Commit
·
b7137b1
1
Parent(s):
1370c40
Create app.py
Browse files
app.py
ADDED
|
@@ -0,0 +1,108 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import time
|
| 2 |
+
import streamlit as st
|
| 3 |
+
import torch
|
| 4 |
+
import string
|
| 5 |
+
|
| 6 |
+
from transformers import BertTokenizer, BertForMaskedLM
|
| 7 |
+
|
| 8 |
+
@st.cache()
|
| 9 |
+
def load_bert_model(model_name):
|
| 10 |
+
try:
|
| 11 |
+
bert_tokenizer = BertTokenizer.from_pretrained(model_name)
|
| 12 |
+
bert_model = BertForMaskedLM.from_pretrained(model_name).eval()
|
| 13 |
+
return bert_tokenizer,bert_model
|
| 14 |
+
except Exception as e:
|
| 15 |
+
pass
|
| 16 |
+
|
| 17 |
+
|
| 18 |
+
|
| 19 |
+
|
| 20 |
+
def decode(tokenizer, pred_idx, top_clean):
|
| 21 |
+
ignore_tokens = string.punctuation + '[PAD]'
|
| 22 |
+
tokens = []
|
| 23 |
+
for w in pred_idx:
|
| 24 |
+
token = ''.join(tokenizer.decode(w).split())
|
| 25 |
+
if token not in ignore_tokens:
|
| 26 |
+
tokens.append(token.replace('##', ''))
|
| 27 |
+
return '\n'.join(tokens[:top_clean])
|
| 28 |
+
|
| 29 |
+
def encode(tokenizer, text_sentence, add_special_tokens=True):
|
| 30 |
+
text_sentence = text_sentence.replace('<mask>', tokenizer.mask_token)
|
| 31 |
+
# if <mask> is the last token, append a "." so that models dont predict punctuation.
|
| 32 |
+
if tokenizer.mask_token == text_sentence.split()[-1]:
|
| 33 |
+
text_sentence += ' .'
|
| 34 |
+
|
| 35 |
+
input_ids = torch.tensor([tokenizer.encode(text_sentence, add_special_tokens=add_special_tokens)])
|
| 36 |
+
mask_idx = torch.where(input_ids == tokenizer.mask_token_id)[1].tolist()[0]
|
| 37 |
+
return input_ids, mask_idx
|
| 38 |
+
|
| 39 |
+
def get_all_predictions(text_sentence, top_clean=5):
|
| 40 |
+
# ========================= BERT =================================
|
| 41 |
+
input_ids, mask_idx = encode(bert_tokenizer, text_sentence)
|
| 42 |
+
with torch.no_grad():
|
| 43 |
+
predict = bert_model(input_ids)[0]
|
| 44 |
+
bert = decode(bert_tokenizer, predict[0, mask_idx, :].topk(top_k).indices.tolist(), top_clean)
|
| 45 |
+
return {'bert': bert}
|
| 46 |
+
|
| 47 |
+
def get_bert_prediction(input_text,top_k):
|
| 48 |
+
try:
|
| 49 |
+
input_text += ' <mask>'
|
| 50 |
+
res = get_all_predictions(input_text, top_clean=int(top_k))
|
| 51 |
+
return res
|
| 52 |
+
except Exception as error:
|
| 53 |
+
pass
|
| 54 |
+
|
| 55 |
+
try:
|
| 56 |
+
|
| 57 |
+
st.title("Qualitative evaluation of Pretrained BERT models")
|
| 58 |
+
st.markdown("""
|
| 59 |
+
<a href="https://ajitrajasekharan.github.io/2021/01/02/my-first-post.html"><small style="font-size:18px; color: #8f8f8f">This app is used to qualitatively examine the performance of pretrained models to do NER , <b>with no fine tuning</b></small></a>
|
| 60 |
+
""", unsafe_allow_html=True)
|
| 61 |
+
st.write("Incomplete. Work in progress...")
|
| 62 |
+
#st.write("https://ajitrajasekharan.github.io/2021/01/02/my-first-post.html")
|
| 63 |
+
st.write("CLS vectors as well as the model prediction for a blank position are examined")
|
| 64 |
+
|
| 65 |
+
top_k = 10
|
| 66 |
+
print(top_k)
|
| 67 |
+
|
| 68 |
+
|
| 69 |
+
bert_tokenizer, bert_model = load_bert_model('ajitrajasekharan/biomedical')
|
| 70 |
+
default_text = "Imatinib is used to treat"
|
| 71 |
+
|
| 72 |
+
|
| 73 |
+
input_text = st.text_area(
|
| 74 |
+
label="Original text",
|
| 75 |
+
value=default_text,
|
| 76 |
+
)
|
| 77 |
+
|
| 78 |
+
start = None
|
| 79 |
+
if st.button("Submit"):
|
| 80 |
+
start = time.time()
|
| 81 |
+
with st.spinner("Computing"):
|
| 82 |
+
|
| 83 |
+
|
| 84 |
+
|
| 85 |
+
|
| 86 |
+
try:
|
| 87 |
+
res = get_bert_prediction(default_text,top_k)
|
| 88 |
+
|
| 89 |
+
|
| 90 |
+
st.header("JSON:")
|
| 91 |
+
|
| 92 |
+
st.json(res)
|
| 93 |
+
|
| 94 |
+
except Exception as e:
|
| 95 |
+
st.error("Some error occured!" + str(e))
|
| 96 |
+
st.stop()
|
| 97 |
+
|
| 98 |
+
st.write("---")
|
| 99 |
+
|
| 100 |
+
|
| 101 |
+
|
| 102 |
+
if start is not None:
|
| 103 |
+
st.text(f"prediction took {time.time() - start:.2f}s")
|
| 104 |
+
|
| 105 |
+
except Exception as e:
|
| 106 |
+
print("SOME PROBLEM OCCURED")
|
| 107 |
+
|
| 108 |
+
|