app.py

import streamlit as st
import transformers as tf
import pandas as pd
from datetime import datetime
from plotly import graph_objects as go

from overview import NQDOverview

import torch

cuda_available = torch.cuda.is_available()
print(f"Is CUDA available: {cuda_available}")
if cuda_available:
    print(f"CUDA device: {torch.cuda.get_device_name(torch.cuda.current_device())}")

# Function to load and cache models
@st.experimental_singleton(show_spinner=False)
def load_model(username, prefix, model_name):
    p = tf.pipeline('text-classification', f'{username}/{prefix}-{model_name}', return_all_scores=True)
    return p

@st.experimental_singleton(show_spinner=False)
def load_pickle(f):
    return pd.read_pickle(f)

def get_results(model, c):
    res = model(c)[0]
    scores = [r['score'] for r in res]
    label = max(range(len(scores)), key=lambda i: scores[i])
    # label = float(res['label'].split('_')[1])
    # scores = res['score']
    return {'label': label, 'scores': scores}

def run_models(model_names, models, c):
    results = {}
    for mn in model_names:
        results[mn] = get_results(models[mn], c)
    return results


st.title('Assess the *QuAL*ity of your feedback')
st.caption(
"""Medical education requires high-quality *written* feedback, 
but evaluating these *supervisor narrative comments* is time-consuming. 
The QuAL score has validity evidence for measuring the quality of short 
comments in  this context. We developed a NLP/ML-powered tool to 
assess written comment quality via the QuAL score with high accuracy. 

*Try it for yourself!*
""")

### Load models
# Specify which models to load 
USERNAME = 'maxspad'
PREFIX = 'nlp-qual'
models_to_load = ['qual', 'q1', 'q2i', 'q3i']
n_models = float(len(models_to_load))
models = {}
# Show a progress bar while models are downloading, 
# then hide it when done
lc_placeholder = st.empty()
loader_container = lc_placeholder.container()
loader_container.caption('Loading models... please wait...')
pbar = loader_container.progress(0.0)
for i, mn in enumerate(models_to_load):
    pbar.progress((i+1.0) / n_models)
    models[mn] = load_model(USERNAME, PREFIX, mn)
lc_placeholder.empty()

### Load example data
examples = load_pickle('test.pkl')

### Process input
ex = examples['comment'].sample(1, random_state=int(datetime.now().timestamp())).tolist()[0]
try:
    ex = ex.strip().replace('_x000D_', '').replace('nan', 'blank')
except:
    ex = 'blank'
if 'comment' not in st.session_state:
    st.session_state['comment'] = ex
with st.form('comment_form'):
    comment = st.text_area('Try a comment:', value=st.session_state['comment'])
    left_col, right_col = st.columns([1,9], gap='medium')
    submitted = left_col.form_submit_button('Submit')
    trying_example = right_col.form_submit_button('Try an example!')
    
    if submitted:
        st.session_state['button_clicked'] = 'submit'
        st.session_state['comment'] = comment
        st.experimental_rerun()
    elif trying_example:
        st.session_state['button_clicked'] = 'example'
        st.session_state['comment'] = ex
        st.experimental_rerun()
    
results = run_models(models_to_load, models, st.session_state['comment'])
# Modify results to sum the QuAL score and to ignore Q3 if Q2 no suggestion
# if results['q2i']['label'] == 1:
    # results['q3i']['label'] = 1 # can't have connection if no suggestion
# results['qual']['label'] = results['q1']['label'] + (not results['q2i']['label']) + (not results['q3i']['label'])

overview = NQDOverview(st, results)
overview.draw()