Hi,

I tried to get into the docker image after downloading and found there's nothing in there. see below:

..\ngq>docker run -it 2caa29d6a3b3 /bin/bash
root@3de7bdd7d8ec:/workspace# ls
root@3de7bdd7d8ec:/workspace# exit
exit

I used your command docker pull magic282/pytorch:0.4.0 to pull the image down. I tried to mount into your image and also found empty in the mounted container. Please help!

Hi, thanks for your great work!
I was wondering how to train this model with my own corpus

when i run " wget https://res.qyzhou.me/redistribute.zip",
i got this error:

--11:28:30-- https://res.qyzhou.me/redistribute.zip
=> `redistribute.zip'
正在解析主机 res.qyzhou.me... 104.31.76.139, 104.31.77.139, 2400:cb00:2048:1::681f:4d8b, ...
Connecting to res.qyzhou.me|104.31.76.139|:443... 已连接。
OpenSSL: error:14077410:SSL routines:SSL23_GET_SERVER_HELLO:sslv3 alert handshake failure
无法建立 SSL 连接。

I am done with training the model.Confused with how to generate questions for my own input. Cou;d you please help me with the file that has to be run?

Would it be possible to provide the code or maybe some code snippets for the NQG++ architecture with shared embedding matrix and pre-trained word embeddings?
That would really be a great help!

def forward(self, input, bio, feats, hidden=None):
        .....
        .....
        featsEmb = [self.feat_lut(feat) for feat in feats[0]] 
        featsEmb = torch.cat(featsEmb, dim=-1)

While trying to run this above code I got this error:
RuntimeError: invalid argument 0: Sizes of tensors must match except in dimension 2. Got 6 and 1 in dimension 0

So I used pad_sequence to match the dimension

featsEmb = [self.feat_lut(feat) for feat in feats[0]]


featsEmb = (pad_sequence(featsEmb, batch_first=True))
featsEmb = torch.cat(tuple(featsEmb), dim=-1)

But now I am getting another dimension error:
RuntimeError: invalid argument 0: Sizes of tensors must match except in dimension 2. Got 6 and 64 in dimension 0

I printed and checked the dimensions and all the dimensions are correct.

print("Shape:------------>")
print((len(featsEmb))
        for i in range(len(featsEmb)):
            print("feat[%d]: %d" % (i, len(featsEmb[i]) ))

            for j in range(len(featsEmb[i])):
                print("feat[%d][%d]: %d" % (i,j, len(featsEmb[i][j]) ))
            print()

Shape:------------>
3

feat[0]: 6
feat[0][0]: 64
feat[0][1]: 64
feat[0][2]: 64
feat[0][3]: 64
feat[0][4]: 64
feat[0][5]: 64

feat[1]: 6
feat[1][0]: 64
feat[1][1]: 64
feat[1][2]: 64
feat[1][3]: 64
feat[1][4]: 64
feat[1][5]: 64

feat[2]: 6
feat[2][0]: 64
feat[2][1]: 64
feat[2][2]: 64
feat[2][3]: 64
feat[2][4]: 64
feat[2][5]: 64

I think it is maybe because of different version of Pytorch.
How can I remove this error and concatenate the tensors correctly? I am running the latest version of PyTorch and the code was written in Pytorch v0.4.0

Any help would be greatly appreciated. Thank you in advance.

I'm trying to reproduce the results reported in the paper but am getting a considerably lower BLEU score (by almost 1 BLEU point) on both the dev and test set.

I have run run_squad_qg.sh with the pre-defined parameters.
I used the generated model model_e20.pt and ran translate.py with the following parameters

python translate.py \
       -model "${MODELPATH}/model_e20.pt" \
       -src "${DATAPATH}/test/dev.txt.shuffle.test.source.txt" \
       -bio "${DATAPATH}/test/dev.txt.shuffle.test.bio" \
       -feats "${DATAPATH}/test/dev.txt.shuffle.test.pos" "${DATAPATH}/test/dev.txt.shuffle.test.ner" "${DATAPATH}/test/dev.txt.shuffle.test.case" \
       -tgt "${DATAPATH}/test/dev.txt.shuffle.test.target.txt" \
       -output "${SAVEPATH}/pred.txt" \
       -replace_unk \
       -verbose \
       -n_best 10 \
       -gpu 0

I then ran test.py in PyBLEU as

python3 test.py ../../../../data/redistribute/QG/test/dev.txt.shuffle.test.target.txt ../../../../data/generated/pred.txt

which resulted in a BLEU score of 11.26 which is almost one bleu score lower than the reported NQG+ BLEU score for the test set in the paper (12.18).

Do you have an explanation or an intuition for this considerable discrepancy? Maybe the parameters in run_squad_qg.sh are not those that were used to compute the results for the paper?

And thanks a lot for making this code public :)

Hello,
Firstly thanks for the code and instruction, it really helps me a lot!
I am wondering that how to implement the "NGQ-POS" module in the paper by modifying this repo.
(So the input will be only a paragraph, ner, bio and Case)

Hello,
I was wondering how do you specify the 'pre_word_vecs_enc' for the pretrained word embeddings.
The Readme file did not contain the specific formatting instructions.
Thank you

Hi,
Firstly thanks for the code and instruction
I have trained the model following the given instruction but I am unable to test it on real data.

If you can give me instructions on how to test my trained model (i.e, taking a paragraph(source), pos, ner,bio,Case tags) and generate a question as output it would be very help ful.

Thanks