Feliks Zaslavskiy
made default model
a95ffe4
raw
history blame
8.27 kB
import math
import streamlit as st
import pandas as pd
import numpy as np
#import torch
#from transformers import AlbertTokenizer, AlbertModel
from sentence_transformers import SentenceTransformer
from sklearn.metrics.pairwise import cosine_similarity
from io import BytesIO
# base is smaller, vs large
#model_size='base'
#tokenizer = AlbertTokenizer.from_pretrained('albert-' + model_size + '-v2')
#model = AlbertModel.from_pretrained('albert-' + model_size + '-v2')
# For baseline 'sentence-transformers/paraphrase-albert-base-v2'
model_name = 'output/training_OnlineConstrativeLoss-2023-03-14_01-24-44'
model_name = 'output/training_OnlineConstrativeLoss-2023-03-17_23-15-52'
model_name = 'output/training_OnlineConstrativeLoss-2023-03-17_23-50-15'
model_name = 'output/training_OnlineConstrativeLoss-2023-03-25_23-48-35' # had 89% accur on eval set
model_name = 'sentence-transformers/paraphrase-albert-base-v2'
similarity_threshold = 0.9
# for regular burt 0.98
model_sbert = SentenceTransformer(model_name)
def get_sbert_embedding(input_text):
embedding = model_sbert.encode(input_text)
return embedding.tolist()
#def get_embedding(input_text):
# encoded_input = tokenizer(input_text, return_tensors='pt')
# input_ids = encoded_input.input_ids
# #input_num_tokens = input_ids.shape[1]
#
# #print( "Number of input tokens: " + str(input_num_tokens))
# #print("Length of input: " + str(len(input_text)))
#
# list_of_tokens = tokenizer.convert_ids_to_tokens(input_ids.view(-1).tolist())
#
# #print( "Tokens : " + ' '.join(list_of_tokens))
# with torch.no_grad():
#
# outputs = model(**encoded_input)
# last_hidden_states = outputs[0]
# sentence_embedding = torch.mean(last_hidden_states[0], dim=0)
# #sentence_embedding = output.last_hidden_state[0][0]
# return sentence_embedding.tolist()
st.set_page_config(layout="wide")
st.title('Upload the Address Dataset')
st.markdown('Upload an Excel file to view the data in a table.')
uploaded_file = st.file_uploader('Choose a file', type='xlsx')
if uploaded_file is not None:
data_caqh = pd.read_excel(uploaded_file, sheet_name='CAQH', dtype=str)
data_ndb = pd.read_excel(uploaded_file, sheet_name='NDB', dtype=str)
# Data cleaning CAQH
# if you need to format with 00000-0000
# lambda x: x[:5] + '-' + x[5:] if len(x) > 5 and not '-' in x else x
data_caqh['postalcode'] = data_caqh['postalcode'].astype(str).apply(lambda x: x[:5])
data_caqh['full-addr'] = data_caqh['address1'].astype(str) + ', ' \
+ np.where(data_caqh['address2'].isnull(), '' , data_caqh['address2'].astype(str)+ ', ') \
+ data_caqh['city'].astype(str) + ', '\
+ data_caqh['state'].astype(str) + ' ' \
+ data_caqh['postalcode'].astype(str)
data_caqh['full-addr'] = data_caqh['full-addr'].str.upper()
st.write(f"CAQH before duplicate removal {len(data_caqh)}")
data_caqh.drop_duplicates(subset='full-addr',inplace=True)
data_caqh = data_caqh.reset_index(drop=True) # reset the index.
st.write(f"CAQH after duplicate removal {len(data_caqh)}")
# Data cleaning NDB
#data_ndb['zip_pls_4_cd'] = data_ndb['zip_pls_4_cd'].astype(str).apply(lambda x: x if (x[-1] != '0' and x[-1] != '1') else '')
#data_ndb['zip_cd_zip_pls_4_cd'] = data_ndb['zip_cd'].astype(str) +\
# np.where( data_ndb['zip_pls_4_cd'] == '', '', '-' \
# + data_ndb['zip_pls_4_cd'].astype(str))
data_ndb['full-addr'] = data_ndb['adr_ln_1_txt'].astype(str).str.strip() + ', ' \
+ data_ndb['cty_nm'].astype(str).str.strip() + ', ' \
+ data_ndb['st_cd'].astype(str) + ' ' + data_ndb['zip_cd'].astype(str)
data_ndb['full-addr'] = data_ndb['full-addr'].str.upper()
# Calculate similarity For CAQH
num_items = len(data_caqh)
progress_bar = st.progress(0)
total_steps = 100
step_size = math.ceil(num_items / total_steps)
data_caqh['embedding'] = 0
embedding_col_index = data_caqh.columns.get_loc('embedding')
full_addr_col_index = data_caqh.columns.get_loc('full-addr')
for i in range(total_steps):
# Update progress bar
progress = (i + 1) / total_steps
# Process a batch of rows
start = i * step_size
end = start + step_size
stop_iter = False
if end >= num_items:
end = num_items
stop_iter = True
data_caqh.iloc[start:end, embedding_col_index] = data_caqh.iloc[start:end, full_addr_col_index].apply(get_sbert_embedding)
progress_bar.progress(value=progress, text=f"CAQH embeddings: {(i + 1) * step_size} processed out of {num_items}")
if stop_iter:
break
st.write(f"Embeddings for CAQH calculated")
# Calculate similarity For NDB
num_items = len(data_ndb)
progress_bar = st.progress(0)
total_steps = 100
step_size = math.ceil(num_items / total_steps)
data_ndb['embedding'] = 0
embedding_col_index = data_ndb.columns.get_loc('embedding')
full_addr_col_index = data_ndb.columns.get_loc('full-addr')
for i in range(total_steps):
# Update progress bar
progress = (i + 1) / total_steps
# Process a batch of rows
start = i * step_size
end = start + step_size
stop_iter = False
if end >= num_items:
end = num_items
stop_iter = True
# or get_embedding
data_ndb.iloc[start:end, embedding_col_index] = data_ndb.iloc[start:end, full_addr_col_index].apply(get_sbert_embedding)
progress_bar.progress(value=progress, text=f"NDB embeddings: {(i + 1) * step_size} processed out of {num_items}")
if stop_iter:
break
st.write(f"Embeddings for NDB calculated... matching")
progress_bar = st.progress(0)
num_items = len(data_caqh)
for i, row in data_caqh.iterrows():
max_similarity = 0
matched_row = None
for j, ndb_row in data_ndb.iterrows():
sim = cosine_similarity([row['embedding']], [ndb_row['embedding']])
if sim > max_similarity:
max_similarity = sim
matched_row = ndb_row
if max_similarity >= similarity_threshold:
data_caqh.at[i, 'matched-addr'] = matched_row['full-addr']
data_caqh.at[i, 'similarity-score'] = max_similarity
else:
print(f"max similarity was {max_similarity}")
data_caqh.at[i, 'matched-addr'] = 'No Matches'
progress = i / num_items
if progress > 1.0:
progress = 1.0
progress_bar.progress(value=progress, text=f"matching similarities - {i} done out of {num_items}")
# Drop columns not needed for display
data_caqh.drop(columns=['embedding'], inplace=True)
data_ndb.drop(columns=['embedding'], inplace=True)
st.header('CAQH addresses and matches')
st.dataframe(data_caqh, use_container_width=True)
# Calculate stats.
total_items = len(data_caqh)
item_without_matches = data_caqh['matched-addr'].value_counts().get('No Matches', 0)
items_with_matches = total_items - item_without_matches;
percent_matched = (items_with_matches/total_items)*100.0
st.write(f"From total matches {total_items}, {items_with_matches} items matched, {item_without_matches} items did not match, {percent_matched:.2f}% matched")
# Create an in-memory binary stream
output = BytesIO()
# Save the DataFrame to the binary stream as an Excel file
with pd.ExcelWriter(output, engine='xlsxwriter') as writer:
data_caqh.to_excel(writer, sheet_name='Sheet1', index=False)
writer.save()
# Get the binary data from the stream
data = output.getvalue()
# Add a download button for the Excel file
st.download_button(
label='Download CAQH matches as Excel file',
data=data,
file_name='data.xlsx',
mime='application/vnd.openxmlformats-officedocument.spreadsheetml.sheet'
)
st.header('NDB data')
st.dataframe(data_ndb, use_container_width=True)