) as its input before fix, it would require data with shape (batch_size, x, y, z) after fix.

Looking forward to your reply. Btw, I am very glad to create a PR to fix it if you are too busy.

diff_loss中的转置是否存在问题？

fudan_mtl_reviews/src/models/mtl_model.py

Lines 71 to 72 in 3cb081d

	correlation_matrix = tf.matmul(
	task_feat, shared_feat, transpose_a=True)

task_feat的size是(bsz, feature_size), shared_feat的size也是(bsz, feature_size), 这样乘出来的结果岂不是(feature, bsz) x (bsz, feature) --> (feature, feature)。
但是我个人感觉这个size应该是(bsz, bsz)更合理些？

是我搞错了, task_feat & shared_feat 的feature_size不一定相同。