Hafez: an Interactive Poetry Generation System

This is the code for the whole pipeline and web interface described in Hafez: an Interactive Poetry Generation System.

For Rhyme Generation Code, please find it here For RNN with FSA decoding code, please find it here

Preparation

Unzip the RNN model files in models: Due to the model's large size (more than 1GB), please download here: lyrics.tl.nn.gz and lyrics.tl.topdown.nn.gz

cd models
gunzip lyrics.tl.nn.gz
gunzip lyrics.tl.topdown.nn.gz

Copy the rhyme generation code into the folder sonnet-project-for-server/

cd sonnet-project-for-server
git clone https://github.com/Marjan-GH/Topical_poetry

Follow the instruction to start the server for Rhyme Generation.

Generate a 4-line poem from command line

In this section, we describe how to generate a four line poem from command line.

Follow the instruction to generate related files given any topic. Here, we pre-generate all the related files about topic Mountain, please check example/4line/ folder:

source.txt the source file
poem.fsa the FSA file
encourage.txt the encourange words list
rhyme.txt the rhyme words information. (You don't need this in following steps)

Run the script:

cd example/4line
bash run_standaline.sh

It will print the command it actually call:

<ROOT_DIR>/exec/ZOPH_RNN_GPU_EXPAND --adjacent-repeat-penalty -2.0 --repeat-penalty -3.0 -L 60 -b 50 --legacy-model 1 -k 1 <ROOT_DIR>/example/4line/../..//models/lyrics.tl.nn <ROOT_DIR>/example/4line/temp.txt --fsa <ROOT_DIR>/example/4line/poem.fsa --encourage-list <ROOT_DIR>/example/4line/encourage.txt --encourage-weight 1.0 --dec-ratio 0.0 100.0 --decode-main-data-files <ROOT_DIR>/example/4line/source.txt

After 1-2 minutes (most time spent on loading the model), it will generate the following 4 line poem and print in STDOUT:

as long as you can see that you are grassy !
this land is filled with highest elevation ,
we took a ride across the river valley ,
and now i got to find my own location .

Setup Poem Generation Server

In this section, you'll setup a backend to generate the poems and build a series a webpages as the frontend.

You need a server with GPU and CUDA 8.0 installed.

Start the RNN server:

cd sh
# start the normal server
bash run_server.sh
# start the interactive server
bash run_server_interactive.sh

(Optional) You need to register a google clound datastore to log all the poems generated by the system.
Start the web server: If you didn't follow step 2, you may need to comments the related lines in py/client.py. (Happy Debugging !) Change the host in py/client.py to reflect your host address, Then run:

cd sh
bash run_client.sh

Go to folder jekyll/poem/data/js/, for each javascript file, locate the variable api_host and change to your host server address. By default, we will use port 8080 for all our API call.
Build the webpage:

cd jekyll/poem
jekyll build

Then the built web pages are in jekyll/poem/_site/

Default Hyper-parameters

For auto mode

The server starts with:

<ROOT_FOLDER>/exec/ZOPH_RNN_GPU_EXPAND --interactive 1 --adjacent-repeat-penalty -2.0 --repeat-penalty -3.0 -b 50 -L 160 --decode-main-data-files <ROOT_FOLDER>/models/source.fake.txt -k 1 <ROOT_FOLDER>/models/lyrics.tl.nn <ROOT_FOLDER>/run/kbest10010.txt --fsa <ROOT_FOLDER>/models/fsa.fake.txt --print-beam 1 --dec-ratio 0.0 100.0 --encourage-list dummy --encourage-weight 1.0 --legacy-model 1

and the client communicate with:

k:1 source_file:<ROOT_FOLDER>/fsas/source.txt fsa_file:<ROOT_FOLDER>/fsas/poem.fsa encourage_list_files:<ROOT_FOLDER>/fsas/encourage.txt,<ROOT_FOLDER>/models/curse.txt,<ROOT_FOLDER>/models/mono.txt encourage_weights:1.0,-5.0,-5.0 repetition:0.0 alliteration:0.0 wordlen:0.0

For ensemble mode

The server starts with:

<ROOT_FOLDER>/exec/ZOPH_RNN_GPU_EXPAND --interactive 1 --interactive-line 1 --adjacent-repeat-penalty -2.0 --repeat-penalty -3.0 -b 50 -L 160 --decode-main-data-files <ROOT_FOLDER>/models/source.fake.txt <ROOT_FOLDER>/models/source.fake.txt -k 1 <ROOT_FOLDER>/models/lyrics.tl.topdown.nn <ROOT_FOLDER>/models/lyrics.tl.nn <ROOT_FOLDER>/run/kbest10010.txt --fsa <ROOT_FOLDER>/models/fsa.fake.txt --print-beam 1 --dec-ratio 0.0 100.0 --encourage-list dummy --encourage-weight 1.0 --legacy-model 1

and the client communicate with:

source <source_file>

words_ensemble word11 word12 word13 ___sep___ word21 word22 word23 ___sep___

fsaline <fsa_file> encourage_list_files:<ROOT_FOLDER>/fsas/encourage.txt,<ROOT_FOLDER>/models/curse.txt,<ROOT_FOLDER>/models/mono.txt encourage_weights:1.0,-5.0,-5.0 repetition:0.0 alliteration:0.0 wordlen:0.0

Run by Docker

Preparation:

Install [nvidia_docker](https://github.com/NVIDIA/nvidia-docker)
Start the docker:

sudo docker run --runtime=nvidia -p 4000:4000 -p 8080:8080 shixing19910105/poem:1.0 /bin/bash -c 'cd /home/poem/sh/ && bash run_all_docker.sh'

Then you will see message like the following:

Poem Client loaded
localhost [127.0.0.1] 10010 (?) : Connection refused
/usr/bin/python
localhost [127.0.0.1] 10010 (?) : Connection refused
 * Running on http://0.0.0.0:8080/ (Press CTRL+C to quit)
 * Restarting with stat
 * Debugger is active!
 * Debugger PIN: 229-279-491
localhost [127.0.0.1] 10010 (?) : Connection refused
localhost [127.0.0.1] 10010 (?) : Connection refused
Accept----------------------------source kbest file info-----------------------------
Number of source word tokens: 0
Number of source segment pairs (lines in training file): 0
Longest source segment (after removing long sentences for training): 0
Total word tokens thrown out due to sentence cutoff: 0
-------------------------------------------------------------------------------

Poem Server loaded and listen on localhost:10010
localhost [127.0.0.1] 50001 (?) : Connection refused
localhost [127.0.0.1] 50001 (?) : Connection refused
ready to operate ....
Rhyme Server 1 loaded and listen on localhost:50001
localhost [127.0.0.1] 50002 (?) : Connection refused
localhost [127.0.0.1] 50002 (?) : Connection refused
37842
ready to operate ....
Out of dictionary word!
data/related_words_3029Rhyme Server 2 loaded and listen on localhost:50002
localhost [127.0.0.1] 50003 (?) : Connection refused
localhost [127.0.0.1] 50003 (?) : Connection refused
37851
1644534
ready to operate ....
pid0
17
data/rhyme_file_6806Rhyme Server 3 loaded and listen on localhost:50003
Configuration file: /home/poem/jekyll/poem/_config.yml
            Source: /home/poem/jekyll/poem
       Destination: /home/poem/jekyll/poem/_site
 Incremental build: disabled. Enable with --incremental
      Generating...
                    done in 0.221 seconds.
 Auto-regeneration: enabled for '/home/poem/jekyll/poem'
    Server address: http://0.0.0.0:4000/poem//
  Server running... press ctrl-c to stop.

Open your browser, access the website through http://<YOUR_SERVER_IP>:4000/poem/

Any questions, please contact Xing Shi

shixing/poem