ArenaTester / app.py
atrytone's picture
Update app.py
1fa2af5
import gradio as gr
import os
import csv
import json
import uuid
import random
import pickle
from langchain.vectorstores import FAISS
from langchain.embeddings import HuggingFaceEmbeddings
from googleapiclient.discovery import build
from google.oauth2 import service_account
USER_ID = uuid.uuid4()
SERVICE_ACCOUNT_JSON = os.environ.get('GOOGLE_SHEET_CREDENTIALS')
creds = service_account.Credentials.from_service_account_info(json.loads(SERVICE_ACCOUNT_JSON))
SPREADSHEET_ID = '1o0iKPxWYKYKEPjqB2YwrTgrLzvGyb9ULj9tnw_cfJb0'
service = build('sheets', 'v4', credentials=creds)
LEFT_MODEL = None
RIGHT_MODEL = None
PROMPT = None
with open("article_list.pkl","rb") as articles:
article_list = tuple(pickle.load(articles))
INDEXES = ["miread_large", "miread_contrastive", "scibert_contrastive"]
MODELS = [
"biodatlab/MIReAD-Neuro-Large",
"biodatlab/MIReAD-Neuro-Contrastive",
"biodatlab/SciBERT-Neuro-Contrastive",
]
model_kwargs = {'device': 'cpu'}
encode_kwargs = {'normalize_embeddings': False}
faiss_embedders = [HuggingFaceEmbeddings(
model_name=name,
model_kwargs=model_kwargs,
encode_kwargs=encode_kwargs) for name in MODELS]
vecdbs = [FAISS.load_local(index_name, faiss_embedder)
for index_name, faiss_embedder in zip(INDEXES, faiss_embedders)]
def get_matchup():
global LEFT_MODEL, RIGHT_MODEL
choices = INDEXES
left, right = random.sample(choices,2)
LEFT_MODEL, RIGHT_MODEL = left, right
return left, right
def get_comp(prompt):
global PROMPT
left, right = get_matchup()
left_output = inference(PROMPT,left)
right_output = inference(PROMPT,right)
return left_output, right_output
def get_article():
return random.choice(article_list)
def send_result(l_output, r_output, prompt, pick):
global PROMPT
global LEFT_MODEL, RIGHT_MODEL
# with open('results.csv','a') as res_file:
# writer = csv.writer(res_file)
# writer.writerow(row)
if (pick=='left'):
pick = LEFT_MODEL
else:
pick = RIGHT_MODEL
row = [USER_ID,PROMPT,LEFT_MODEL,RIGHT_MODEL,pick]
row = [str(x) for x in row]
body = {'values': [row]}
result = service.spreadsheets().values().append(spreadsheetId=SPREADSHEET_ID, range='A1:E1', valueInputOption='RAW', body=body).execute()
print(f"Appended {result['updates']['updatedCells']} cells.")
new_prompt = get_article()
PROMPT = new_prompt
return new_prompt,gr.State.update(value=new_prompt)
def get_matches(query, db_name="miread_contrastive"):
"""
Wrapper to call the similarity search on the required index
"""
matches = vecdbs[INDEXES.index(
db_name)].similarity_search_with_score(query, k=30)
return matches
def inference(query, model="miread_contrastive"):
"""
This function processes information retrieved by the get_matches() function
Returns - Gradio update commands for the authors, abstracts and journals tablular output
"""
matches = get_matches(query, model)
auth_counts = {}
n_table = []
scores = [round(match[1].item(), 3) for match in matches]
min_score = min(scores)
max_score = max(scores)
def normaliser(x): return round(1 - (x-min_score)/max_score, 3)
i = 1
for match in matches:
doc = match[0]
score = round(normaliser(round(match[1].item(), 3)), 3)
title = doc.metadata['title']
author = doc.metadata['authors'][0].title()
date = doc.metadata.get('date', 'None')
link = doc.metadata.get('link', 'None')
# For authors
record = [score,
author,
title,
link,
date]
if auth_counts.get(author, 0) < 2:
n_table.append([i,]+record)
i += 1
if auth_counts.get(author, 0) == 0:
auth_counts[author] = 1
else:
auth_counts[author] += 1
n_output = gr.Dataframe.update(value=n_table[:10], visible=True)
return n_output
with gr.Blocks(theme=gr.themes.Soft()) as demo:
gr.Markdown("# NBDT Recommendation Engine Arena")
gr.Markdown("NBDT Recommendation Engine Arena is a tool designed to compare neuroscience abstract recommendations by our models. \
We will use this data to compare the performance of models and their preference by various neuroscientists.\
Click on the 'Get Comparision' button to run two random models on the displayed prompt. Then use the correct 'Model X is Better' button to give your vote.\
All models were trained on data provided to us by the NBDT Journal.")
article = get_article()
models = gr.State(value=get_matchup())
prompt = gr.State(value=article)
PROMPT = article
abst = gr.Textbox(value = article, label="Abstract", lines=10)
action_btn = gr.Button(value="Get comparison")
with gr.Group():
with gr.Row().style(equal_height=True):
with gr.Column(scale=1):
l_output = gr.Dataframe(
headers=['No.', 'Score', 'Name', 'Title', 'Link', 'Date'],
datatype=['number', 'number', 'str', 'str', 'str', 'str'],
col_count=(6, "fixed"),
wrap=True,
visible=True,
label='Model A',
show_label = True,
overflow_row_behaviour='paginate',
scale=1
)
with gr.Column(scale=1):
r_output = gr.Dataframe(
headers=['No.', 'Score', 'Name', 'Title', 'Link', 'Date'],
datatype=['number', 'number', 'str', 'str', 'str', 'str'],
col_count=(6, "fixed"),
wrap=True,
visible=True,
label='Model B',
show_label = True,
overflow_row_behaviour='paginate',
scale=1
)
with gr.Row().style(equal_height=True):
l_btn = gr.Button(value="Model A is better",scale=1)
r_btn = gr.Button(value="Model B is better",scale=1)
action_btn.click(fn=get_comp,
inputs=[prompt,],
outputs=[l_output, r_output],
api_name="arena")
l_btn.click(fn=lambda x,y,z: send_result(x,y,z,'left'),
inputs=[l_output,r_output,prompt],
outputs=[abst,],
api_name="feedleft")
r_btn.click(fn=lambda x,y,z: send_result(x,y,z,'right'),
inputs=[l_output,r_output,prompt],
outputs=[abst,prompt],
api_name="feedright")
demo.launch(debug=True)