import time
import streamlit as st
import torch
import string
from transformers import BertTokenizer, BertForMaskedLM
st.set_page_config(page_title='Compare pretrained BERT models qualitatively', page_icon=None, layout='centered', initial_sidebar_state='auto')
@st.cache()
def load_bert_model(model_name):
try:
bert_tokenizer = BertTokenizer.from_pretrained(model_name,do_lower_case
=False)
bert_model = BertForMaskedLM.from_pretrained(model_name).eval()
return bert_tokenizer,bert_model
except Exception as e:
pass
def decode(tokenizer, pred_idx, top_clean):
ignore_tokens = string.punctuation
tokens = []
for w in pred_idx:
token = ''.join(tokenizer.decode(w).split())
if token not in ignore_tokens and len(token) > 1 and not token.startswith('.') and not token.startswith('['):
#tokens.append(token.replace('##', ''))
tokens.append(token)
return '\n'.join(tokens[:top_clean])
def encode(tokenizer, text_sentence, add_special_tokens=True):
text_sentence = text_sentence.replace('<mask>', tokenizer.mask_token)
tokenized_text = tokenizer.tokenize(text_sentence)
input_ids = torch.tensor([tokenizer.encode(text_sentence, add_special_tokens=add_special_tokens)])
if (tokenizer.mask_token in text_sentence.split()):
mask_idx = torch.where(input_ids == tokenizer.mask_token_id)[1].tolist()[0]
else:
mask_idx = 0
return input_ids, mask_idx,tokenized_text
def get_all_predictions(text_sentence, model_name,top_clean=5):
bert_tokenizer = st.session_state['bert_tokenizer']
bert_model = st.session_state['bert_model']
top_k = st.session_state['top_k']
# ========================= BERT =================================
input_ids, mask_idx,tokenized_text = encode(bert_tokenizer, text_sentence)
with torch.no_grad():
predict = bert_model(input_ids)[0]
bert = decode(bert_tokenizer, predict[0, mask_idx, :].topk(top_k*10).indices.tolist(), top_clean)
cls = decode(bert_tokenizer, predict[0, 0, :].topk(top_k*10).indices.tolist(), top_clean)
if ("[MASK]" in text_sentence or "<mask>" in text_sentence):
return {'Input sentence':text_sentence,'Tokenized text': tokenized_text, 'results_count':top_k,'Model':model_name,'Masked position': bert,'[CLS]':cls}
else:
return {'Input sentence':text_sentence,'Tokenized text': tokenized_text,'results_count':top_k,'Model':model_name,'[CLS]':cls}
def get_bert_prediction(input_text,top_k,model_name):
try:
#input_text += ' <mask>'
res = get_all_predictions(input_text,model_name, top_clean=int(top_k))
return res
except Exception as error:
pass
def run_test(sent,top_k,model_name,display_area):
if (st.session_state['bert_tokenizer'] is None):
display_area.text("Loading model:" + st.session_state['model_name'])
st.session_state['bert_tokenizer'], st.session_state['bert_model'] = load_bert_model(st.session_state['model_name'])
display_area.text("Model " + str(st.session_state['model_name']) + " load complete")
try:
display_area.text("Computing fill-mask prediction...")
res = get_bert_prediction(sent,st.session_state['top_k'],st.session_state['model_name'])
display_area.text("Fill-mask prediction complete")
return res
except Exception as e:
st.error("Some error occurred during prediction" + str(e))
st.stop()
return {}
def on_text_change(text,display_area):
return run_test(text,st.session_state['top_k'],st.session_state['model_name'],display_area)
def on_results_count_change():
st.session_state['top_k'] = int(st.session_state.my_slider)
st.info("Results count changed " + str(st.session_state['top_k']))
def on_model_change1():
st.session_state['model_name'] = st.session_state.my_model1
st.info("Pre-selected model chosen: " + st.session_state['model_name'])
st.session_state['bert_tokenizer'], st.session_state['bert_model'] = load_bert_model(st.session_state['model_name'])
def on_model_change2(model_name):
st.session_state['model_name'] = model_name
st.info("Custom model chosen: " + st.session_state['model_name'])
st.session_state['bert_tokenizer'], st.session_state['bert_model'] = load_bert_model(st.session_state['model_name'])
def init_selectbox():
return st.selectbox(
'Choose any of the sentences in pull-down below',
("[MASK] who lives in New York and works for XCorp suffers from Parkinson's", "Lou Gehrig who lives in [MASK] and works for XCorp suffers from Parkinson's","Lou Gehrig who lives in New York and works for [MASK] suffers from Parkinson's","Lou Gehrig who lives in New York and works for XCorp suffers from [MASK]","[MASK] who lives in New York and works for XCorp suffers from Lou Gehrig's", "Parkinson who lives in [MASK] and works for XCorp suffers from Lou Gehrig's","Parkinson who lives in New York and works for [MASK] suffers from Lou Gehrig's","Parkinson who lives in New York and works for XCorp suffers from [MASK]","Lou Gehrig","Parkinson","Lou Gehrigh's is a [MASK]","Parkinson is a [MASK]","New York is a [MASK]","New York","XCorp","XCorp is a [MASK]","acute lymphoblastic leukemia","acute lymphoblastic leukemia is a [MASK]","eGFR is a [MASK]","EGFR is a [MASK]","Trileptal is a [MASK]","no bond or se curity of any kind will be required of any [MASK] of this will","habeas corpus is a [MASK]","modus operandi is a [MASK]","the volunteers were instructed to buy specific systems using our usual [MASK] —anonymously and with cash"),key='my_choice')
def init_session_states():
if 'top_k' not in st.session_state:
st.session_state['top_k'] = 20
if 'bert_tokenizer' not in st.session_state:
st.session_state['bert_tokenizer'] = None
if 'bert_model' not in st.session_state:
st.session_state['bert_model'] = None
if 'model_name' not in st.session_state:
st.session_state['model_name'] = "ajitrajasekharan/biomedical"
def main():
init_session_states()
st.markdown("<h3 style='text-align: center;'>Compare pretrained BERT models qualitatively</h3>", unsafe_allow_html=True)
st.markdown("""
<small style="font-size:20px; color: #2f2f2f"><br/>Why compare pretrained models <b>before fine-tuning</b>?</small><br/><small style="font-size:16px; color: #7f7f7f">Pretrained BERT models can be used as is, <a href="https://ajitrajasekharan.github.io/2021/01/02/my-first-post.html"><b>with no fine tuning to perform tasks like NER.</b><br/></a>This can be done ideally by using both fill-mask and CLS predictions, or just using fill-mask predictions if CLS predictions are poor</small>
""", unsafe_allow_html=True)
st.write("This app can be used to examine both fill-mask predictions as well as the neighborhood of CLS vector")
st.write(" - To examine fill-mask predictions, enter the token [MASK] or <mask> in a sentence")
st.write(" - To examine just the [CLS] vector, enter a word/phrase or sentence. Example: eGFR or EGFR or non small cell lung cancer")
st.write("Pretrained BERT models from three domains (biomedical,PHI [person,location,org, etc.], and legal) are listed on the left. Their performance on domain specific sentences reveal both their strength and weakness.")
st.sidebar.slider("Select count of predictions to display", 1 , 50, 20,key='my_slider',on_change=on_results_count_change) #some times it is possible to have less words
try:
st.sidebar.selectbox(label='Select Model to Apply', options=['ajitrajasekharan/biomedical', 'bert-base-cased','bert-large-cased','microsoft/BiomedNLP-PubMedBERT-base-uncased-abstract-fulltext','allenai/scibert_scivocab_cased','dmis-lab/biobert-v1.1','nlpaueb/legal-bert-base-uncased'], index=0, key = "my_model1",on_change=on_model_change1)
#if (st.session_state['bert_tokenizer'] is None):
# st.session_state['bert_tokenizer'], st.session_state['bert_model'] = load_bert_model(st.session_state['model_name'])
with st.form('my_form'):
selected_sentence = init_selectbox()
text_input = st.text_input("Type any sentence below", "",key='my_text')
custom_model_selection = st.text_input("Model not listed on left? Type the model name (**fill-mask BERT models only**)", "",key="my_model2")
submit_button = st.form_submit_button('Submit')
input_status_area = st.empty()
display_area = st.empty()
display_area.text("Current selections: results count = " + str(st.session_state['top_k']) + " ; " + " Model name = " + st.session_state['model_name'])
if submit_button:
start = time.time()
if (len(text_input) == 0):
text_input = selected_sentence
if (len(custom_model_selection) != 0):
on_model_change2(custom_model_selection)
input_status_area.text("Input sentence: " + text_input)
results = on_text_change(text_input,display_area)
display_area.empty()
with display_area.container():
st.text(f"prediction took {time.time() - start:.2f}s")
st.json(results)
except Exception as e:
st.error("Some error occurred during loading" + str(e))
st.stop()
if __name__ == "__main__":
main()