app.py

import os
import gradio as gr
import pandas as pd
import tensorflow as tf
from tapas.scripts import prediction_utils
from tapas.utils import number_annotation_utils
from tapas.protos import interaction_pb2

# 1) Read CSV and build list-of-lists table
import pandas as pd

df = pd.read_csv("synthetic_profit.csv")
# Ensure all values are strings
df = df.astype(str)
# Build TAPAS-style table: header row + data rows
table = [list(df.columns)] + df.values.tolist()

# 2) Configure TAPAS conversion with aggregation support
from tapas.utils import example_utils as tf_example_utils
config = tf_example_utils.ClassifierConversionConfig(
    vocab_file="tapas_sqa_base/vocab.txt",
    max_seq_length=512,
    max_column_id=512,
    max_row_id=512,
    strip_column_names=False,             # Keep header names
    add_aggregation_candidates=True,     # Propose SUM/AVERAGE operations
)
converter = tf_example_utils.ToClassifierTensorflowExample(config)

# 3) Helper: convert one interaction to model input
def interaction_from_query(question: str):
    interaction = interaction_pb2.Interaction()
    # Add question
    q = interaction.questions.add()
    q.original_text = question
    # Add table columns
    for col in table[0]:
        interaction.table.columns.add().text = col
    # Add table rows/cells
    for row in table[1:]:
        r = interaction.table.rows.add()
        for cell in row:
            r.cells.add().text = cell
    # Annotate numeric values
    number_annotation_utils.add_numeric_values(interaction)
    return interaction

# 4) Instantiate TAPAS model and tokenizer
from transformers import TFAutoModelForSequenceClassification, AutoTokenizer
MODEL = "google/tapas-base-finetuned-wtq"
tokenizer = AutoTokenizer.from_pretrained(MODEL)
model     = TFAutoModelForSequenceClassification.from_pretrained(MODEL)

# 5) Prediction helper
def predict_answer(question: str):
    interaction = interaction_from_query(question)
    # Convert to TensorFlowExample
    tf_example = converter.convert(interaction)
    # Run prediction
    result = model(tf_example.features)
    # Parse answer coordinates
    coords = prediction_utils.parse_coordinates(result.logits)
    # Map coordinates back to table cells
    answers = []
    for r, c in coords:
        answers.append(table[r+1][c])
    return ", ".join(answers)

# 6) Gradio interface
iface = gr.Interface(
    fn=predict_answer,
    inputs=gr.Textbox(lines=2, placeholder="Ask a question…"),
    outputs=gr.Textbox(lines=3),
    title="SAP Profitability Q&A (TAPAS Low-Level)",
    description=(
        "Low-level TAPAS: list-of-lists input, numeric annotations, "
        "aggregation candidates, and coordinate post-processing."
    ),
    allow_flagging="never",
)

if __name__ == "__main__":
    iface.launch(server_name="0.0.0.0", server_port=7860)