Spaces:
Sleeping
Sleeping
File size: 7,114 Bytes
17a78b5 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 | """
Experiment: Local Table QA with TAPAS/TAPEX models on financial data.
Runs a table QA model locally on CPU to answer questions about data
from the PostgreSQL financial database.
Supports two architectures:
- TAPAS (google/tapas-*): cell selection + aggregation
- TAPEX (microsoft/tapex-*): seq2seq text generation (BART-based)
Run:
uv run --extra experiment python scripts/experiment_table_qa.py
uv run --extra experiment python scripts/experiment_table_qa.py --model google/tapas-small-finetuned-wtq
uv run --extra experiment python scripts/experiment_table_qa.py --model microsoft/tapex-base-finetuned-wtq
"""
import argparse
import json
import sys
import time
from datetime import datetime
from pathlib import Path
# Add project root to path so we can import src.*
sys.path.insert(0, str(Path(__file__).resolve().parent.parent))
import pandas as pd
from src.db.connection import get_connection
DEFAULT_MODEL = "google/tapas-mini-finetuned-wtq"
QUERY = """
SELECT
transaction_date::text AS date,
transaction_description AS description,
category_name AS category,
entry_amount::text AS amount,
account_name AS account
FROM v_transaction_details
WHERE category_name IS NOT NULL
ORDER BY transaction_date DESC
LIMIT 15
"""
QUESTIONS = [
"What is the total amount?",
"Which category has the highest amount?",
"How many transactions are there?",
]
RESULTS_FILE = Path(__file__).resolve().parent.parent / "eval_cases" / "table_qa_results.jsonl"
def is_tapex(model_name: str) -> bool:
return "tapex" in model_name.lower()
def fetch_table() -> pd.DataFrame:
print("[STEP 1] Connecting to PostgreSQL...")
with get_connection() as conn:
print(f" -> Connected to: {conn.dsn}")
print(f" -> Executing query:\n{QUERY.strip()}")
with conn.cursor() as cur:
cur.execute(QUERY)
columns = [desc[0] for desc in cur.description]
rows = cur.fetchall()
print(f" -> Raw result: {len(rows)} rows, {len(columns)} columns")
print(f" -> Columns: {columns}")
# Build DataFrame from dict of lists — pandas 2.x defaults to object dtype.
# TAPAS tokenizer mutates cells via iloc with its internal Cell namedtuple,
# which requires object dtype (incompatible with pandas 3.0 StringDtype).
data = {col: [str(row[i]) for row in rows] for i, col in enumerate(columns)}
df = pd.DataFrame.from_dict(data)
print(f" -> Dtypes (should be object):\n{df.dtypes.to_string()}")
print(f" -> Sample row: {dict(df.iloc[0])}")
print()
return df
def load_tapas(model_name):
from transformers import pipeline
tqa = pipeline("table-question-answering", model=model_name, device=-1)
param_count = sum(p.numel() for p in tqa.model.parameters())
print(f" -> Architecture: TAPAS (cell selection + aggregation)")
print(f" -> Model class: {type(tqa.model).__name__}")
print(f" -> Tokenizer: {type(tqa.tokenizer).__name__}")
print(f" -> Model params: {param_count:,}")
return tqa, param_count
def load_tapex(model_name):
from transformers import BartForConditionalGeneration, TapexTokenizer
tokenizer = TapexTokenizer.from_pretrained(model_name)
model = BartForConditionalGeneration.from_pretrained(model_name)
param_count = sum(p.numel() for p in model.parameters())
print(f" -> Architecture: TAPEX (seq2seq text generation, BART-based)")
print(f" -> Model class: {type(model).__name__}")
print(f" -> Tokenizer: {type(tokenizer).__name__}")
print(f" -> Model params: {param_count:,}")
return (tokenizer, model), param_count
def run_tapas(tqa, table, query):
result = tqa(table=table, query=query)
return {
"answer": result["answer"],
"cells": result.get("cells", []),
"aggregator": result.get("aggregator", "NONE"),
}
def run_tapex(tapex_pair, table, query):
tokenizer, model = tapex_pair
encoding = tokenizer(table=table, query=query, return_tensors="pt", truncation=True)
print(f" -> Input token count: {encoding['input_ids'].shape[1]}")
outputs = model.generate(**encoding, max_new_tokens=50)
decoded = tokenizer.batch_decode(outputs, skip_special_tokens=True)
answer = decoded[0] if decoded else ""
return {
"answer": answer,
"cells": [],
"aggregator": "seq2seq",
}
def main():
parser = argparse.ArgumentParser(description="Table QA experiment (TAPAS/TAPEX)")
parser.add_argument("--model", default=DEFAULT_MODEL, help="HuggingFace model name")
args = parser.parse_args()
model_name = args.model
use_tapex = is_tapex(model_name)
run_results = {
"timestamp": datetime.now().isoformat(),
"model": model_name,
"architecture": "tapex" if use_tapex else "tapas",
"questions": [],
}
# --- Model loading ---
print("=" * 60)
print("[STEP 2] Loading model")
print("=" * 60)
print(f" -> Model: {model_name}")
print(f" -> Device: CPU")
print()
t0 = time.time()
if use_tapex:
model_obj, param_count = load_tapex(model_name)
else:
model_obj, param_count = load_tapas(model_name)
load_time = time.time() - t0
print(f" -> Load time: {load_time:.2f}s")
print()
run_results["params"] = param_count
run_results["load_time_s"] = round(load_time, 2)
# --- Data fetching ---
print("=" * 60)
print("[STEP 1] Fetching data from PostgreSQL")
print("=" * 60)
table = fetch_table()
# --- Display table ---
print("=" * 60)
print("TABLE (15 most recent transactions)")
print("=" * 60)
print(table.to_string(index=False))
print()
# --- Q&A ---
run_fn = run_tapex if use_tapex else run_tapas
print("=" * 60)
print("[STEP 3] Running Table QA inference")
print("=" * 60)
for i, q in enumerate(QUESTIONS, 1):
print(f"\n--- Question {i}/{len(QUESTIONS)} ---")
print(f" -> Input query: {q!r}")
print(f" -> Input table shape: {table.shape}")
t0 = time.time()
result = run_fn(model_obj, table, q)
inference_time = time.time() - t0
print(f" -> Answer: {result['answer']}")
print(f" -> Cells: {result.get('cells', [])}")
print(f" -> Aggregator: {result.get('aggregator', 'N/A')}")
print(f" -> Inference time: {inference_time:.3f}s")
run_results["questions"].append({
"query": q,
"answer": result["answer"],
"cells": result.get("cells", []),
"aggregator": result.get("aggregator", "N/A"),
"inference_time_s": round(inference_time, 3),
})
print()
# --- Save results ---
RESULTS_FILE.parent.mkdir(parents=True, exist_ok=True)
with open(RESULTS_FILE, "a") as f:
f.write(json.dumps(run_results, ensure_ascii=False) + "\n")
print(f"Results appended to: {RESULTS_FILE}")
print("=" * 60)
print("Experiment complete.")
print("=" * 60)
if __name__ == "__main__":
main()
|