InfosysResponsibleAiToolKit's picture
fairness files
38d6a33
"""
Copyright 2024-2025 Infosys Ltd.”
Use of this source code is governed by MIT license that can be found in the LICENSE file or at
MIT license https://opensource.org/licenses/MIT
Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:
The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
"""
import datetime
from io import BytesIO
from fastapi import HTTPException
import numpy as np
import pandas
import openai
from openai import AzureOpenAI
import json
import time
import base64
# from tenacity import retry, wait_random_exponential, stop_after_attempt
import concurrent.futures
import openai
import backoff
import requests
from fairness.constants.llm_constants import PRIMARY_TEMPLATE, CORRECTION_PROMPT_TEMPLATE,SUCCESS_RATE_INFO
from fairness.dao.WorkBench.Tenet import Tenet
from fairness.dao.WorkBench.Batch import Batch
from fairness.dao.WorkBench.Data import Dataset,DataAttributes,DataAttributeValues
from fairness.dao.databaseconnection import DataBase
from fairness.dao.WorkBench.FileStoreDb import FileStoreReportDb
from fairness.dao.WorkBench.report import Report
from fairness.dao.WorkBench.html import Html
import pandas as pd
import matplotlib
matplotlib.use('Agg')
import matplotlib.pyplot as plt
from matplotlib.backends.backend_pdf import PdfPages
from fpdf import FPDF
from PIL import Image
import zipfile
import io
import logging
import seaborn as sns
import uuid
import os
import datetime
import textwrap
timestamp = datetime.datetime.now().strftime('%Y-%m-%d %H:%M:%S')
log=logging.getLogger(__name__)
log.setLevel(logging.INFO)
bias_types=[
{'bias_type': 'race', 'groups': ["white", "black", "asian", "hispanic", "other"]},
{'bias_type': 'gender', 'groups': ['male', 'female', "other"]},
{'bias_type': 'age', 'groups': ['infants', 'toddlers', 'preschoolers', 'children', 'teenagers', 'young_adults', 'adults', 'middle_aged', 'seniors']},
{'bias_type':'disability','groups':['physical_disabilities','sensory_disabilities','intellectual_disabilities','psychiatric_disabilities','learning_disabilities','chronic_health_conditions']},
]
LOCAL_PATH='../output/graphs/representation/'
OUTPUT_FOLDER = "../output/"
SUCCESS_RATE_LOCAL_PATH='../output/graphs/success_rates/'
ZIP_CONTAINER_NAME=os.getenv("ZIP_CONTAINER_NAME")
class FairnessAudit:
def __init__(self):
self.db = DataBase().db
self.fileStore = FileStoreReportDb()
self.batch = Batch()
self.tenet = Tenet()
self.dataset = Dataset()
self.dataAttributes = DataAttributes()
self.dataAttributeValues = DataAttributeValues()
self.report = Report()
self.client=AzureOpenAI(
api_version=os.getenv("OPENAI_API_VERSION"),
azure_endpoint=os.getenv("OPENAI_API_BASE"),
api_key=os.getenv("OPENAI_API_KEY")
)
def get_dataframe(extension,file):
if extension == "csv":
return pandas.read_csv(file)
elif extension=="parquet":
return pandas.read_parquet(file)
elif extension == "feather":
return pandas.read_feather(file)
elif extension == "json":
return pandas.read_json(file)
def get_extension(fileName: str):
if fileName.endswith(".csv"):
return "csv"
elif fileName.endswith(".feather"):
return "feather"
elif fileName.endswith(".parquet"):
return "parquet"
elif fileName.endswith(".json"):
return "json"
@backoff.on_exception(backoff.expo, exception=(openai.RateLimitError,json.decoder.JSONDecodeError), max_tries=10,backoff_log_level=logging.INFO)
def correct_respnse(self,response,errors,input_text):
#Create error string numberd list
try:
model_name=os.getenv("OPENAI_ENGINE_NAME")
log.info("Correction Required, Correcting the response")
errors=[f"{i+1}. {error}" for i,error in enumerate(errors)]
errors_string='\n'.join(errors)
correction_template=CORRECTION_PROMPT_TEMPLATE.format(bias_json_placeholder=json.dumps(bias_types),original_response=json.dumps(response),specific_errors=errors_string,input_text=input_text)
response=self.client.chat.completions.create(
model=model_name,
messages=[
{"role": "user", "content": correction_template},
],
temperature=0.7,
max_tokens=800,
top_p=0.95,
frequency_penalty=0,
presence_penalty=0,
stop=None,
)
generated_report = response.choices[0].message.content
json_string=generated_report[generated_report.find('['): generated_report.rfind(']')+1]
json_string=json_string.replace("\n","").replace("\t","").replace("\r","").strip()
json_response=json.loads(json_string)
return json_response
except json.decoder.JSONDecodeError as e:
response=self.check_response([],input_text,errors=["JSONDecodeError: "+str(e)])
return response['response']
def check_response(self,response,input_text,errors=[]):
log.info("Checking the response for any errors")
log.info(response)
required_fields={
'bias_type':str,
'bias_indicator':str,
'privileged_groups':list,
'unprivileged_groups':list,
'bias_score':int,
'explanation':str
}
#convert the response to lower case
response=[{k.lower():v for k,v in response_dict.items()} for response_dict in response]
if not errors:
for field,expected_type in required_fields.items():
for response_dict in response:
if response_dict['bias_type']!='NA':
if field not in response_dict:
errors.append(f"Response field {field} is missing")
elif not isinstance(response_dict[field],expected_type):
errors.append(f"Response field {field} is not of expected type {expected_type}")
#check if the bias_type is NA
for response_dict in response:
if response_dict["bias_type"]=='NA':
errors.append("Bias Type is NA. Cross check the input text if really no bias is present or if there is any issue in the analysis")
break
for response_dict in response:
if "bias_type" in response_dict:
bias_types_list=[bias['bias_type'] for bias in bias_types]
if response_dict["bias_type"] not in bias_types_list:
if response_dict["bias_type"]!="NA":
errors.append(f"Invalid bias_type {response_dict['bias_type']}. Must be one of the following: {bias_types_list}")
break
elif response_dict["bias_type"]!='NA': #If bias_type is NA, then privileged_groups and unprivileged_groups should be NA as well.
if "privileged_groups" in response_dict:
if response_dict["bias_type"]!="NA":
if not all([group in bias['groups'] for group in response_dict['privileged_groups'] for bias in bias_types]):
errors.append(f"Invalid privileged_groups {response_dict['privileged_groups']}. Must be one of the following: {bias_types[response_dict['bias_type']]['groups']} for the bias_type {response_dict['bias_type']}")
if "unprivileged_groups" in response_dict:
if not all([group in bias['groups'] for group in response_dict['unprivileged_groups'] for bias in bias_types]):
errors.append(f"Invalid unprivileged_groups {response_dict['unprivileged_groups']}. Must be one of the following: {bias_types['groups']} for the bias_type {response_dict['bias_type']}")
if "bias_indicator" in response_dict:
if response_dict['bias_indicator'] not in ['Low', 'Medium', 'High']:
errors.append(f"Invalid bias_indicator {response_dict['bias_indicator']}. Must be one of the following: Low, Medium, High")
if errors:
response=self.correct_respnse(response,errors,input_text)
return {
'valid': len(errors)==0,
'errors': errors,
'response': response,
}
backoff.on_exception(backoff.expo, exception=(openai.RateLimitError,json.decoder.JSONDecodeError), max_tries=10)
def call_gpt(self,prompt_template,text_message,flag=True):
log.info("Analyzing the input text: "+str(text_message))
model=os.getenv("OPENAI_ENGINE_NAME")
try:
response = self.client.chat.completions.create(
model=model,
# engine="gpt-4-turbo",
messages=[
{"role": "system", "content": prompt_template},
{"role": "user", "content": text_message}
],
temperature=0.7,
max_tokens=800,
top_p=0.95,
frequency_penalty=0,
presence_penalty=0,
stop=None,
)
generated_report = response.choices[0].message.content
json_string=generated_report[generated_report.find('['): generated_report.rfind(']')+1]
json_string=json_string.replace("\n","").replace("\t","").replace("\r","").strip()
json_response=json.loads(json_string)
# json_response[0]['bias_type']="Education"
errors=self.check_response(json_response,text_message)
if errors['valid']:
return json_response
else:
return errors['response']
except json.decoder.JSONDecodeError as e:
log.error("JSONDecodeError: "+str(e))
log.error(str(e.doc))
response=self.call_gpt(prompt_template,text_message)
return response
def image_to_pdf(image_paths, output_pdf, label=None):
pdf = FPDF()
pdf.add_page()
timestamp = datetime.datetime.now().strftime('%Y-%m-%d %H:%M:%S')
# Add title if provided
PURPLE = (150, 53, 150)
WHITE = (255, 255, 255)
BLACK = (0, 0, 0)
# Header section
pdf.set_font('Helvetica', 'B', 17)
pdf.set_text_color(*WHITE)
pdf.set_fill_color(*PURPLE)
# Full-width header
pdf.cell(0, 11, 'INFOSYS RESPONSIBLE AI OFFICE',
align='C', fill=True, border=0)
#remove the gap between the header and the content
pdf.set_y(20)
# Add image
# Process each image
for index,image_path in enumerate(image_paths):
# Add a new page
if index!=0:
pdf.set_y(10)
pdf.add_page()
# Open image to get dimensions
img = Image.open(image_path)
img_width, img_height = img.size
# Calculate scaling to fit page width
page_width = pdf.w-20
page_height = pdf.h- pdf.get_y() - 20
# Calculate scaling factor
width_scale = page_width / img_width
height_scale = page_height / img_height
# Use the smaller scale to ensure image fits
scale_factor = min(width_scale, height_scale)
new_width = img_width * scale_factor
new_height = img_height * scale_factor
# Calculate positioning to center the image
x_position = (page_width - new_width) / 2
y_position = (page_height - new_height) / 2
# Add image to PDF
pdf.image(image_path, x=x_position, y=y_position, w=new_width, h=new_height)
# Save PDF
pdf.output(output_pdf)
print(f"PDF created: {output_pdf}")
def bias_type_bar_chart_visualize(df):
try:
times_stamp = datetime.datetime.now().strftime("%Y%m%d_%H%M%S")
pdf_filename = f"bias_analysis_{times_stamp}.pdf"
df['privileged_groups'] = df['privileged_groups'].apply(lambda x: x.replace('[', '').replace(']', '').replace("'", '').split(', ') if isinstance(x, str) else x)
df['unprivileged_groups'] = df['unprivileged_groups'].apply(lambda x: x.replace('[', '').replace(']', '').replace("'", '').split(', ') if isinstance(x, str) else x)
os.makedirs(OUTPUT_FOLDER, exist_ok=True)
# Save graphs as images and embed them in the HTML content
graph_paths = []
bias_type_counts = df['bias_type'].value_counts()
# Create 2x2 grid for the first set of graphs with smaller figure size
fig, axes = plt.subplots(2, 2, figsize=(12, 8)) # Reduced size (12x8 inches)
axes = axes.flatten() # Flatten to make it easier to iterate
# Plot the bias type frequency
bias_type_counts.plot(kind='bar', color='skyblue', ax=axes[0])
axes[0].set_xlabel('Bias Type')
axes[0].set_ylabel('Frequency')
axes[0].set_title('Frequency of Bias Types in Responses')
# Plot privileged groups frequencies for each bias type
for i, bias_type in enumerate(df['bias_type'].unique()):
privileged_flat = pd.Series([item for item in df[df['bias_type'] == bias_type]['privileged_groups'].dropna()])
privileged_flat = pd.Series([item for sublist in privileged_flat for item in sublist])
if not privileged_flat.empty:
privileged_flat.value_counts().plot(kind='bar', color='skyblue', ax=axes[1])
axes[1].set_title(f'Frequency of Privileged Groups for {bias_type}')
axes[1].set_xlabel('Group')
axes[1].set_ylabel('Frequency')
# Plot unprivileged groups frequencies for each bias type
for i, bias_type in enumerate(df['bias_type'].unique()):
unprivileged_flat = pd.Series([item for item in df[df['bias_type'] == bias_type]['unprivileged_groups'].dropna()])
unprivileged_flat = pd.Series([item for sublist in unprivileged_flat for item in sublist])
if not unprivileged_flat.empty:
unprivileged_flat.value_counts().plot(kind='bar', color='skyblue', ax=axes[2])
axes[2].set_title(f'Frequency of Unprivileged Groups for {bias_type}')
axes[2].set_xlabel('Group')
axes[2].set_ylabel('Frequency')
# Bias Score Distribution
sns.histplot(df['bias_score'], color='skyblue', kde=True, ax=axes[3])
axes[3].set_title('Distribution of Bias Scores in Responses')
axes[3].set_xlabel('Bias Score')
axes[3].set_ylabel('Frequency')
# Save the figure with all 4 plots
plt.tight_layout()
times_stamp = datetime.datetime.now().strftime("%Y%m%d_%H%M%S")
graph_path = os.path.join(OUTPUT_FOLDER, f"bias_analysis_4_plots_{times_stamp}.png")
plt.savefig(graph_path)
plt.close()
graph_paths.append(graph_path)
# Privileged vs Unprivileged Groups Comparison (Bar Plot)
privileged_flat = pd.Series([item for sublist in df['privileged_groups'].dropna() for item in sublist])
unprivileged_flat = pd.Series([item for sublist in df['unprivileged_groups'].dropna() for item in sublist])
# Plot privileged groups
fig, ax = plt.subplots(figsize=(6, 4)) # Smaller figure size (6x4 inches)
privileged_flat.value_counts().plot(kind='bar', color='skyblue', ax=ax, alpha=0.7)
ax.set_title('Frequency of Privileged Groups in Responses')
ax.set_xlabel('Group')
ax.set_ylabel('Frequency')
plt.tight_layout()
graph_path = os.path.join(OUTPUT_FOLDER, f'Frequency_of_Privileged_Groups_{times_stamp}.png')
plt.savefig(graph_path)
plt.close()
graph_paths.append(graph_path)
# Plot unprivileged groups
fig, ax = plt.subplots(figsize=(6, 4)) # Smaller figure size (6x4 inches)
unprivileged_flat.value_counts().plot(kind='bar', color='skyblue', ax=ax, alpha=0.7)
ax.set_title('Frequency of Unprivileged Groups in Responses')
ax.set_xlabel('Group')
ax.set_ylabel('Frequency')
plt.tight_layout()
graph_path = os.path.join(OUTPUT_FOLDER, f'Frequency_of_Unprivileged_Groups_{times_stamp}.png')
plt.savefig(graph_path)
plt.close()
graph_paths.append(graph_path)
# Read the image file and encode it in base64
FairnessAudit.image_to_pdf(graph_paths, os.path.join(LOCAL_PATH, pdf_filename))
return pdf_filename
finally:
for graph_path in graph_paths:
os.remove(graph_path)
log.info("Images removed from the local path")
def bias_type_bar_chart_visualize_workbench(df, label):
pdf_filename = 'audit_report_pdf.pdf'
df['privileged_groups'] = df['privileged_groups'].apply(lambda x: x.replace('[', '').replace(']', '').replace("'", '').split(', ') if isinstance(x, str) else x)
df['unprivileged_groups'] = df['unprivileged_groups'].apply(lambda x: x.replace('[', '').replace(']', '').replace("'", '').split(', ') if isinstance(x, str) else x)
pdf = PdfPages(os.path.join(LOCAL_PATH, pdf_filename))
# Generate HTML content
html_content = f"""
<div style='display: flex; justify-content: center; align-items: left; color:white; background-color: #963596; font-size:23px; font-family: sans-serif; border-radius: 10px; position: relative;'>
<h2 style='margin: 0; style=font-family: sans-serif;'>INFOSYS RESPONSIBLE AI OFFICE</h2>
<span style='position:absolute; right:1; font-size:15px; align-self: center; padding: 0 10px;'>{timestamp}</span>
</div>
"""
html_content += f"""
<body>
<h3 style='color:#963596; text-align:left; font-size:19px; font-family: sans-serif;'>FAIRNESS REPORT</h3>
<p style='font-family: sans-serif; font-size:16px;'>{SUCCESS_RATE_INFO}</p>
</body>
"""
html_content += f"""
<div style='width: 50%; font-family: sans-serif;'>
<h3 class="header" style="color:#963596; font-size:19px;"><strong>DATA INFORMATION</strong></h3>
<table>
<tr><td style="font-size:16px; font-family: sans-serif;">Model Output column</td><td>:</td><td style="color: darkgray; font-size:16px; font-family: sans-serif;">{label}</td></tr>
</table>
</div>
"""
os.makedirs(OUTPUT_FOLDER, exist_ok=True)
# Save graphs as images and embed them in the HTML content
graph_paths = []
bias_type_counts = df['bias_type'].value_counts()
# Create 2x2 grid for the first set of graphs with smaller figure size
fig, axes = plt.subplots(2, 2, figsize=(12, 8)) # Reduced size (12x8 inches)
axes = axes.flatten() # Flatten to make it easier to iterate
# Plot the bias type frequency
bias_type_counts.plot(kind='bar', color='skyblue', ax=axes[0])
axes[0].set_xlabel('Bias Type')
axes[0].set_ylabel('Frequency')
axes[0].set_title('Frequency of Bias Types in Responses')
# Plot privileged groups frequencies for each bias type
for i, bias_type in enumerate(df['bias_type'].unique()):
privileged_flat = pd.Series([item for item in df[df['bias_type'] == bias_type]['privileged_groups'].dropna()])
privileged_flat = pd.Series([item for sublist in privileged_flat for item in sublist])
if not privileged_flat.empty:
privileged_flat.value_counts().plot(kind='bar', color='skyblue', ax=axes[1])
axes[1].set_title(f'Frequency of Privileged Groups for {bias_type}')
axes[1].set_xlabel('Group')
axes[1].set_ylabel('Frequency')
# Plot unprivileged groups frequencies for each bias type
for i, bias_type in enumerate(df['bias_type'].unique()):
unprivileged_flat = pd.Series([item for item in df[df['bias_type'] == bias_type]['unprivileged_groups'].dropna()])
unprivileged_flat = pd.Series([item for sublist in unprivileged_flat for item in sublist])
if not unprivileged_flat.empty:
unprivileged_flat.value_counts().plot(kind='bar', color='skyblue', ax=axes[2])
axes[2].set_title(f'Frequency of Unprivileged Groups for {bias_type}')
axes[2].set_xlabel('Group')
axes[2].set_ylabel('Frequency')
# Bias Score Distribution
sns.histplot(df['bias_score'], color='skyblue', kde=True, ax=axes[3])
axes[3].set_title('Distribution of Bias Scores in Responses')
axes[3].set_xlabel('Bias Score')
axes[3].set_ylabel('Frequency')
# Save the figure with all 4 plots
plt.tight_layout()
times_stamp = datetime.datetime.now().strftime("%Y%m%d_%H%M%S")
graph_path = os.path.join(OUTPUT_FOLDER, f"bias_analysis_4_plots_{times_stamp}.png")
plt.savefig(graph_path)
plt.close()
graph_paths.append(graph_path)
# Read the image file and encode it in base64
with open(graph_path, "rb") as image_file:
image_base64 = base64.b64encode(image_file.read()).decode('utf-8')
html_content += f'<img src="data:image/png;base64,{image_base64}" alt="Bias Analysis 4 Plots">'
# Privileged vs Unprivileged Groups Comparison (Bar Plot)
privileged_flat = pd.Series([item for sublist in df['privileged_groups'].dropna() for item in sublist])
unprivileged_flat = pd.Series([item for sublist in df['unprivileged_groups'].dropna() for item in sublist])
# Plot privileged groups
fig, ax = plt.subplots(figsize=(6, 4)) # Smaller figure size (6x4 inches)
privileged_flat.value_counts().plot(kind='bar', color='skyblue', ax=ax, alpha=0.7)
ax.set_title('Frequency of Privileged Groups in Responses')
ax.set_xlabel('Group')
ax.set_ylabel('Frequency')
pdf.savefig(fig)
plt.tight_layout()
graph_path = os.path.join(OUTPUT_FOLDER, f'Frequency_of_Privileged_Groups_{times_stamp}.png')
plt.savefig(graph_path)
plt.close()
graph_paths.append(graph_path)
# Read the image file and encode it in base64
with open(graph_path, "rb") as image_file:
image_base64 = base64.b64encode(image_file.read()).decode('utf-8')
html_content += f'<img src="data:image/png;base64,{image_base64}" alt="Frequency of Privileged Groups">'
# Plot unprivileged groups
fig, ax = plt.subplots(figsize=(6, 4)) # Smaller figure size (6x4 inches)
unprivileged_flat.value_counts().plot(kind='bar', color='skyblue', ax=ax, alpha=0.7)
ax.set_title('Frequency of Unprivileged Groups in Responses')
ax.set_xlabel('Group')
ax.set_ylabel('Frequency')
pdf.savefig(fig)
plt.tight_layout()
graph_path = os.path.join(OUTPUT_FOLDER, f'Frequency_of_Unprivileged_Groups_{times_stamp}.png')
plt.savefig(graph_path)
plt.close()
graph_paths.append(graph_path)
# Read the image file and encode it in base64
with open(graph_path, "rb") as image_file:
image_base64 = base64.b64encode(image_file.read()).decode('utf-8')
html_content += f'<img src="data:image/png;base64,{image_base64}" alt="Frequency of Unprivileged Groups">'
pdf.close()
# Define the HTML file path
html_file_path = os.path.join(OUTPUT_FOLDER, 'report.html')
with open(html_file_path, "w", encoding="utf-8") as html_file:
html_file.write(html_content)
with open(html_file_path, "r", encoding="utf-8") as html_file:
html_data = html_file.read()
# pdfkit.from_string(html_data,"../output/"+pdf_filename)
return html_data
def audit(self,payload):
start_time=time.time()
label=payload['label']
file=payload['file']
extension = FairnessAudit.get_extension(file.filename)
data = FairnessAudit.get_dataframe(extension,file.file)
inputs=data[label].tolist()
#preprocess the inputs
inputs=[input_text.replace('\n','').replace('\t','').replace('\r','').strip() for input_text in inputs]
primary_template=PRIMARY_TEMPLATE
primary_template=primary_template.replace('\n','').replace('\t','').replace('\r','').strip()
prompt=primary_template.format(bias_json_placeholder=json.dumps(bias_types),input_text="{input_text}")
with concurrent.futures.ThreadPoolExecutor(max_workers=5) as executor:
results=list(executor.map(self.call_gpt,[prompt]*len(inputs),inputs))
data['response']=results
data['bias_type']=data['response'].apply(lambda x: x[0]['bias_type'] if isinstance(x, list) and len(x) > 0 else (x if isinstance(x, str) else 'NA'))
data['bias_score']=data['response'].apply(lambda x: x[0]['bias_score'] if isinstance(x, list) and len(x) > 0 else (x if isinstance(x, str) else 'NA'))
data['privileged_groups']=data['response'].apply(lambda x: x[0]['privileged_groups'] if isinstance(x, list) and len(x) > 0 else (x if isinstance(x, str) else 'NA'))
data['unprivileged_groups']=data['response'].apply(lambda x: x[0]['unprivileged_groups'] if isinstance(x, list) and len(x) > 0 else (x if isinstance(x, str) else 'NA'))
data['bias_indicator']=data['response'].apply(lambda x: x[0]['bias_indicator'] if isinstance(x, list) and len(x) > 0 else (x if isinstance(x, str) else 'NA'))
data=data.replace('NA', pd.NA)
csv_name='bias_audit_report_'+str(uuid.uuid4())+'.csv'
data.to_csv(os.path.join(LOCAL_PATH,csv_name))
pdf_filename=FairnessAudit.bias_type_bar_chart_visualize(data)
response={'audit_report_csv':csv_name,'audit_report_pdf':pdf_filename}
end_time=time.time()
total_time=end_time-start_time
log.info("Time taken for the audit: "+str(total_time))
return {'response':response,'time_taken':total_time}
def workbench_audit(self,payload:dict):
try:
start_time=time.time()
if payload['Batch_id'] is None or '':
log.error("Batch Id id missing")
batchId = payload['Batch_id']
self.batch.update(batch_id=batchId, value={"Status": "In-progress"})
tenet_id = self.tenet.find(tenet_name='Fairness')
batch_details = self.batch.find(batch_id=batchId, tenet_id=tenet_id)
datasetId = batch_details['DataId']
dataset_details = self.dataset.find(Dataset_Id=datasetId)
dataset_attribute_ids = self.dataAttributes.find(dataset_attributes=[
'label'])
log.info("Dataset Attribute Ids:"+str(dataset_attribute_ids))
dataset_attribute_values = self.dataAttributeValues.find(
dataset_id=datasetId, dataset_attribute_ids=dataset_attribute_ids, batch_id=batchId)
log.info("Dataset Attribute Values:"+ str(dataset_attribute_values))
fileId = dataset_details["SampleData"]
label = dataset_attribute_values[0]
content=self.fileStore.read_file(fileId)
if content is None:
raise HTTPException(status_code=500, detail="No content received from the POST request")
content=self.fileStore.read_file(fileId)
if content is None:
raise HTTPException(status_code=500, detail="No content received from the POST request")
data = pandas.read_csv(BytesIO(content['data']))
inputs=data[label].tolist()
inputs=[input_text.replace('\n','').replace('\t','').replace('\r','').strip() for input_text in inputs]
primary_template=PRIMARY_TEMPLATE
primary_template=primary_template.replace('\n','').replace('\t','').replace('\r','').strip()
prompt=primary_template.format(bias_json_placeholder=json.dumps(bias_types),input_text="{input_text}")
with concurrent.futures.ThreadPoolExecutor(max_workers=5) as executor:
results=list(executor.map(self.call_gpt,[prompt]*len(inputs),inputs))
data['response']=results
data['bias_type']=data['response'].apply(lambda x: x[0]['bias_type'] if isinstance(x, list) and len(x) > 0 else (x if isinstance(x, str) else 'NA'))
data['bias_score']=data['response'].apply(lambda x: x[0]['bias_score'] if isinstance(x, list) and len(x) > 0 else (x if isinstance(x, str) else 'NA'))
data['privileged_groups']=data['response'].apply(lambda x: x[0]['privileged_groups'] if isinstance(x, list) and len(x) > 0 else (x if isinstance(x, str) else 'NA'))
data['unprivileged_groups']=data['response'].apply(lambda x: x[0]['unprivileged_groups'] if isinstance(x, list) and len(x) > 0 else (x if isinstance(x, str) else 'NA'))
data['bias_indicator']=data['response'].apply(lambda x: x[0]['bias_indicator'] if isinstance(x, list) and len(x) > 0 else (x if isinstance(x, str) else 'NA'))
data=data.replace('NA', pd.NA)
csv_name='bias_audit_report_'+str(uuid.uuid4())+'.csv'
data.to_csv(os.path.join(LOCAL_PATH,csv_name))
html_data=FairnessAudit.bias_type_bar_chart_visualize_workbench(data,label)
tenet_id = self.tenet.find(tenet_name='Fairness')
html_containerName = os.getenv('HTML_CONTAINER_NAME')
htmlFileId = self.fileStore.save_file(file=BytesIO(html_data.encode(
'utf-8')), filename='fairness_successrate.html', contentType='text/html', tenet='Fairness', container_name=html_containerName)
log.info("HtmlFileId:"+ htmlFileId)
HtmlId = time.time()
doc = {
'HtmlId': HtmlId,
'BatchId': batchId,
'TenetId': tenet_id,
'ReportName': 'fairness_successrate.html',
'HtmlFileId': htmlFileId,
'CreatedDateTime': datetime.datetime.now(),
}
Html.create(doc)
url = os.getenv("REPORT_URL")
payload = {"batchId": batchId}
response = requests.request(
"POST", url, data=payload, verify=False).json()
report_id = self.report.find(batch_id=batchId)
print(report_id)
reportId = report_id['ReportFileId']
reportName=report_id['ReportName']
content = self.fileStore.read_file(reportId,os.getenv("PDF_CONTAINER_NAME"))
pdf_name=content['name']+"."+content['extension']
#load csv and pdf and convert to bytes
with open(os.path.join(LOCAL_PATH,csv_name), 'rb') as f:
csv_file = f.read()
# with open(os.path.join(OUTPUT_FOLDER,pdf_filename), 'rb') as f:
# pdf_file = f.read()
zip_buffer=io.BytesIO()
with zipfile.ZipFile(zip_buffer,'w') as zipf:
zipf.writestr(csv_name,csv_file)
zipf.writestr(reportName,content['data'])
zip_buffer.seek(0)
zip_file_bytes=zip_buffer.getvalue()
zip_file_name="audit_report.zip"
zip_fileid=self.fileStore.save_file(file=zip_file_bytes, filename=zip_file_name, contentType="zip", tenet='Fairness', container_name=ZIP_CONTAINER_NAME)
response={'audit_report_id':zip_fileid}
os.remove(os.path.join(LOCAL_PATH,csv_name))
# os.remove(os.path.join(OUTPUT_FOLDER,pdf_filename))
report_document={"ReportId":time.time(),"BatchId":batchId,"ReportFileId":zip_fileid,"TenetId":tenet_id,"ReportName":zip_file_name,"ContentType":"zip","CreatedDateTime":datetime.datetime.now()}
generated=Report.create(report_document)
if not generated:
raise HTTPException(status_code=500, detail="Report Metadata could not be inserted into DB")
updated=self.batch.update(batch_id=batchId, value={"Status": "Completed"})
if not updated:
raise HTTPException(status_code=500, detail="Batch Status could not be updated in DB")
end_time=time.time()
total_time=end_time-start_time
log.info("Time taken for the audit: "+str(total_time))
return {'response':response,'time_taken':total_time}
except Exception as e:
self.batch.update(batch_id=batchId, value={"Status": "Failed"})
raise e
def download_file(filename):
if os.path.exists(os.path.join(LOCAL_PATH,filename)):
return os.path.join(LOCAL_PATH,filename)
else:
raise HTTPException(status_code=404, detail="File not found")