Skip to content

Latest commit

 

History

History
165 lines (122 loc) · 5.72 KB

readme.md

File metadata and controls

165 lines (122 loc) · 5.72 KB
Raw

A set of tools for text search over CommonPool (aka DataComp) dataset.

Overall pipeline:

  1. Clone repo and install requirements
  2. Download CommonPool .parquet files (up to 3 TB)
  3. Set up Solr search engine
  4. Index files with Solr
  5. Run server and open search.html file to search

Install requirements

git clone [email protected]:sedol1339/commonpool_search.git
cd commonpool_search
python3.8 -m venv venv
source venv/bin/activate
pip install -r requirements.txt

Downloading CommonPool .parquet files

download_commonpool.py
Argument Description Possible values Default value
--subset CommonPool subset to download. small, medium, large, xlarge xlarge
--folder Folder to download. any path (equals subset name)
--part Optional. If downloading xlarge subset, set this to download only a part of the subset. Part can be from 0 to 7. Value -1 means downloading all parts (3 TB). -1, 0, ..., 7 -1

After downloading, files will be placed into {folder}/hf_cache subfolder, and relative symlinks to parquet files will be created in {folder}. If you downloaded xlarge subset, symlinks are placed in subfolders {folder}/part_{i}. Then you need to move them to the {folder} after finishing donwnloading:

for dir in xlarge/part_*/; do
	for file in $dir*; do
		echo $file
		ln -s $(readlink -f $file) xlarge/${file##*/}
	done
done

rm -r xlarge/part_*

NOTE: if you remove .parquet symlinks from {folder}, you can run download_commonpool.py again - this will create symlinks to hf_cache again without re-downloading the files.

Setting up Solr search engine and importing .csv files

Let your linux username is linuxuser, then run the following commands:

cd ~
wget https://www.apache.org/dyn/closer.lua/solr/solr/9.3.0/solr-9.3.0.tgz?action=download -O solr-9.3.0.tgz
tar xzf solr-9.3.0.tgz solr-9.3.0/bin/install_solr_service.sh --strip-components=2
sudo bash ./install_solr_service.sh solr-9.3.0.tgz -u linuxuser

By default, the script extracts the distribution archive into /opt, configures Solr to write files into /var/solr, and runs Solr as service. If you do not want to start the service immediately, pass the -n option to install script.

The following instruction will allow access to Solr from web. This is a security risk, so better to run this in the internal network.

Firstly, we allow allow cross-origin JS requests. In file /opt/solr/server/solr-webapp/webapp/WEB-INF/web.xml, add the following lines as the first child of tag:

 <filter>
   <filter-name>cross-origin</filter-name>
   <filter-class>org.eclipse.jetty.servlets.CrossOriginFilter</filter-class>
   <init-param>
     <param-name>allowedOrigins</param-name>
     <param-value>*</param-value>
   </init-param>
   <init-param>
     <param-name>allowedMethods</param-name>
     <param-value>GET,POST,OPTIONS,DELETE,PUT,HEAD</param-value>
   </init-param>
   <init-param>
     <param-name>allowedHeaders</param-name>
     <param-value>origin, content-type, accept</param-value>
   </init-param>
 </filter>
 <filter-mapping>
   <filter-name>cross-origin</filter-name>
   <url-pattern>/*</url-pattern>
 </filter-mapping>

Then restart the service:

sudo service solr restart

Append the following lines in the end of the file /etc/default/solr.in.sh:

SOLR_JETTY_HOST="0.0.0.0"
SOLR_JAVA_MEM="-Xms2g -Xmx2g"

(change -Xms -Xmx to the desired maximum memory allocation pool for a Java Virtual Machine)

Move to the commonpool_search folder, then run the following command to Move xmls folder contents to Solr data folder.

cp xmls/* /var/solr/data/ -r

Create commonpool Solr core:

curl "http://localhost:8983/solr/admin/cores?action=CREATE&name=commonpool"

Just to check, the following request should return a json with "response" and "responseHeader" fields:

curl "http://localhost:8983/solr/commonpool/select?defType=edismax&df=Text&q=badger&start=0&rows=10"

Index files with Solr

NOTE: this will require a long time.

Run the following script:

python index_commonpool_to_solr.py --folder xlarge --min_width 500 --min_height 500 \
 --solr_core commonpool --remove_parquet_on_success --shard_size 1

Argument of the index_commonpool_to_solr.py script:

Argument Description Possible values Default value
--folder Folder with .parquet files. any path xlarge
--remove_parquet_on_success Whether to remove .parquet files on successful indexing. true not specified (=false)
--min_width Skip images with width lower than specified. any integer 0
--min_height Skip images with height lower than specified. any integer 0
--solr_path Path to Solr folder, from which bin/post file will be run. any path /opt/solr
--solr_core Name of the Solr core. any string commonpool
--shard_size One post request per {shard_size} .parquet files. Warning: high values may crash Solr. any positive integer 1

At any time, to make POSTed files searchable make COMMIT with the following request:

curl http://localhost:8983/solr/commonpool/update?commit=true

Run server and open search.html file to search

Run the server:

cd commonpool_search
screen -S webserver
python -m http.server 8080

Then press Ctrl-A, Ctrl-D to detach from screen.

Then connect to your server, port 8080, open search.html and do a search. Solr request URLs will appear in the development console in your browser.

Searching for required objects in other datasets

In folder inspector, you can find scripts for text-searching for any object in several existing datasets.