arvind385801 / paraphrasegen Goto Github PK

Hi!
Would you give us more details on the way you prepare the outputs of the test to measure the metrics. I'm using MULTEVAL and I'm getting this results for Quora 50k (METEOR is slightly different and TER is different in ~13 units
n=5 BLEU (s_sel/s_opt/p) METEOR (s_sel/s_opt/p) TER (s_sel/s_opt/p)
baseline 17.6 (0.4/0.1/-) 19.6 (0.2/0.0/-) 74.4 (0.4/0.3/-)

I am very interested in your work and want to follow your research. I read the code and have some questions. I would be very grateful if you can get your advice.

1. The code needs two embeddings, one for the original sentences and the other for the generated sentences, why? If we use the same embedding, how will it affect the performance?
2. In the line 156 in the rvae.py, you set loss = 79 * cross_entropy + kld_coef(i) * kld. And for kld_coef(i), the function defined two constants, 3500 and 1000. What factors are taken into account to set these parameters?
3. Are the training document 'train.txt' and the test document 'test.txt' the same to the documents used in the paper (50K train data and 4K test data)? Can you give me 100K and 150K corpus?

Hello,

Is there an issue with the beam_decode function in rvae.py? I think the output of decoder be converted to log_softmax instead of softmax out = F.log_softmax(self.decoder.fc(dec_out)).unsqueeze(0) in sample_decode function.

Thanks,

Jie

Hello！
When I was training the model, the program reported an error at the 1560th iteration, causing the operation to fail. I want to kown the reason. Thanks

I am very interested in your work and want to follow your research. I read the code and have some questions. I would be very grateful if you can get your advice.

1. The code needs two embeddings, one for the original sentences and the other for the generated sentences, why? If we use the same embedding, how will it affect the performance?
2. In the line 156 in the rvae.py, you set loss = 79 * cross_entropy + kld_coef(i) * kld. And for kld_coef(i), the function defined two constants, 3500 and 1000. What factors are taken into account to set these parameters?
3. Are the training document 'train.txt' and the test document 'test.txt' the same to the documents used in the paper (50K train data and 4K test data)? Can you give me 100K and 150K corpus?

Hi, when i ran train.py it returns this error

[root @ ~]$ python2.7 train.py --num-iterations 140000
preprocessed data was found and loaded
preprocessed data was found and loaded
Traceback (most recent call last):
  File "train.py", line 102, in <module>
    cross_entropy, kld, coef = train_step(iteration, args.batch_size, args.use_cuda, args.dropout, start_index)
  File "./paraphraseGen/model/rvae.py", line 124, in train
    input = [Variable(t.from_numpy(var)) for var in input]
TypeError: can't convert np.ndarray of type numpy.object_. The only supported types are: double, float, float16, int64, int32, and uint8.

Pytorch does not support Python2.7 now.