Feat/knowledge grounding service (#44)

* Create requirements.txt

* feat: add knowledge grounding service

* knowledge grounding service: update runtests.sh

* knowledge grounding service: update configs

* speedup fix

* update kubernetes config file

* chmod +x test.sh

* codestyle fixes

* fix: codestyle

* fix: formatting

* fix: app route respond

* feat: knowledge grounding service to gpu with quest gen

* fix: batch processing

* fix: responses' batch is a list of str

* feat/knowledge_grounding_service: change nltk downloader in dockerfile for safe way

* added service to proxy.yml

* fix: try-except scope and docker-compose wait-hosts

* fix: typo

* fix: empty knowledge handling

Co-authored-by: dilyararimovna <dilyara.rimovna@gmail.com>
Co-authored-by: Денис Кузнецов <kuznetsov.den.p@gmail.com>

cpu.yml

@@ -33,3 +33,6 @@ services:
   question-generator:
     environment:
       CUDA_VISIBLE_DEVICES: ""
+  knowledge-grounding:
+    environment:
+      CUDA_VISIBLE_DEVICES: ""

dev.yml

@@ -410,4 +410,10 @@ services:
       - "./services/question_generator:/src"
     ports:
       - 8079:8079
+  knowledge-grounding:
+    env_file: [.env.dev]
+    volumes:
+      - "./services/knowledge_grounding:/src"
+    ports:
+      - 8083:8083
 version: "3.7"

docker-compose.yml

@@ -15,7 +15,8 @@ services:
         eliza:8047, emotion-skill:8049, dummy-skill-dialog:8052, comet-atomic:8053, meta-script-skill:8054,
         oscar-skill:8055, coronavirus-skill:8061, small-talk-skill:8062, game-cooperative-skill:8068, program-y-wide:8064,
         comet-conceptnet:8065, news-api-skill:8066, short-story-skill:8057, greeting-skill:8070, factoid-qa:8071, kbqa:8072,
-        factoid-classification:8073, spelling-preprocessing:8074, entity-linking:8075, wiki-parser:8077, odqa:8078"
+        factoid-classification:8073, spelling-preprocessing:8074, entity-linking:8075, wiki-parser:8077, odqa:8078,
+        knowledge-grounding:8083"
       WAIT_HOSTS_TIMEOUT: ${WAIT_TIMEOUT:-480}
 
   cobotqa:
@@ -466,5 +467,11 @@ services:
     environment:
       - CUDA_VISIBLE_DEVICES=0
       - DECODING=greedy
+  knowledge-grounding:
+    build:
+      context: ./services/knowledge_grounding/
+    command: gunicorn --workers=1 server:app -b 0.0.0.0:8083
+    environment:
+      - CUDA_VISIBLE_DEVICES=0
 
 version: '3.7'

one-worker.yml

@@ -85,4 +85,6 @@ services:
     command: gunicorn --workers=1 server:app -b 0.0.0.0:8074 --timeout=100
   grounding-skill:
     command: gunicorn --workers=1 server:app -b 0.0.0.0:8080 --timeout=100
+  knowledge_grounding:
+    command: gunicorn --workers=1 server:app -b 0.0.0.0:8083
 version: '3.7'

proxy.yml

@@ -471,4 +471,12 @@ services:
       - PROXY_PASS=10.11.1.1:8079
       - PORT=8079
     command: ["nginx", "-g", "daemon off;"]
+  knowledge-grounding:
+    build:
+      context: dp/proxy/
+      dockerfile: Dockerfile
+    environment:
+      - PROXY_PASS=10.11.1.1:8083
+      - PORT=8083
+    command: ["nginx", "-g", "daemon off;"]
 version: '3.7'

services/knowledge_grounding/Dockerfile

@@ -0,0 +1,58 @@
+FROM pytorch/pytorch:1.5-cuda10.1-cudnn7-runtime
+
+RUN apt-get update && apt-get install -y --allow-unauthenticated wget && rm -rf /var/lib/apt/lists/*
+
+WORKDIR /src
+
+#create and activate venv
+#ENV VIRTUAL_ENV=/parlaivenv
+#RUN python3 -m venv $VIRTUAL_ENV
+#ENV PATH="$VIRTUAL_ENV/bin:$PATH"
+
+# install parlai
+RUN pip install parlai
+
+#create dir for redditgk task
+RUN mkdir -p /opt/conda/lib/python3.7/site-packages/parlai/tasks/redditgk
+#create dir to use default parlai DATAPATH
+RUN mkdir -p /opt/conda/lib/python3.7/site-packages/data
+#create dir for data file for redditgk task
+RUN mkdir -p /opt/conda/lib/python3.7/site-packages/data/redditgk
+#create dir for wow model file
+RUN mkdir -p /opt/conda/lib/python3.7/site-packages/data/models/wizard_of_wikipedia
+#create dir for courier agent
+RUN mkdir -p /opt/conda/lib/python3.7/site-packages/parlai/agents/courier
+#
+
+#wget redditgk scripts from cloud
+#move task files to tasks/redditgk
+RUN wget http://lnsigo.mipt.ru/export/alexaprize_data/parlai_grounding_knowledge/redditgk/__init__.py -q -P /opt/conda/lib/python3.7/site-packages/parlai/tasks/redditgk
+RUN wget http://lnsigo.mipt.ru/export/alexaprize_data/parlai_grounding_knowledge/redditgk/agents.py -q -P /opt/conda/lib/python3.7/site-packages/parlai/tasks/redditgk
+RUN wget http://lnsigo.mipt.ru/export/alexaprize_data/parlai_grounding_knowledge/redditgk/worlds.py -q -P /opt/conda/lib/python3.7/site-packages/parlai/tasks/redditgk
+RUN wget http://lnsigo.mipt.ru/export/alexaprize_data/parlai_grounding_knowledge/redditgk/test.py -q -P /opt/conda/lib/python3.7/site-packages/parlai/tasks/redditgk
+
+#delete old task_list, copy new from cloud
+RUN wget http://lnsigo.mipt.ru/export/alexaprize_data/parlai_grounding_knowledge/task_list.py -q -O /opt/conda/lib/python3.7/site-packages/parlai/tasks/task_list.py
+
+#wget courier agent scripts from cloud
+#move agent files to agents/courier
+RUN wget http://lnsigo.mipt.ru/export/alexaprize_data/parlai_grounding_knowledge/courier/__init__.py -q -P /opt/conda/lib/python3.7/site-packages/parlai/agents/courier
+RUN wget http://lnsigo.mipt.ru/export/alexaprize_data/parlai_grounding_knowledge/courier/courier.py -q -P /opt/conda/lib/python3.7/site-packages/parlai/agents/courier
+
+#unzip jsons to the DATAPATH/redditgk
+RUN wget http://lnsigo.mipt.ru/export/alexaprize_data/parlai_grounding_knowledge/parlai_redditgk_data.tar.gz -q -P /opt/conda/lib/python3.7/site-packages/data/redditgk
+RUN tar -xvzf /opt/conda/lib/python3.7/site-packages/data/redditgk/parlai_redditgk_data.tar.gz -C /opt/conda/lib/python3.7/site-packages/data/redditgk
+#RUN rm parlai_redditgk_data.tar.gz
+
+#get wow model tar.gz
+RUN wget http://lnsigo.mipt.ru/export/alexaprize_data/parlai_grounding_knowledge/end2end_generator_0.tar.gz -q -P /opt/conda/lib/python3.7/site-packages/data/models/wizard_of_wikipedia
+
+WORKDIR /src
+
+COPY ./requirements.txt /src/requirements.txt
+RUN pip install -r /src/requirements.txt
+RUN python -c "import nltk; nltk.download('punkt')"
+
+COPY . /src
+
+CMD gunicorn --workers=1 server:app -b 0.0.0.0:8083

services/knowledge_grounding/requirements.txt

@@ -0,0 +1,4 @@
+flask==1.1.1
+gunicorn==19.9.0
+requests==2.22.0
+sentry-sdk[flask]==0.14.1

services/knowledge_grounding/server.py

@@ -0,0 +1,145 @@
+import logging
+import os
+import random
+import time
+
+import sentry_sdk
+import torch
+from flask import Flask, request, jsonify
+from parlai.core.params import ParlaiParser
+from parlai.core.agents import create_agent
+from parlai.core.worlds import create_task
+from parlai.core.script import ParlaiScript, register_script
+from parlai.agents.courier.courier import CourierAgent
+from sentry_sdk.integrations.flask import FlaskIntegration
+
+sentry_sdk.init(dsn=os.getenv('SENTRY_DSN'), integrations=[FlaskIntegration()])
+
+
+logging.basicConfig(format='%(asctime)s - %(name)s - %(levelname)s - %(message)s',
+                    level=logging.INFO)
+logger = logging.getLogger(__name__)
+
+
+cuda = torch.cuda.is_available()
+if cuda:
+    torch.cuda.set_device(0)  # singe gpu
+    device = torch.device('cuda')
+else:
+    device = torch.device('cpu')
+
+logger.info(f'knowledge grounding is set to run on {device}')
+
+
+logger.info('knowledge grounding script is preparing...')
+
+
+@register_script('get model response')
+class GetModelResponse(ParlaiScript):
+    @classmethod
+    def setup_args(cls):
+        parser = ParlaiParser(True, True, 'Get response from model in knowledge grounded conversation')
+        parser.add_argument(
+            '-it',
+            '--interactive-task',
+            type='bool',
+            default=True,
+            help='Create interactive version of task',
+        )
+        parser.add_argument(
+            '--user-input-topic',
+            type=str,
+            default='',
+            help='User input topic',
+        )
+        parser.add_argument(
+            '--user-input-knowledge',
+            type=str,
+            default='',
+            help='User input knowledge',
+        )
+        parser.add_argument(
+            '--user-input-text',
+            type=str,
+            default='',
+            help='User input text',
+        )
+        parser.add_argument(
+            '--user-input-history',
+            type=str,
+            default='',
+            help='User input history',
+        )
+        parser.set_defaults(interactive_mode=True, task='interactive')
+        return parser
+
+    def run(self):
+        opt = self.opt
+        if isinstance(self.opt, ParlaiParser):
+            logging.error('opt should be passed, not Parser')
+            opt = self.opt.parse_args()
+        # Create model and courier and assign them to the specified task
+        agent = create_agent(opt, requireModelExists=True)
+        courier_agent = CourierAgent(opt)
+        world = create_task(opt, [courier_agent, agent])
+        user_input = {
+            'topic': opt['user_input_topic'],
+            'knowledge': opt['user_input_knowledge'],
+            'text': opt['user_input_text'],
+            'history': opt['user_input_history'].split('\n') if opt['user_input_history'] else ['']
+        }
+        response = world.parley(user_input)
+        courier_agent.finished = True
+        return response['text']
+
+
+try:
+    GetModelResponse.main(
+        task='redditgk',
+        datatype='test',
+        user_input_topic='',
+        user_input_knowledge='.',
+        user_input_text='hi',
+        user_input_history='',
+        split_lines=False,
+        model_file='zoo:wizard_of_wikipedia/end2end_generator/model',
+    )
+except Exception as e:
+    sentry_sdk.capture_exception(e)
+    logger.exception(e)
+
+logger.info(f'knowledge grounding script is ready')
+
+app = Flask(__name__)
+
+
+@app.route("/respond", methods=['POST'])
+def respond():
+    batch = request.json['batch']
+    responses = []
+    random.seed(42)
+    for sample in batch:
+        response = ""
+        st_time = time.time()
+        if sample['knowledge']:
+            try:
+                response = GetModelResponse.main(
+                    task='redditgk',
+                    datatype='test',
+                    user_input_topic=sample['topic'],
+                    user_input_knowledge=sample['knowledge'],
+                    user_input_text=sample['text'],
+                    user_input_history=sample['history'],
+                    split_lines=False,
+                    model_file='zoo:wizard_of_wikipedia/end2end_generator/model',
+                )
+            except Exception as e:
+                sentry_sdk.capture_exception(e)
+                logger.exception(e)
+            logger.info(f'Current sample response: {response}')
+        else:
+            logger.info(f'Sample knowledge is empty, returning empty response')
+        total_time = time.time() - st_time
+        logger.info(f'knowledge grounding: one sample from batch exec time: {total_time:.3f}s')
+        responses.append(response)
+    return jsonify(responses)

services/knowledge_grounding/test.py

@@ -0,0 +1,41 @@
+import requests
+
+
+def test_knowledge_grounding():
+    url = 'http://0.0.0.0:8083/respond'
+
+    topic = "financial endowment"
+    knowledge = "<h1> financial endowment </h1> <h2> <anchor> criticisms </anchor> </h2> <p> officials in charge of " \
+                "the endowments of some universities have been criticized for ' hoarding ' and reinvesting too much " \
+                "of the endowment's income . \ngiven a historical endowment performance of 10 – 11 % , and a payout " \
+                "rate of 5 % , around half of the endowment's income is reinvested . \nroughly 3 % of the " \
+                "reinvestment is used to keep pace with inflation , leaving an inflation-adjusted 2 % annual " \
+                "growth of the endowment . \nof course , many endowments fail to earn 10 – 11 % . \n</p> <p> " \
+                "two arguments against inflation-adjusted endowment growth are : </p> <h3> hoarding money </h3> <p> " \
+                "large endowments have been criticized for ' hoarding ' money . \nmost philanthropies are required " \
+                "by federal law to distribute 5 % of their assets per year , but university endowments are not " \
+                "required to spend anything . \nmany universities with very large endowments would require less " \
+                "than 5 % to pay full tuition for all their students . \nfor example , it has been estimated that " \
+                "if in 2006 all the harvard students had paid the maximum in tuition and fees , it would have " \
+                "amounted to less than $ 300 million . \nin 2007 , if harvard <h3> size </h3> <p> financial " \
+                "endowments range in size depending on the size of the institution and the level of community " \
+                "support . \nat the large end of the spectrum , the total endowment can be over one billion " \
+                "dollars at many leading private universities . \nharvard university has the largest endowment " \
+                "in the world with $ 37.6 billion in assets as of june 30 , 2015 . \neach university typically " \
+                "has numerous endowments , each of which are frequently restricted to funding very specific areas " \
+                "of the university . \nthe most common examples are endowed professorships , and endowed " \
+                "scholarships or fellowships <h3> socially and environmentally responsible investing </h3> <p> " \
+                "many college and university endowments have come under fire in recent years for practices such " \
+                "as investing in fossil fuels , ' land grabs ' in poor countries and high-risk , high-return " \
+                "investment practices that led to the financial crisis . </p>"
+    text = "wow do you know about financial endowment?"
+    history = "hello how are you\n fine just got from work \n me too what do you do for living? \n i am a financist"
+
+    request_data = {'batch': [{'topic': topic, 'knowledge': knowledge, 'text': text, 'history': history}]}
+    result = requests.post(url, json=request_data).json()[0]
+    assert result != '', f'Got empty string as a result'
+    print('Got\n{}\nSuccess'.format(result))
+
+
+if __name__ == '__main__':
+    test_knowledge_grounding()

services/knowledge_grounding/test.sh

@@ -0,0 +1,3 @@
+#!/bin/bash
+
+python test.py

staging.yml

@@ -250,6 +250,10 @@ services:
     image: ${DOCKER_REGISTRY}/grounding-skill:${VERSION}
     env_file:
       - ${ENV_FILE}
+  knowledge-grounding:
+    image: ${DOCKER_REGISTRY}/knowledge-grounding:${VERSION}
+    env_file:
+      - ${ENV_FILE}
 volumes:
   external-logs:
   external-deeppavlov:

test.yml

@@ -29,6 +29,9 @@ services:
   comet-conceptnet:
     environment:
       - CUDA_VISIBLE_DEVICES=6
+  knowledge-grounding:
+    environment:
+      - CUDA_VISIBLE_DEVICES=8
   wiki-parser:
     command: [ "nginx", "-g", "daemon off;" ]
     build:

tests/runtests.sh

@@ -139,7 +139,8 @@ if [[ "$MODE" == "test_skills" || "$MODE" == "all" ]]; then
                      dummy-skill-dialog intent-catcher short-story-skill comet-atomic \
                      comet-conceptnet convers-evaluation-selector emotion-skill game-cooperative-skill \
                      entity-linking odqa convert-reddit question-generator grounding-skill \
-                     cobot-topics cobot-dialogact cobot-convers-evaluator-annotator; do
+                     cobot-topics cobot-dialogact cobot-convers-evaluator-annotator \
+                     knowledge-grounding; do
 
         echo "Run tests for $container"
         dockercompose_cmd exec -T -u $(id -u) $container ./test.sh