WIP
This commit is contained in:
Yaojia Wang
103
scripts/analyze_v3.py
Normal file
103
scripts/analyze_v3.py
Normal file
@@ -0,0 +1,103 @@
|
||||
#!/usr/bin/env python3
|
||||
"""Analyze batch inference v3 results (Round 2 fixes)."""
|
||||
|
||||
import json
|
||||
from collections import Counter
|
||||
|
||||
with open("scripts/inference_results_v3.json") as f:
|
||||
results = json.load(f)
|
||||
|
||||
total = len(results)
|
||||
success = sum(1 for r in results if r["status"] == 200)
|
||||
print(f"Total PDFs: {total}, Successful: {success}")
|
||||
print()
|
||||
|
||||
# Summary table
|
||||
header = f"{'PDF':<40} {'Det':<4} {'Fld':<4} {'Time':<7} Fields"
|
||||
print(header)
|
||||
print("-" * 140)
|
||||
for r in results:
|
||||
fn = r["filename"][:39]
|
||||
data = r.get("data", {})
|
||||
result_data = data.get("result", {})
|
||||
fields = result_data.get("fields", {})
|
||||
dets = len(result_data.get("detections", []))
|
||||
nfields = len(fields)
|
||||
t = r["time_seconds"]
|
||||
parts = []
|
||||
for k, v in fields.items():
|
||||
sv = str(v)
|
||||
if len(sv) > 30:
|
||||
sv = sv[:27] + "..."
|
||||
parts.append(f"{k}={sv}")
|
||||
field_str = ", ".join(parts)
|
||||
print(f"{fn:<40} {dets:<4} {nfields:<4} {t:<7} {field_str}")
|
||||
|
||||
print()
|
||||
|
||||
# Field coverage
|
||||
field_counts: Counter = Counter()
|
||||
conf_sums: Counter = Counter()
|
||||
ok_count = 0
|
||||
for r in results:
|
||||
if r["status"] != 200:
|
||||
continue
|
||||
ok_count += 1
|
||||
result_data = r["data"]["result"]
|
||||
for k in result_data.get("fields", {}):
|
||||
field_counts[k] += 1
|
||||
for k, v in (result_data.get("confidence") or {}).items():
|
||||
conf_sums[k] += v
|
||||
|
||||
print(f"Field Coverage ({ok_count} successful PDFs):")
|
||||
hdr = f"{'Field':<35} {'Present':<10} {'Rate':<10} {'Avg Conf':<10}"
|
||||
print(hdr)
|
||||
print("-" * 65)
|
||||
for field in [
|
||||
"InvoiceNumber", "InvoiceDate", "InvoiceDueDate", "OCR",
|
||||
"Amount", "Bankgiro", "Plusgiro",
|
||||
"supplier_organisation_number", "customer_number", "payment_line",
|
||||
]:
|
||||
cnt = field_counts.get(field, 0)
|
||||
rate = cnt / ok_count * 100 if ok_count else 0
|
||||
avg_conf = conf_sums.get(field, 0) / cnt if cnt else 0
|
||||
flag = ""
|
||||
if rate < 30:
|
||||
flag = " <<<"
|
||||
elif rate < 60:
|
||||
flag = " !!"
|
||||
print(f"{field:<35} {cnt:<10} {rate:<10.1f} {avg_conf:<10.3f}{flag}")
|
||||
|
||||
# Fallback count
|
||||
fb_count = 0
|
||||
for r in results:
|
||||
if r["status"] == 200:
|
||||
result_data = r["data"]["result"]
|
||||
if result_data.get("fallback_used"):
|
||||
fb_count += 1
|
||||
print(f"\nFallback used: {fb_count}/{ok_count}")
|
||||
|
||||
# Low-confidence fields
|
||||
print("\nLow-confidence extractions (< 0.7):")
|
||||
for r in results:
|
||||
if r["status"] != 200:
|
||||
continue
|
||||
result_data = r["data"]["result"]
|
||||
for k, v in (result_data.get("confidence") or {}).items():
|
||||
if v < 0.7:
|
||||
fv = result_data.get("fields", {}).get(k, "?")
|
||||
print(f" [{v:.3f}] {k:<25} = {str(fv)[:40]:<40} ({r['filename'][:36]})")
|
||||
|
||||
# PDFs with very few fields (possible issues)
|
||||
print("\nPDFs with <= 2 fields extracted:")
|
||||
for r in results:
|
||||
if r["status"] != 200:
|
||||
continue
|
||||
result_data = r["data"]["result"]
|
||||
fields = result_data.get("fields", {})
|
||||
if len(fields) <= 2:
|
||||
print(f" {r['filename']}: {len(fields)} fields - {list(fields.keys())}")
|
||||
|
||||
# Avg time
|
||||
avg_time = sum(r["time_seconds"] for r in results) / len(results)
|
||||
print(f"\nAverage processing time: {avg_time:.2f}s")
|
||||
92
scripts/batch_inference_v3.py
Normal file
92
scripts/batch_inference_v3.py
Normal file
@@ -0,0 +1,92 @@
|
||||
#!/usr/bin/env python3
|
||||
"""Batch inference v3 - 30 random PDFs for Round 2 validation."""
|
||||
|
||||
import json
|
||||
import os
|
||||
import random
|
||||
import time
|
||||
|
||||
import requests
|
||||
|
||||
PDF_DIR = "/mnt/c/Users/yaoji/git/Billo/Billo.Platform.Document/Billo.Platform.Document.AdminAPI/downloads/to_check"
|
||||
API_URL = "http://localhost:8000/api/v1/infer"
|
||||
OUTPUT_FILE = "/mnt/c/Users/yaoji/git/ColaCoder/invoice-master-poc-v2/scripts/inference_results_v3.json"
|
||||
SAMPLE_SIZE = 30
|
||||
|
||||
|
||||
def main():
|
||||
random.seed(99_2026) # New seed for Round 3
|
||||
|
||||
all_pdfs = [f for f in os.listdir(PDF_DIR) if f.lower().endswith(".pdf")]
|
||||
selected = random.sample(all_pdfs, min(SAMPLE_SIZE, len(all_pdfs)))
|
||||
|
||||
print(f"Selected {len(selected)} random PDFs for inference")
|
||||
|
||||
results = []
|
||||
for i, filename in enumerate(selected, 1):
|
||||
filepath = os.path.join(PDF_DIR, filename)
|
||||
filesize = os.path.getsize(filepath)
|
||||
print(f"[{i}/{len(selected)}] Processing {filename}...", end=" ", flush=True)
|
||||
|
||||
start = time.time()
|
||||
try:
|
||||
with open(filepath, "rb") as f:
|
||||
resp = requests.post(
|
||||
API_URL,
|
||||
files={"file": (filename, f, "application/pdf")},
|
||||
timeout=120,
|
||||
)
|
||||
elapsed = round(time.time() - start, 2)
|
||||
|
||||
if resp.status_code == 200:
|
||||
data = resp.json()
|
||||
field_count = sum(
|
||||
1 for k, v in data.items()
|
||||
if k not in (
|
||||
"DocumentId", "confidence", "success", "fallback_used",
|
||||
"bboxes", "cross_validation", "processing_time_ms",
|
||||
"line_items", "vat_summary", "vat_validation",
|
||||
"raw_detections", "detection_classes", "detection_count",
|
||||
)
|
||||
and v is not None
|
||||
)
|
||||
det_count = data.get("detection_count", "?")
|
||||
print(f"OK ({elapsed}s) - {field_count} fields, {det_count} detections")
|
||||
results.append({
|
||||
"filename": filename,
|
||||
"status": resp.status_code,
|
||||
"time_seconds": elapsed,
|
||||
"filesize": filesize,
|
||||
"data": data,
|
||||
})
|
||||
else:
|
||||
print(f"HTTP {resp.status_code} ({elapsed}s)")
|
||||
results.append({
|
||||
"filename": filename,
|
||||
"status": resp.status_code,
|
||||
"time_seconds": elapsed,
|
||||
"filesize": filesize,
|
||||
"error": resp.text[:200],
|
||||
})
|
||||
except Exception as e:
|
||||
elapsed = round(time.time() - start, 2)
|
||||
print(f"FAILED ({elapsed}s) - {e}")
|
||||
results.append({
|
||||
"filename": filename,
|
||||
"status": -1,
|
||||
"time_seconds": elapsed,
|
||||
"filesize": filesize,
|
||||
"error": str(e),
|
||||
})
|
||||
|
||||
with open(OUTPUT_FILE, "w", encoding="utf-8") as f:
|
||||
json.dump(results, f, ensure_ascii=False, indent=2)
|
||||
print(f"\nResults saved to {OUTPUT_FILE}")
|
||||
|
||||
success = sum(1 for r in results if r["status"] == 200)
|
||||
failed = len(results) - success
|
||||
print(f"Total: {len(results)}, Success: {success}, Failed: {failed}")
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
main()
|
||||
2986
scripts/inference_results_v3.json
Normal file
2986
scripts/inference_results_v3.json
Normal file
File diff suppressed because it is too large
Load Diff
@@ -1,387 +0,0 @@
|
||||
#!/usr/bin/env python3
|
||||
"""
|
||||
PP-StructureV3 Line Items Extraction POC
|
||||
|
||||
Tests line items extraction from Swedish invoices using PP-StructureV3.
|
||||
Parses HTML table structure to extract structured line item data.
|
||||
|
||||
Run with invoice-sm120 conda environment.
|
||||
"""
|
||||
|
||||
import sys
|
||||
import re
|
||||
from pathlib import Path
|
||||
from html.parser import HTMLParser
|
||||
from dataclasses import dataclass
|
||||
|
||||
# Add project root to path
|
||||
project_root = Path(__file__).parent.parent
|
||||
sys.path.insert(0, str(project_root / "packages" / "backend"))
|
||||
|
||||
from paddleocr import PPStructureV3
|
||||
import fitz # PyMuPDF
|
||||
|
||||
|
||||
@dataclass
|
||||
class LineItem:
|
||||
"""Single line item from invoice."""
|
||||
row_index: int
|
||||
article_number: str | None
|
||||
description: str | None
|
||||
quantity: str | None
|
||||
unit: str | None
|
||||
unit_price: str | None
|
||||
amount: str | None
|
||||
vat_rate: str | None
|
||||
confidence: float = 0.9
|
||||
|
||||
|
||||
class TableHTMLParser(HTMLParser):
|
||||
"""Parse HTML table into rows and cells."""
|
||||
|
||||
def __init__(self):
|
||||
super().__init__()
|
||||
self.rows: list[list[str]] = []
|
||||
self.current_row: list[str] = []
|
||||
self.current_cell: str = ""
|
||||
self.in_td = False
|
||||
self.in_thead = False
|
||||
self.header_row: list[str] = []
|
||||
|
||||
def handle_starttag(self, tag, attrs):
|
||||
if tag == "tr":
|
||||
self.current_row = []
|
||||
elif tag in ("td", "th"):
|
||||
self.in_td = True
|
||||
self.current_cell = ""
|
||||
elif tag == "thead":
|
||||
self.in_thead = True
|
||||
|
||||
def handle_endtag(self, tag):
|
||||
if tag in ("td", "th"):
|
||||
self.in_td = False
|
||||
self.current_row.append(self.current_cell.strip())
|
||||
elif tag == "tr":
|
||||
if self.current_row:
|
||||
if self.in_thead:
|
||||
self.header_row = self.current_row
|
||||
else:
|
||||
self.rows.append(self.current_row)
|
||||
elif tag == "thead":
|
||||
self.in_thead = False
|
||||
|
||||
def handle_data(self, data):
|
||||
if self.in_td:
|
||||
self.current_cell += data
|
||||
|
||||
|
||||
# Swedish column name mappings
|
||||
# Note: Some headers may contain multiple column names merged together
|
||||
COLUMN_MAPPINGS = {
|
||||
'article_number': ['art nummer', 'artikelnummer', 'artikel', 'artnr', 'art.nr', 'art nr'],
|
||||
'description': ['beskrivning', 'produktbeskrivning', 'produkt', 'tjänst', 'text', 'benämning', 'vara/tjänst', 'vara'],
|
||||
'quantity': ['antal', 'qty', 'st', 'pcs', 'kvantitet'],
|
||||
'unit': ['enhet', 'unit'],
|
||||
'unit_price': ['á-pris', 'a-pris', 'pris', 'styckpris', 'enhetspris', 'à pris'],
|
||||
'amount': ['belopp', 'summa', 'total', 'netto', 'rad summa'],
|
||||
'vat_rate': ['moms', 'moms%', 'vat', 'skatt', 'moms %'],
|
||||
}
|
||||
|
||||
|
||||
def normalize_header(header: str) -> str:
|
||||
"""Normalize header text for matching."""
|
||||
return header.lower().strip().replace(".", "").replace("-", " ")
|
||||
|
||||
|
||||
def map_columns(headers: list[str]) -> dict[int, str]:
|
||||
"""Map column indices to field names."""
|
||||
mapping = {}
|
||||
for idx, header in enumerate(headers):
|
||||
normalized = normalize_header(header)
|
||||
|
||||
# Skip empty headers
|
||||
if not normalized.strip():
|
||||
continue
|
||||
|
||||
best_match = None
|
||||
best_match_len = 0
|
||||
|
||||
for field, patterns in COLUMN_MAPPINGS.items():
|
||||
for pattern in patterns:
|
||||
# Require exact match or pattern must be a significant portion
|
||||
if pattern == normalized:
|
||||
# Exact match - use immediately
|
||||
best_match = field
|
||||
best_match_len = len(pattern) + 100 # Prioritize exact
|
||||
break
|
||||
elif pattern in normalized and len(pattern) > best_match_len:
|
||||
# Pattern found in header - use longer matches
|
||||
if len(pattern) >= 3: # Minimum pattern length
|
||||
best_match = field
|
||||
best_match_len = len(pattern)
|
||||
|
||||
if best_match_len > 100: # Was exact match
|
||||
break
|
||||
|
||||
if best_match:
|
||||
mapping[idx] = best_match
|
||||
|
||||
return mapping
|
||||
|
||||
|
||||
def parse_table_html(html: str) -> tuple[list[str], list[list[str]]]:
|
||||
"""Parse HTML table and return header and rows."""
|
||||
parser = TableHTMLParser()
|
||||
parser.feed(html)
|
||||
return parser.header_row, parser.rows
|
||||
|
||||
|
||||
def detect_header_row(rows: list[list[str]]) -> tuple[int, list[str], bool]:
|
||||
"""
|
||||
Detect which row is the header based on content patterns.
|
||||
|
||||
Returns (header_row_index, header_row, is_at_end).
|
||||
is_at_end indicates if header is at the end (table is reversed).
|
||||
Returns (-1, [], False) if no header detected.
|
||||
"""
|
||||
header_keywords = set()
|
||||
for patterns in COLUMN_MAPPINGS.values():
|
||||
for p in patterns:
|
||||
header_keywords.add(p.lower())
|
||||
|
||||
best_match = (-1, [], 0)
|
||||
|
||||
for i, row in enumerate(rows):
|
||||
# Skip empty rows
|
||||
if all(not cell.strip() for cell in row):
|
||||
continue
|
||||
|
||||
# Check if row contains header keywords
|
||||
row_text = " ".join(cell.lower() for cell in row)
|
||||
matches = sum(1 for kw in header_keywords if kw in row_text)
|
||||
|
||||
# Track the best match
|
||||
if matches > best_match[2]:
|
||||
best_match = (i, row, matches)
|
||||
|
||||
if best_match[2] >= 2:
|
||||
header_idx = best_match[0]
|
||||
is_at_end = header_idx == len(rows) - 1 or header_idx > len(rows) // 2
|
||||
return header_idx, best_match[1], is_at_end
|
||||
|
||||
return -1, [], False
|
||||
|
||||
|
||||
def extract_line_items(html: str) -> list[LineItem]:
|
||||
"""Extract line items from HTML table."""
|
||||
header, rows = parse_table_html(html)
|
||||
|
||||
is_reversed = False
|
||||
if not header:
|
||||
# Try to detect header row from content
|
||||
header_idx, detected_header, is_at_end = detect_header_row(rows)
|
||||
if header_idx >= 0:
|
||||
header = detected_header
|
||||
if is_at_end:
|
||||
# Header is at the end - table is reversed
|
||||
is_reversed = True
|
||||
rows = rows[:header_idx] # Data rows are before header
|
||||
else:
|
||||
rows = rows[header_idx + 1:] # Data rows start after header
|
||||
elif rows:
|
||||
# Fall back to first non-empty row
|
||||
for i, row in enumerate(rows):
|
||||
if any(cell.strip() for cell in row):
|
||||
header = row
|
||||
rows = rows[i + 1:]
|
||||
break
|
||||
|
||||
column_map = map_columns(header)
|
||||
|
||||
items = []
|
||||
for row_idx, row in enumerate(rows):
|
||||
item_data = {
|
||||
'row_index': row_idx,
|
||||
'article_number': None,
|
||||
'description': None,
|
||||
'quantity': None,
|
||||
'unit': None,
|
||||
'unit_price': None,
|
||||
'amount': None,
|
||||
'vat_rate': None,
|
||||
}
|
||||
|
||||
for col_idx, cell in enumerate(row):
|
||||
if col_idx in column_map:
|
||||
field = column_map[col_idx]
|
||||
item_data[field] = cell if cell else None
|
||||
|
||||
# Only add if we have at least description or amount
|
||||
if item_data['description'] or item_data['amount']:
|
||||
items.append(LineItem(**item_data))
|
||||
|
||||
return items
|
||||
|
||||
|
||||
def render_pdf_to_image(pdf_path: str, dpi: int = 200) -> bytes:
|
||||
"""Render first page of PDF to image bytes."""
|
||||
doc = fitz.open(pdf_path)
|
||||
page = doc[0]
|
||||
mat = fitz.Matrix(dpi / 72, dpi / 72)
|
||||
pix = page.get_pixmap(matrix=mat)
|
||||
img_bytes = pix.tobytes("png")
|
||||
doc.close()
|
||||
return img_bytes
|
||||
|
||||
|
||||
def test_line_items_extraction(pdf_path: str) -> dict:
|
||||
"""Test line items extraction on a PDF."""
|
||||
print(f"\n{'='*70}")
|
||||
print(f"Testing Line Items Extraction: {Path(pdf_path).name}")
|
||||
print(f"{'='*70}")
|
||||
|
||||
# Render PDF to image
|
||||
print("Rendering PDF to image...")
|
||||
img_bytes = render_pdf_to_image(pdf_path)
|
||||
|
||||
# Save temp image
|
||||
temp_img_path = "/tmp/test_invoice.png"
|
||||
with open(temp_img_path, "wb") as f:
|
||||
f.write(img_bytes)
|
||||
|
||||
# Initialize PP-StructureV3
|
||||
print("Initializing PP-StructureV3...")
|
||||
pipeline = PPStructureV3(
|
||||
device="gpu:0",
|
||||
use_doc_orientation_classify=False,
|
||||
use_doc_unwarping=False,
|
||||
)
|
||||
|
||||
# Run detection
|
||||
print("Running table detection...")
|
||||
results = pipeline.predict(temp_img_path)
|
||||
|
||||
all_line_items = []
|
||||
table_details = []
|
||||
|
||||
for result in results if results else []:
|
||||
table_res_list = result.get("table_res_list") if hasattr(result, "get") else None
|
||||
|
||||
if table_res_list:
|
||||
print(f"\nFound {len(table_res_list)} tables")
|
||||
|
||||
for i, table_res in enumerate(table_res_list):
|
||||
html = table_res.get("pred_html", "")
|
||||
ocr_pred = table_res.get("table_ocr_pred", {})
|
||||
|
||||
print(f"\n--- Table {i+1} ---")
|
||||
|
||||
# Debug: show full HTML for first table
|
||||
if i == 0:
|
||||
print(f" Full HTML:\n{html}")
|
||||
|
||||
# Debug: inspect table_ocr_pred structure
|
||||
if isinstance(ocr_pred, dict):
|
||||
print(f" table_ocr_pred keys: {list(ocr_pred.keys())}")
|
||||
# Check if rec_texts exists (actual OCR text)
|
||||
if "rec_texts" in ocr_pred:
|
||||
texts = ocr_pred["rec_texts"]
|
||||
print(f" OCR texts count: {len(texts)}")
|
||||
print(f" Sample OCR texts: {texts[:5]}")
|
||||
elif isinstance(ocr_pred, list):
|
||||
print(f" table_ocr_pred is list with {len(ocr_pred)} items")
|
||||
if ocr_pred:
|
||||
print(f" First item type: {type(ocr_pred[0])}")
|
||||
print(f" First few items: {ocr_pred[:3]}")
|
||||
|
||||
# Parse HTML
|
||||
header, rows = parse_table_html(html)
|
||||
print(f" HTML Header (from thead): {header}")
|
||||
print(f" HTML Rows: {len(rows)}")
|
||||
|
||||
# Try to detect header if not in thead
|
||||
detected_header = None
|
||||
is_reversed = False
|
||||
if not header and rows:
|
||||
header_idx, detected_header, is_at_end = detect_header_row(rows)
|
||||
if header_idx >= 0:
|
||||
is_reversed = is_at_end
|
||||
print(f" Detected header at row {header_idx}: {detected_header}")
|
||||
print(f" Table is {'REVERSED (header at bottom)' if is_reversed else 'normal'}")
|
||||
header = detected_header
|
||||
|
||||
if rows:
|
||||
print(f" First row: {rows[0]}")
|
||||
if len(rows) > 1:
|
||||
print(f" Second row: {rows[1]}")
|
||||
|
||||
# Check if this looks like a line items table
|
||||
column_map = map_columns(header) if header else {}
|
||||
print(f" Column mapping: {column_map}")
|
||||
|
||||
is_line_items_table = (
|
||||
'description' in column_map.values() or
|
||||
'amount' in column_map.values() or
|
||||
'article_number' in column_map.values()
|
||||
)
|
||||
|
||||
if is_line_items_table:
|
||||
print(f" >>> This appears to be a LINE ITEMS table!")
|
||||
items = extract_line_items(html)
|
||||
print(f" Extracted {len(items)} line items:")
|
||||
for item in items:
|
||||
print(f" - {item.description}: {item.quantity} x {item.unit_price} = {item.amount}")
|
||||
all_line_items.extend(items)
|
||||
else:
|
||||
print(f" >>> This is NOT a line items table (summary/payment)")
|
||||
|
||||
table_details.append({
|
||||
"index": i,
|
||||
"header": header,
|
||||
"row_count": len(rows),
|
||||
"is_line_items": is_line_items_table,
|
||||
"column_map": column_map,
|
||||
})
|
||||
|
||||
print(f"\n{'='*70}")
|
||||
print(f"EXTRACTION SUMMARY")
|
||||
print(f"{'='*70}")
|
||||
print(f"Total tables: {len(table_details)}")
|
||||
print(f"Line items tables: {sum(1 for t in table_details if t['is_line_items'])}")
|
||||
print(f"Total line items: {len(all_line_items)}")
|
||||
|
||||
return {
|
||||
"pdf": pdf_path,
|
||||
"tables": table_details,
|
||||
"line_items": all_line_items,
|
||||
}
|
||||
|
||||
|
||||
def main():
|
||||
import argparse
|
||||
parser = argparse.ArgumentParser(description="Test line items extraction")
|
||||
parser.add_argument("--pdf", type=str, help="Path to PDF file")
|
||||
args = parser.parse_args()
|
||||
|
||||
if args.pdf:
|
||||
# Test specific PDF
|
||||
pdf_path = Path(args.pdf)
|
||||
if not pdf_path.exists():
|
||||
# Try relative to project root
|
||||
pdf_path = project_root / args.pdf
|
||||
if not pdf_path.exists():
|
||||
print(f"PDF not found: {args.pdf}")
|
||||
return
|
||||
test_line_items_extraction(str(pdf_path))
|
||||
else:
|
||||
# Test default invoice
|
||||
default_pdf = project_root / "exampl" / "Faktura54011.pdf"
|
||||
if default_pdf.exists():
|
||||
test_line_items_extraction(str(default_pdf))
|
||||
else:
|
||||
print(f"Default PDF not found: {default_pdf}")
|
||||
print("Usage: python ppstructure_line_items_poc.py --pdf <path>")
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
main()
|
||||
@@ -1,154 +0,0 @@
|
||||
#!/usr/bin/env python3
|
||||
"""
|
||||
PP-StructureV3 POC Script
|
||||
|
||||
Tests table detection on real Swedish invoices using PP-StructureV3.
|
||||
Run with invoice-sm120 conda environment.
|
||||
"""
|
||||
|
||||
import sys
|
||||
from pathlib import Path
|
||||
|
||||
# Add project root to path
|
||||
project_root = Path(__file__).parent.parent
|
||||
sys.path.insert(0, str(project_root / "packages" / "backend"))
|
||||
|
||||
from paddleocr import PPStructureV3
|
||||
import fitz # PyMuPDF
|
||||
|
||||
|
||||
def render_pdf_to_image(pdf_path: str, dpi: int = 200) -> bytes:
|
||||
"""Render first page of PDF to image bytes."""
|
||||
doc = fitz.open(pdf_path)
|
||||
page = doc[0]
|
||||
mat = fitz.Matrix(dpi / 72, dpi / 72)
|
||||
pix = page.get_pixmap(matrix=mat)
|
||||
img_bytes = pix.tobytes("png")
|
||||
doc.close()
|
||||
return img_bytes
|
||||
|
||||
|
||||
def test_table_detection(pdf_path: str) -> dict:
|
||||
"""Test PP-StructureV3 table detection on a PDF."""
|
||||
print(f"\n{'='*60}")
|
||||
print(f"Testing: {Path(pdf_path).name}")
|
||||
print(f"{'='*60}")
|
||||
|
||||
# Render PDF to image
|
||||
print("Rendering PDF to image...")
|
||||
img_bytes = render_pdf_to_image(pdf_path)
|
||||
|
||||
# Save temp image
|
||||
temp_img_path = "/tmp/test_invoice.png"
|
||||
with open(temp_img_path, "wb") as f:
|
||||
f.write(img_bytes)
|
||||
print(f"Saved temp image: {temp_img_path}")
|
||||
|
||||
# Initialize PP-StructureV3
|
||||
print("Initializing PP-StructureV3...")
|
||||
pipeline = PPStructureV3(
|
||||
device="gpu:0",
|
||||
use_doc_orientation_classify=False,
|
||||
use_doc_unwarping=False,
|
||||
)
|
||||
|
||||
# Run detection
|
||||
print("Running table detection...")
|
||||
results = pipeline.predict(temp_img_path)
|
||||
|
||||
# Parse results - PaddleX 3.x returns dict-like LayoutParsingResultV2
|
||||
tables_found = []
|
||||
all_elements = []
|
||||
|
||||
for result in results if results else []:
|
||||
# Get table results from the new API
|
||||
table_res_list = result.get("table_res_list") if hasattr(result, "get") else None
|
||||
|
||||
if table_res_list:
|
||||
print(f" Found {len(table_res_list)} tables in table_res_list")
|
||||
for i, table_res in enumerate(table_res_list):
|
||||
# Debug: show all keys in table_res
|
||||
if isinstance(table_res, dict):
|
||||
print(f" Table {i+1} keys: {list(table_res.keys())}")
|
||||
else:
|
||||
print(f" Table {i+1} attrs: {[a for a in dir(table_res) if not a.startswith('_')]}")
|
||||
|
||||
# Extract table info - use correct key names from PaddleX 3.x
|
||||
cell_boxes = table_res.get("cell_box_list", [])
|
||||
html = table_res.get("pred_html", "") # HTML is in pred_html
|
||||
ocr_text = table_res.get("table_ocr_pred", [])
|
||||
region_id = table_res.get("table_region_id", -1)
|
||||
bbox = [] # bbox is stored elsewhere in parsing_res_list
|
||||
|
||||
print(f" Table {i+1}:")
|
||||
print(f" - Cells: {len(cell_boxes) if cell_boxes is not None else 0}")
|
||||
print(f" - Region ID: {region_id}")
|
||||
print(f" - HTML length: {len(html) if html else 0}")
|
||||
print(f" - OCR texts: {len(ocr_text) if ocr_text else 0}")
|
||||
|
||||
if html:
|
||||
print(f" - HTML preview: {html[:300]}...")
|
||||
|
||||
if ocr_text and len(ocr_text) > 0:
|
||||
print(f" - First few OCR texts: {ocr_text[:3]}")
|
||||
|
||||
tables_found.append({
|
||||
"index": i,
|
||||
"cell_count": len(cell_boxes) if cell_boxes is not None else 0,
|
||||
"region_id": region_id,
|
||||
"html": html[:1000] if html else "",
|
||||
"ocr_count": len(ocr_text) if ocr_text else 0,
|
||||
})
|
||||
|
||||
# Get parsing results for all layout elements
|
||||
parsing_res_list = result.get("parsing_res_list") if hasattr(result, "get") else None
|
||||
|
||||
if parsing_res_list:
|
||||
print(f"\n Layout elements from parsing_res_list:")
|
||||
for elem in parsing_res_list[:10]: # Show first 10
|
||||
label = elem.get("label", "unknown") if isinstance(elem, dict) else getattr(elem, "label", "unknown")
|
||||
bbox = elem.get("bbox", []) if isinstance(elem, dict) else getattr(elem, "bbox", [])
|
||||
print(f" - {label}: {bbox}")
|
||||
all_elements.append({"label": label, "bbox": bbox})
|
||||
|
||||
print(f"\nSummary:")
|
||||
print(f" Tables detected: {len(tables_found)}")
|
||||
print(f" Layout elements: {len(all_elements)}")
|
||||
|
||||
return {"pdf": pdf_path, "tables": tables_found, "elements": all_elements}
|
||||
|
||||
|
||||
def main():
|
||||
# Find test PDFs
|
||||
pdf_dir = Path("/mnt/c/Users/yaoji/git/ColaCoder/invoice-master-poc-v2/data/admin_uploads")
|
||||
pdf_files = list(pdf_dir.glob("*.pdf"))[:5] # Test first 5
|
||||
|
||||
if not pdf_files:
|
||||
print("No PDF files found in admin_uploads directory")
|
||||
return
|
||||
|
||||
print(f"Found {len(pdf_files)} PDF files")
|
||||
|
||||
all_results = []
|
||||
for pdf_file in pdf_files:
|
||||
result = test_table_detection(str(pdf_file))
|
||||
all_results.append(result)
|
||||
|
||||
# Summary
|
||||
print(f"\n{'='*60}")
|
||||
print("FINAL SUMMARY")
|
||||
print(f"{'='*60}")
|
||||
total_tables = sum(len(r["tables"]) for r in all_results)
|
||||
print(f"Total PDFs tested: {len(all_results)}")
|
||||
print(f"Total tables detected: {total_tables}")
|
||||
|
||||
for r in all_results:
|
||||
pdf_name = Path(r["pdf"]).name
|
||||
table_count = len(r["tables"])
|
||||
print(f" {pdf_name}: {table_count} tables")
|
||||
for t in r["tables"]:
|
||||
print(f" - Table {t['index']+1}: {t['cell_count']} cells")
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
main()
|
||||
54
scripts/render_pdfs_v3.py
Normal file
54
scripts/render_pdfs_v3.py
Normal file
@@ -0,0 +1,54 @@
|
||||
#!/usr/bin/env python3
|
||||
"""Render selected PDFs from v3 batch for visual comparison."""
|
||||
|
||||
import os
|
||||
|
||||
import fitz # PyMuPDF
|
||||
|
||||
PDF_DIR = "/mnt/c/Users/yaoji/git/Billo/Billo.Platform.Document/Billo.Platform.Document.AdminAPI/downloads/to_check"
|
||||
OUTPUT_DIR = "/mnt/c/Users/yaoji/git/ColaCoder/invoice-master-poc-v2/scripts/pdf_renders_v3"
|
||||
|
||||
# Select 10 PDFs covering different scenarios:
|
||||
SELECTED = [
|
||||
# Potentially wrong Amount (81648164.00 - too high?)
|
||||
"b84c7d70-821d-4a1a-9be7-d7bb2392bd91.pdf",
|
||||
# Only 2 fields extracted
|
||||
"072571e2-da5f-4268-b1a8-f0e5a85a3ec4.pdf",
|
||||
# InvoiceNumber=5085 (suspiciously short, same as BG prefix?)
|
||||
"6a83ba35-afdf-4c13-ade1-25513e213637.pdf",
|
||||
# InvoiceNumber=450 (very short, might be wrong)
|
||||
"8551b540-d93d-459d-b7eb-e9ee086f9f16.pdf",
|
||||
# InvoiceNumber=134 (very short, same as BG prefix)
|
||||
"cb1bd3b1-63d0-4140-930f-e4a7ae2b6cd5.pdf",
|
||||
# Large Amount=172904.52, InvoiceNumber=89902
|
||||
"d121a5ee-7382-41d8-8010-63880def1f96.pdf",
|
||||
# Good 9-field PDF for positive check
|
||||
"6cb90895-e52b-4831-b57b-7cb968bcdd54.pdf",
|
||||
# Amount=2026.00 (same as year - could be confused?)
|
||||
"d376c5b5-0dc5-4ccf-b787-0d481eef8577.pdf",
|
||||
# 8 fields, good coverage
|
||||
"f3f5da6f-7552-4ec6-8625-3629042fbfd0.pdf",
|
||||
# Low confidence Amount=596.49
|
||||
"5783e4af-eef3-411c-84b1-3a8f4694fed8.pdf",
|
||||
]
|
||||
|
||||
os.makedirs(OUTPUT_DIR, exist_ok=True)
|
||||
|
||||
for pdf_name in SELECTED:
|
||||
pdf_path = os.path.join(PDF_DIR, pdf_name)
|
||||
if not os.path.exists(pdf_path):
|
||||
print(f"SKIP {pdf_name} - not found")
|
||||
continue
|
||||
|
||||
doc = fitz.open(pdf_path)
|
||||
page = doc[0]
|
||||
mat = fitz.Matrix(150 / 72, 150 / 72)
|
||||
pix = page.get_pixmap(matrix=mat)
|
||||
|
||||
out_name = pdf_name.replace(".pdf", ".png")
|
||||
out_path = os.path.join(OUTPUT_DIR, out_name)
|
||||
pix.save(out_path)
|
||||
print(f"Rendered {pdf_name} -> {out_name} ({pix.width}x{pix.height})")
|
||||
doc.close()
|
||||
|
||||
print(f"\nAll renders saved to {OUTPUT_DIR}")
|
||||
Reference in New Issue
Block a user