dataset/annotation/utilities/embeddings.py

83 lines
3.2 KiB
Python

# --- Import librerie ---
import pandas as pd
from openai import AzureOpenAI
from sentence_transformers import SentenceTransformer
import numpy as np
import re
from openpyxl.styles import PatternFill
from openpyxl import load_workbook
from collections import Counter
from prompts.prompt import build_prompt_local
import warnings
import logging
from sentence_transformers import SentenceTransformer
import numpy as np
import pickle
import unicodedata
# ----- Caricare datasets -----
df_labeled = pd.read_excel("main/datasets/annotated_dataset.xlsx")
df_labeled = df_labeled.dropna(how="all") # rimuove righe completamente vuote
df_unlabeled = pd.read_excel("main/datasets/unlabeled_dataset.xlsx")
df_unlabeled = df_unlabeled.dropna(how="all")
print("***STEP 1***")
print("Dataset etichettato caricato. Numero righe:", len(df_labeled))
print("Dataset non etichettato caricato. Numero righe:", len(df_unlabeled))
# ----- Pulizia colonne ----
df_labeled = pd.read_excel("main/datasets/annotated_dataset.xlsx").dropna(how="all")
df_unlabeled = pd.read_excel("main/datasets/unlabeled_dataset.xlsx").dropna(how="all")
def clean_str(x):
if pd.isna(x):
return ""
s = str(x).strip().lower()
s = unicodedata.normalize("NFKC", s)
# rimuove tutti i caratteri non alfanumerici e spazi multipli, lascia solo lettere, numeri e spazi
s = re.sub(r'[^a-z0-9 ]+', '', s)
s = re.sub(r'\s+', ' ', s) # spazi multipli → 1 spazio
return s
# Applica pulizia su automation_id e folder
for df in [df_labeled, df_unlabeled]:
df["automation_id"] = df["automation_id"].apply(clean_str)
df["folder"] = df["folder"].apply(clean_str)
unlabeled_pairs = set(zip(df_unlabeled["automation_id"], df_unlabeled["folder"]))
# Filtro: rimuove dal dataset non etichettato le righe già presenti in df_labeled
labeled_pairs = set(zip(df_labeled["automation_id"], df_labeled["folder"]))
mask = ~df_unlabeled[["automation_id", "folder"]].apply(tuple, axis=1).isin(labeled_pairs)
df_unlabeled_filtered = df_unlabeled[mask]
print("Numero righe df_unlabeled dopo aver rimosso quelle etichettate:", len(df_unlabeled_filtered))
# Trova coppie mancanti (debug)
missing_pairs = labeled_pairs - unlabeled_pairs
print("Numero righe etichettate non trovate nel dataset non etichettato:", len(missing_pairs))
if missing_pairs:
print("Coppie mancanti:")
for p in missing_pairs:
print(p)
# ----- Step 2: embeddings -----
# Silenzia warning generici
warnings.filterwarnings("ignore")
# Silenzia logging di transformers / sentence-transformers / HF hub
logging.getLogger("sentence_transformers").setLevel(logging.ERROR)
logging.getLogger("transformers").setLevel(logging.ERROR)
logging.getLogger("huggingface_hub").setLevel(logging.ERROR)
model = SentenceTransformer("all-MiniLM-L6-v2")
texts = df_labeled["automation"].tolist()
embeddings = model.encode(texts, show_progress_bar=True, convert_to_numpy=True, normalize_embeddings=True)
embeddings = embeddings.astype("float32")
print("Shape embeddings ricalcolati:", embeddings.shape)
# ----- Step 3: salvare embeddings -----
with open("main/labeled_embeddings2.pkl", "wb") as f:
pickle.dump({"embeddings": embeddings, "automation_id": df_labeled["automation_id"].tolist()}, f)
print("Embeddings salvati con successo!")