dir-vhred's Introduction

Dirichlet Latent Variable Hierarchical Recurrent Encoder-Decoder in dialogue generation（EMNLP2019）

Description

This repository hosts the Dir-VHRED model for dialogue generation as described by Min and Yisen et al.2019

requirements

pandas==0.20.3
numpy==1.14.0
gensim==3.1.0
spacy==1.9.0
tqdm==4.15.0
nltk==3.2.5
tensorboardX==1.1
torch==0.4

1.Prepare the dataset:

1.download the ubuntu corpus in http://dataset.cs.mcgill.ca/ubuntu-corpus-1.0/ & movie corpus in https://www.cs.cornell.edu/~cristian/Cornell_Movie-Dialogs_Corpus.html

2.split the dataset with train/validation/test rate: 0.8，0.1，0.1

3.run code: python3 cornell_preprocess.py python3 ubuntu_preprocess.py

2.Train the model:

python3 train.py --data=ubuntu --batch_size=40 --eval_batch_size=40 --kl_annealing_iter=100000 --word_drop=0.25 --z_sent_size=3

python3 train.py --data=cornell --batch_size=40 --eval_batch_size=40 --kl_annealing_iter=20000 --word_drop=0.25 --z_sent_size=3

3.Evaluate the model(negative log-likelihood):

python3 eval.py --data=ubuntu --batch_size=40 --eval_batch_size=40 --z_sent_size=3 --checkpoint=xxxx

python3 eval.py --data=cornell --batch_size=40 --eval_batch_size=40 --z_sent_size=3 --checkpoint=xxxx

4.Evaluate the model(word-embedding metric):

python3 eval_embed.py --data=ubuntu --batch_size=40 --eval_batch_size=40 --z_sent_size=3 --checkpoint=xxxx --beam_size=5 --n_sample_step=3 #n_sample_step=1 means 1-turn while n_sample_step=3 means 3-turn responses

python3 eval_embed.py --data=cornell --batch_size=40 --eval_batch_size=40 --z_sent_size=3 --checkpoint=xxxx --beam_size=5 --n_sample-step=3

5.Generate the response:

python3 generate_sentence.py --data=ubuntu --batch_size=40 --kl_annealing_iter=100000 --word_drop=0.25 --z_sent_size=3 --checkpoint=xxx

python3 generate_sentence.py --data=cornell --batch_size=40 --kl_annealing_iter=20000 --word_drop=0.25 --z_sent_size=3 --checkpoint=xxx

dir-vhred's People

Contributors

Stargazers

Watchers

dir-vhred's Issues

Problems about the beam search code

Hi, first of all, thank you so much for releasing this nice codes, especially the beam search.

And I have one question about the beam search. It seems that you implemented the beam search from scratch. But, why don't you use the PriorityQueue module in the python? Is there any special reason for this?

Hope to respond to me. Thank you.

Originally posted by @passing2961 in #1 (comment)

Problems about paper and dataset

Hi, first of all, I really thank you for the open-source codes of this wonderful repo.
I searched all the HRED-based codes on GitHub but nothing can help me.
Your pytorch-based HRED codes seem to help me a lot.
So I really appreciate you about this repo.

But I still have some questions. First, I cannot find your paper on the web resource except the EMNLP 2019 accepted paper page and I also want to get in touch with your ideas.
If you are willing to send your paper to me, I will be very happy. (my email address can be found on my GitHub homepage).

Then, I'm also very curious about the dataset. Can I load other datasets to run your code?
Is there anything I should care about?

Hope to respond to me. Thank you!

The 'mvlgamma' function not consist in torch<=0.4.1

It appear that:

module 'torch' has no attribute 'mvlgamma'.

I found that 'mvlgamma' function added in torch>=1.0.0, so did other function could replace maybe?

Recommend Projects

cloversjtu / dir-vhred Goto Github PK