document-vqa-v2 / main.py
MJobe's picture
Update main.py
694da95 verified
import fitz
import io
from fastapi import FastAPI, File, UploadFile, Form, HTTPException
from fastapi.responses import JSONResponse
from transformers import pipeline
from PIL import Image
from io import BytesIO
from starlette.middleware import Middleware
from starlette.middleware.cors import CORSMiddleware
from pdf2image import convert_from_bytes
from pydub import AudioSegment
import numpy as np
import json
import torchaudio
import torch
from pydub import AudioSegment
import speech_recognition as sr
import logging
import asyncio
from concurrent.futures import ThreadPoolExecutor
import re
from pydantic import BaseModel
from typing import List, Dict, Any
app = FastAPI()
# Set up CORS middleware
origins = ["*"] # or specify your list of allowed origins
app.add_middleware(
CORSMiddleware,
allow_origins=origins,
allow_credentials=True,
allow_methods=["*"],
allow_headers=["*"],
)
nlp_qa = pipeline("document-question-answering", model="jinhybr/OCR-DocVQA-Donut")
nlp_qa_v2 = pipeline("document-question-answering", model="faisalraza/layoutlm-invoices", ignore_mismatched_sizes=True)
nlp_qa_v3 = pipeline("question-answering", model="deepset/roberta-base-squad2")
nlp_classification = pipeline("text-classification", model="distilbert/distilbert-base-uncased-finetuned-sst-2-english")
nlp_classification_v2 = pipeline("text-classification", model="cardiffnlp/twitter-roberta-base-sentiment-latest")
nlp_speech_to_text = pipeline("automatic-speech-recognition", model="facebook/wav2vec2-base-960h")
nlp_sequence_classification = pipeline("zero-shot-classification", model="facebook/bart-large-mnli")
nlp_main_classification = pipeline("zero-shot-classification", model="roberta-large-mnli")
description = """
## Image-based Document QA
This API performs document question answering using a LayoutLMv2-based model.
### Endpoints:
- **POST /uploadfile/:** Upload an image file to extract text and answer provided questions.
- **POST /pdfQA/:** Provide a PDF file to extract text and answer provided questions.
"""
app = FastAPI(docs_url="/", description=description)
@app.post("/uploadfile/", description="Upload an image file to extract text and answer provided questions.")
async def perform_document_qa(
file: UploadFile = File(...),
questions: str = Form(...),
):
try:
# Read the uploaded file as bytes
contents = await file.read()
# Open the image using PIL
image = Image.open(BytesIO(contents))
# Perform document question answering for each question using LayoutLMv2-based model
answers_dict = {}
for question in questions.split(','):
result = nlp_qa(
image,
question.strip()
)
# Access the 'answer' key from the first item in the result list
answer = result[0]['answer']
# Format the question as a string without extra characters
formatted_question = question.strip("[]")
answers_dict[formatted_question] = answer
return answers_dict
except Exception as e:
return JSONResponse(content=f"Error processing file: {str(e)}", status_code=500)
@app.post("/uploadfilev2/", description="Upload an image file to extract text and answer provided questions.")
async def perform_document_qa(
file: UploadFile = File(...),
questions: str = Form(...),
):
try:
# Read the uploaded file as bytes
contents = await file.read()
# Open the image using PIL
image = Image.open(BytesIO(contents))
# Perform document question answering for each question using LayoutLMv2-based model
answers_dict = {}
for question in questions.split(','):
result = nlp_qa_v2(
image,
question.strip()
)
# Access the 'answer' key from the first item in the result list
answer = result[0]['answer']
# Format the question as a string without extra characters
formatted_question = question.strip("[]")
answers_dict[formatted_question] = answer
return answers_dict
except Exception as e:
return JSONResponse(content=f"Error processing file: {str(e)}", status_code=500)
@app.post("/uploadfilev3/", description="Upload an image file to extract text and answer provided questions.")
async def perform_document_qa(
context: str = Form(...),
question: str = Form(...),
):
try:
QA_input = {
'question': question,
'context': context
}
res = nlp_qa_v3(QA_input)
return res['answer']
except Exception as e:
return JSONResponse(content=f"Error processing file: {str(e)}", status_code=500)
@app.post("/classify/", description="Classify the provided text.")
async def classify_text(text: str = Form(...)):
try:
# Perform text classification using the pipeline
result = nlp_classification(text)
# Return the classification result
return result
except Exception as e:
return JSONResponse(content=f"Error classifying text: {str(e)}", status_code=500)
@app.post("/test_classify/", description="Classify the provided text with positive, neutral, or negative sentiment.")
async def test_classify_text(text: str = Form(...)):
try:
# Perform text classification using the updated model that returns positive, neutral, or negative
result = nlp_classification_v2(text)
# Print the raw label for debugging purposes (can be removed later)
raw_label = result[0]['label']
print(f"Raw label from model: {raw_label}")
# Map the model labels to human-readable format
label_map = {
"negative": "Negative",
"neutral": "Neutral",
"positive": "Positive"
}
# Get the readable label from the map
formatted_label = label_map.get(raw_label, "Unknown")
return {"label": formatted_label, "score": result[0]['score']}
except Exception as e:
return JSONResponse(content=f"Error classifying text: {str(e)}", status_code=500)
@app.post("/transcribe_and_answer/", description="Transcribe audio and answer provided questions based on the transcription.")
async def transcribe_and_answer(
file: UploadFile = File(...),
questions: str = Form(...)
):
try:
# Ensure correct file format
if file.content_type not in ["audio/wav", "audio/mpeg", "audio/mp3", "audio/webm"]:
raise HTTPException(status_code=400, detail="Unsupported audio format. Please upload a WAV or MP3 file.")
logging.info(f"Received file type: {file.content_type}")
logging.info(f"Received questions: {questions}")
# Convert uploaded file to WAV if needed
audio_data = await file.read()
audio_file = io.BytesIO(audio_data)
if file.content_type in ["audio/mpeg", "audio/mp3"]:
audio = AudioSegment.from_file(audio_file, format="mp3")
audio_wav = io.BytesIO()
audio.export(audio_wav, format="wav")
audio_wav.seek(0)
elif file.content_type == "audio/webm":
audio = AudioSegment.from_file(audio_file, format="webm")
audio_wav = io.BytesIO()
audio.export(audio_wav, format="wav")
audio_wav.seek(0)
else:
audio_wav = audio_file
# Transcription
recognizer = sr.Recognizer()
with sr.AudioFile(audio_wav) as source:
audio = recognizer.record(source)
transcription_text = recognizer.recognize_google(audio)
# Parse questions JSON
try:
questions_dict = json.loads(questions)
except json.JSONDecodeError as e:
raise HTTPException(status_code=400, detail="Invalid JSON format for questions")
# Answer each question
answers_dict = {}
for key, question in questions_dict.items():
QA_input = {
'question': question,
'context': transcription_text
}
# Add error handling here for model-based Q&A
try:
result = nlp_qa_v3(QA_input) # Ensure this is defined or imported correctly
answers_dict[key] = result['answer']
except Exception as e:
logging.error(f"Error in question answering model: {e}")
answers_dict[key] = "Error in answering this question."
# Return transcription + answers
return {
"transcription": transcription_text,
"answers": answers_dict
}
except Exception as e:
logging.error(f"General error: {e}")
raise HTTPException(status_code=500, detail="Internal Server Error")
@app.post("/test-transcription/", description="Upload an audio file to test transcription using speech_recognition.")
async def test_transcription(file: UploadFile = File(...)):
try:
# Check if the file format is supported
if file.content_type not in ["audio/wav", "audio/mpeg", "audio/mp3"]:
raise HTTPException(status_code=400, detail="Unsupported audio format. Please upload a WAV or MP3 file.")
# Convert uploaded file to WAV if necessary for compatibility with SpeechRecognition
audio_data = await file.read()
audio_file = io.BytesIO(audio_data)
if file.content_type in ["audio/mpeg", "audio/mp3"]:
# Convert MP3 to WAV
audio = AudioSegment.from_file(audio_file, format="mp3")
audio_wav = io.BytesIO()
audio.export(audio_wav, format="wav")
audio_wav.seek(0)
else:
audio_wav = audio_file
# Transcribe audio using speech_recognition
recognizer = sr.Recognizer()
with sr.AudioFile(audio_wav) as source:
audio = recognizer.record(source)
transcription = recognizer.recognize_google(audio)
# Return the transcription
return {"transcription": transcription}
except Exception as e:
raise HTTPException(status_code=500, detail=f"Error during transcription: {str(e)}")
# Define the ThreadPoolExecutor globally to manage asynchronous execution
executor = ThreadPoolExecutor(max_workers=10)
# Predefined classifications
labels = [
"All Pricing copy quote requested",
"Change to quote",
"Change to quote & Status Check",
"Change to quote (Items missed?)",
"Confirmation",
"Copy quote requested",
"Cost copy quote requested",
"MRSP copy quote requested",
"MSRP & All Pricing copy quote requested",
"MSRP & Cost copy quote requested",
"No narrative in email",
"Notes not clear",
"Retail copy quote requested",
"Status Check (possibly)"
]
@app.post("/fast_classify/", description="Quickly classify text into predefined categories.")
async def fast_classify_text(statement: str = Form(...)):
try:
# Use run_in_executor to handle the synchronous model call asynchronously
loop = asyncio.get_running_loop()
result = await loop.run_in_executor(
executor,
lambda: nlp_sequence_classification(statement, labels, multi_label=False)
)
# Extract the best label and score
best_label = result["labels"][0]
best_score = result["scores"][0]
return {"classification": best_label, "confidence": best_score}
except asyncio.TimeoutError:
# Handle timeout
return JSONResponse(content="Classification timed out. Try a shorter input or increase timeout.", status_code=504)
except HTTPException as http_exc:
# Handle HTTP errors
return JSONResponse(content=f"HTTP error: {http_exc.detail}", status_code=http_exc.status_code)
except Exception as e:
# Handle general errors
return JSONResponse(content=f"Error in classification pipeline: {str(e)}", status_code=500)
# Predefined classifications
labels = [
"All Pricing copy quote requested",
"Change to quote",
"Change to quote & Status Check",
"Change to quote (Items missed?)",
"Confirmation",
"Copy quote requested",
"Cost copy quote requested",
"MRSP copy quote requested",
"MSRP & All Pricing copy quote requested",
"MSRP & Cost copy quote requested",
"No narrative in email",
"Notes not clear",
"Retail copy quote requested",
"Status Check (possibly)"
]
@app.post("/fast_classify_v2/", description="Quickly classify text into predefined categories.")
async def fast_classify_text(statement: str = Form(...)):
try:
# Use run_in_executor to handle the synchronous model call asynchronously
loop = asyncio.get_running_loop()
result = await loop.run_in_executor(
executor,
lambda: nlp_sequence_classification(statement, labels, multi_label=False)
)
# Extract all labels and their scores
all_labels = result["labels"]
all_scores = result["scores"]
# Extract the best label and score
best_label = all_labels[0]
best_score = all_scores[0]
# Prepare the response
full_response = {
"classification": best_label,
"confidence": best_score,
"all_labels": {label: score for label, score in zip(all_labels, all_scores)}
}
return full_response
except asyncio.TimeoutError:
# Handle timeout
return JSONResponse(content="Classification timed out. Try a shorter input or increase timeout.", status_code=504)
except HTTPException as http_exc:
# Handle HTTP errors
return JSONResponse(content=f"HTTP error: {http_exc.detail}", status_code=http_exc.status_code)
except Exception as e:
# Handle general errors
return JSONResponse(content=f"Error in classification pipeline: {str(e)}", status_code=500)
# Labels for main classifications
main_labels = [
"Change to quote",
"Copy quote requested",
"Expired Quote",
"Notes not clear"
]
# Define a model for the response
class ClassificationResponse(BaseModel):
classification: str
sub_classification: str
confidence: float
scores: Dict[str, float]
# Keyword dictionaries for overriding classifications
change_to_quote_keywords = ["Per ATP", "Add", "Revised", "Remove", "Advise"]
copy_quote_requested_keywords = ["MSRP", "Send Quote", "Copy", "All pricing", "Retail"]
sub_classification_keywords = {
"MRSP": ["MSRP"],
"Direct": ["Direct"],
"All": ["All pricing"],
"MRSP & All": ["MSRP", "All pricing"]
}
# Helper function to check for keywords in a case-insensitive way
def check_keywords(statement: str, keywords: List[str]) -> bool:
return any(re.search(rf"\b{keyword}\b", statement, re.IGNORECASE) for keyword in keywords)
# Function to determine sub-classification based on keywords
def get_sub_classification(statement: str) -> str:
for sub_label, keywords in sub_classification_keywords.items():
if all(check_keywords(statement, [keyword]) for keyword in keywords):
return sub_label
return "None" # Default to "None" if no keywords match
@app.post("/classify_with_subcategory/", response_model=ClassificationResponse, description="Classify text into main categories with subcategories.")
async def classify_with_subcategory(statement: str = Form(...)) -> ClassificationResponse:
try:
# Check if the statement is empty or "N/A"
if not statement or statement.strip().lower() == "n/a":
return ClassificationResponse(
classification="Notes not clear",
sub_classification="None",
confidence=1.0,
scores={"main": 1.0}
)
# Keyword-based classification override
if check_keywords(statement, change_to_quote_keywords):
main_best_label = "Change to quote"
main_best_score = 1.0 # High confidence since it's a direct match
elif check_keywords(statement, copy_quote_requested_keywords):
main_best_label = "Copy quote requested"
main_best_score = 1.0
else:
# If no keywords matched, perform the main classification using the model
loop = asyncio.get_running_loop()
main_classification_result = await loop.run_in_executor(
None,
lambda: nlp_sequence_classification(statement, main_labels, multi_label=False)
)
# Extract the best main classification label and confidence score
main_best_label = main_classification_result["labels"][0]
main_best_score = main_classification_result["scores"][0]
# Perform sub-classification only if the main classification is "Copy quote requested"
if main_best_label == "Copy quote requested":
best_sub_label = get_sub_classification(statement)
else:
best_sub_label = "None"
# Gather the scores for response
scores = {"main": main_best_score}
if best_sub_label != "None":
scores[best_sub_label] = 1.0 # Assign full confidence to sub-classification matches
return ClassificationResponse(
classification=main_best_label,
sub_classification=best_sub_label,
confidence=main_best_score,
scores=scores
)
except asyncio.TimeoutError:
# Handle timeout errors
return JSONResponse(content="Classification timed out. Try a shorter input or increase timeout.", status_code=504)
except HTTPException as http_exc:
# Handle HTTP errors
return JSONResponse(content=f"HTTP error: {http_exc.detail}", status_code=http_exc.status_code)
except Exception as e:
# Handle any other errors
return JSONResponse(content=f"Error in classification pipeline: {str(e)}", status_code=500)
# Set up CORS middleware
origins = ["*"] # or specify your list of allowed origins
app.add_middleware(
CORSMiddleware,
allow_origins=origins,
allow_credentials=True,
allow_methods=["*"],
allow_headers=["*"],
)