Repository navigation
Expand file tree
/
Copy pathhyperparameters.py
More file actions
114 lines (88 loc) · 4.04 KB
/
Copy pathhyperparameters.py
File metadata and controls
114 lines (88 loc) · 4.04 KB
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
import spacy
import pandas as pd
import numpy as np
import unicodedata
from datasets import Dataset
from setfit import SetFitModel, Trainer
from optuna import Trial
def preparer_donnees():
# Importer le jeu de données
df = pd.read_excel("jeu_de_donnees.xlsx")
# On supprime les virgules, les "|" ainsi que les nombres dans la colonne Category Label
df['Category Label'] = df['Category Label'].str.replace(r'[)(\-,|/\d]+', '', regex=True)
df['Question Text'] = df['Question Text'].str.replace(r'[)(]', '', regex=True)
# On fusionne les colonnes Question Text et Category Label
# Replace None values with empty strings before concatenating
df['Question Text'] = df['Question Text'].fillna('')
df['Category Label'] = df['Category Label'].fillna('')
df = df.drop(columns=['Category Label'])
# S'il n'y a ni question, ni modalité, alors on met le nom du label
df['Question Text'] = df['Question Text'].replace('', np.nan)
df['Question Text'] = df['Question Text'].fillna(df['Variable Label'])
# On supprime les colonnes redondantes
df = df.drop(["Title", "Variable Label", "concept2_validé", "concept3_validé"], axis=1)
# On renomme les colonnes
df = df.rename(columns={df.columns[2]: 'text', df.columns[3]: 'label'})
# On garde 6 exemples max par concept
# df = df.groupby("label").head(6)
# Nettoyage spaCy
spacy.cli.download("fr_core_news_md")
nlp = spacy.load("fr_core_news_md")
new_stop_words = [",", ".", "?", "…", "(", ")", ":", "-"]
nlp.Defaults.stop_words.update(new_stop_words)
for word in new_stop_words:
nlp.vocab[word].is_stop = True
def remove_stop_words_spacy(text):
# First normalize string to remove any unicode characters
text = unicodedata.normalize('NFKD', text)
if pd.isna(text):
return text
doc = nlp(text)
return " ".join([token.lemma_ for token in doc if not token.is_stop])
df['text'] = df['text'].apply(remove_stop_words_spacy)
# Filtrer les labels avec moins de 2 exemples
label_counts = df['label'].value_counts()
valid_labels = label_counts[label_counts >= 2].index
df = df[df['label'].isin(valid_labels)]
return Dataset.from_pandas(df[["text", "label"]])
def model_init(params):
params = params or {}
max_iter = params.get("max_iter", 100)
solver = params.get("solver", "liblinear")
return SetFitModel.from_pretrained(
"sentence-transformers/paraphrase-multilingual-mpnet-base-v2",
head_params={"max_iter": max_iter, "solver": solver},
)
def hp_space(trial: Trial):
return {
"body_learning_rate": trial.suggest_float("body_learning_rate", 1e-6, 1e-3, log=True),
"num_epochs": trial.suggest_int("num_epochs", 1, 3),
"batch_size": trial.suggest_categorical("batch_size", [8, 16, 32]),
"seed": trial.suggest_int("seed", 1, 40),
"max_iter": trial.suggest_int("max_iter", 50, 300),
"solver": trial.suggest_categorical("solver", ["newton-cg", "lbfgs"]),
}
def main():
dataset_hf = preparer_donnees()
# Split train / eval
dataset_hf = dataset_hf.class_encode_column("label")
split_dataset = dataset_hf.train_test_split(test_size=0.2, seed=42, stratify_by_column="label")
train_dataset = split_dataset["train"]
eval_dataset = split_dataset["test"]
# Trainer avec model_init
trainer = Trainer(
train_dataset=train_dataset,
eval_dataset=eval_dataset,
model_init=model_init,
)
# Recherche d'hyperparamètres
best_run = trainer.hyperparameter_search(direction="maximize", hp_space=hp_space, n_trials=10)
print(best_run)
# Réentraînement final
trainer.apply_hyperparameters(best_run.hyperparameters, final_model=True)
trainer.train()
# Sauvegarde
trainer.model.save_pretrained("mon_modele_setfit_final")
print("Modèle sauvegardé dans 'mon_modele_setfit_final'")
if __name__ == "__main__":
main()