tag
This commit is contained in:
parent
ee9fe1bca2
commit
92032f67a4
BIN
Rapport2.pdf
BIN
Rapport2.pdf
Binary file not shown.
445
Rapport2.typ
445
Rapport2.typ
@ -1,445 +0,0 @@
|
|||||||
// Using native Typst table instead of tablex for compatibility
|
|
||||||
|
|
||||||
#set document(title: "Projet CS:GO - Pipeline MLOps", author: "Équipe MLOps")
|
|
||||||
#set page(margin: 2cm, numbering: "1")
|
|
||||||
#set text(size: 11pt)
|
|
||||||
#set heading(numbering: "1.1")
|
|
||||||
|
|
||||||
#align(center)[
|
|
||||||
#text(18pt, weight: "bold")[Projet CS:GO Esports Intelligence Platform]
|
|
||||||
#v(0.5cm)
|
|
||||||
#text(14pt)[Pipeline MLOps et Stratégie de Monitoring]
|
|
||||||
#v(0.3cm)
|
|
||||||
#line(length: 100%)
|
|
||||||
#v(0.5cm)
|
|
||||||
|
|
||||||
#grid(
|
|
||||||
columns: (1fr, 1fr),
|
|
||||||
[*Équipe : Paul Roost, Axelle Desthombes, Alexis Bruneteau* ], [*Date :* #datetime.today().display()]
|
|
||||||
)
|
|
||||||
|
|
||||||
#v(0.2cm)
|
|
||||||
*Dataset :* CS:GO Professional Matches (Kaggle - 25K+ matches) \
|
|
||||||
*Objectif :* Prédiction des résultats de matchs et optimisation des stratégies esports
|
|
||||||
]
|
|
||||||
|
|
||||||
#v(1cm)
|
|
||||||
|
|
||||||
= Atelier 1 : Pipeline du Fil Rouge
|
|
||||||
|
|
||||||
== Architecture Générale du Pipeline
|
|
||||||
|
|
||||||
#figure(
|
|
||||||
image("images/pipeline2.svg", width: 60%),
|
|
||||||
caption: [Architecture complète du pipeline MLOps CS:GO]
|
|
||||||
) <pipeline-arch>
|
|
||||||
|
|
||||||
|
|
||||||
|
|
||||||
== Étapes Détaillées du Pipeline
|
|
||||||
|
|
||||||
=== Collecte et Ingestion des Données
|
|
||||||
|
|
||||||
*Sources de données :*
|
|
||||||
- *HLTV.org* : Résultats historiques, classements équipes
|
|
||||||
- *Steam API* : Données joueurs en temps réel
|
|
||||||
- *Tournament APIs* : Calendriers, formats de compétition
|
|
||||||
|
|
||||||
*Pipeline d'ingestion automatisé avec Apache Airflow :*
|
|
||||||
|
|
||||||
```python
|
|
||||||
@dag(schedule_interval="@hourly", start_date=datetime(2024,1,1))
|
|
||||||
def csgo_data_ingestion():
|
|
||||||
|
|
||||||
extract_hltv_matches = PythonOperator(
|
|
||||||
task_id='extract_hltv',
|
|
||||||
python_callable=scrape_hltv_matches
|
|
||||||
)
|
|
||||||
|
|
||||||
validate_data = PythonOperator(
|
|
||||||
task_id='validate_raw_data',
|
|
||||||
python_callable=validate_match_schema
|
|
||||||
)
|
|
||||||
|
|
||||||
store_s3 = PythonOperator(
|
|
||||||
task_id='store_to_s3',
|
|
||||||
python_callable=upload_to_s3
|
|
||||||
)
|
|
||||||
|
|
||||||
extract_hltv_matches >> validate_data >> store_s3
|
|
||||||
```
|
|
||||||
|
|
||||||
=== Feature Engineering Multi-Niveaux
|
|
||||||
|
|
||||||
#table(
|
|
||||||
columns: (2fr, 3fr),
|
|
||||||
stroke: 0.5pt,
|
|
||||||
[*Catégorie*], [*Features*],
|
|
||||||
[*Team-level*], [
|
|
||||||
• `recent_form_10_matches` - Ratio W/L récent \
|
|
||||||
• `map_pool_strength` - Win rate par map \
|
|
||||||
• `clutch_success_rate` - Performance clutch \
|
|
||||||
• `eco_round_conversion` - Gestion économique
|
|
||||||
],
|
|
||||||
[*Context*], [
|
|
||||||
• `tournament_tier` - Prestige de l'événement \
|
|
||||||
• `prize_pool_amount` - Facteur de pression \
|
|
||||||
• `head_to_head_record` - Historique direct \
|
|
||||||
• `current_game_patch` - Version meta game
|
|
||||||
],
|
|
||||||
[*Live*], [
|
|
||||||
• `current_score_difference` - Score en cours \
|
|
||||||
• `momentum_last_5_rounds` - Élan récent \
|
|
||||||
• `economy_advantage` - Avantage économique
|
|
||||||
]
|
|
||||||
)
|
|
||||||
|
|
||||||
=== Entraînement Multi-Target
|
|
||||||
|
|
||||||
Architecture d'apprentissage multitâche avec PyTorch :
|
|
||||||
|
|
||||||
```python
|
|
||||||
class CSGOPredictor(nn.Module):
|
|
||||||
def __init__(self, input_dim):
|
|
||||||
super().__init__()
|
|
||||||
self.shared_layers = nn.Sequential(
|
|
||||||
nn.Linear(input_dim, 256),
|
|
||||||
nn.ReLU(),
|
|
||||||
nn.Dropout(0.3),
|
|
||||||
nn.Linear(256, 128)
|
|
||||||
)
|
|
||||||
|
|
||||||
# Têtes spécialisées par tâche
|
|
||||||
self.match_winner = nn.Linear(128, 2) # Classification binaire
|
|
||||||
self.final_score = nn.Linear(128, 2) # Régression scores
|
|
||||||
self.total_maps = nn.Linear(128, 4) # Nombre de maps
|
|
||||||
|
|
||||||
def forward(self, x):
|
|
||||||
shared_repr = self.shared_layers(x)
|
|
||||||
return {
|
|
||||||
'match_winner': self.match_winner(shared_repr),
|
|
||||||
'final_score': self.final_score(shared_repr),
|
|
||||||
'total_maps': self.total_maps(shared_repr)
|
|
||||||
}
|
|
||||||
```
|
|
||||||
|
|
||||||
== Automatisation et Points de Contrôle
|
|
||||||
|
|
||||||
=== Stratégie d'Automatisation
|
|
||||||
|
|
||||||
#table(
|
|
||||||
columns: (2fr, 1fr, 3fr),
|
|
||||||
stroke: 0.5pt,
|
|
||||||
[*Étape*], [*Status*], [*Justification*],
|
|
||||||
[*Ingestion données*], [AUTO], [Nouveaux matchs quotidiens, obsolescence rapide],
|
|
||||||
[*Feature Engineering*], [AUTO], [Features dépendent de données temps-réel],
|
|
||||||
[*Model Retraining*], [AUTO], [Meta game évolue (patches, transferts)],
|
|
||||||
[*Deployment*], [AUTO], [Évite erreurs humaines, rollback rapide],
|
|
||||||
[*Model Selection*], [MANUEL], [Décisions business complexes nécessitant expertise]
|
|
||||||
)
|
|
||||||
|
|
||||||
=== Points de Contrôle Critiques
|
|
||||||
|
|
||||||
*Validation des Données :*
|
|
||||||
```python
|
|
||||||
def validate_match_data(df):
|
|
||||||
"""Validation avant feature engineering"""
|
|
||||||
checks = [
|
|
||||||
('schema_compliance', validate_schema(df)),
|
|
||||||
('completeness', check_missing_values(df, threshold=0.05)),
|
|
||||||
('consistency', validate_team_names(df)),
|
|
||||||
('freshness', check_data_age(df, max_hours=24)),
|
|
||||||
('volume', validate_daily_match_count(df, min_matches=50))
|
|
||||||
]
|
|
||||||
|
|
||||||
for check_name, result in checks:
|
|
||||||
if not result.passed:
|
|
||||||
raise DataValidationError(f"{check_name} failed")
|
|
||||||
```
|
|
||||||
|
|
||||||
*Validation des Performances :*
|
|
||||||
```python
|
|
||||||
def validate_model_performance(model, validation_data):
|
|
||||||
"""Validation avant déploiement"""
|
|
||||||
metrics = evaluate_model(model, validation_data)
|
|
||||||
|
|
||||||
# Seuils minimaux
|
|
||||||
assert metrics['accuracy'] > 0.65, "Accuracy insuffisante"
|
|
||||||
assert metrics['roi_betting'] > 1.05, "ROI non profitable"
|
|
||||||
assert metrics['upset_detection'] > 0.20, "Détection upsets faible"
|
|
||||||
|
|
||||||
return True
|
|
||||||
```
|
|
||||||
|
|
||||||
=== Difficultés Techniques et Solutions
|
|
||||||
|
|
||||||
*Défi 1 : Concept Drift Extrême*
|
|
||||||
|
|
||||||
Les mises à jour du jeu modifient significativement les stratégies et l'équilibre, ce qui peut rendre les modèles existants moins performants.
|
|
||||||
|
|
||||||
*Solution :* Détection automatisée de drift + retraining d'urgence
|
|
||||||
```python
|
|
||||||
def detect_meta_shift(recent_matches, baseline):
|
|
||||||
"""Détecte changements post-patch"""
|
|
||||||
map_rates = calculate_map_win_rates(recent_matches)
|
|
||||||
baseline_rates = baseline['map_win_rates']
|
|
||||||
|
|
||||||
for map_name in map_rates:
|
|
||||||
ks_stat, p_value = ks_2samp(map_rates[map_name],
|
|
||||||
baseline_rates[map_name])
|
|
||||||
if p_value < 0.01: # Drift significatif
|
|
||||||
return True
|
|
||||||
return False
|
|
||||||
```
|
|
||||||
|
|
||||||
*Défi 2 : Cold Start Problem*
|
|
||||||
|
|
||||||
Les nouvelles équipes ou changements de composition ne disposent pas d'historique suffisant pour l'entraînement.
|
|
||||||
|
|
||||||
*Solution :* Transfer learning via embeddings joueurs
|
|
||||||
```python
|
|
||||||
def handle_cold_start_team(roster, player_db):
|
|
||||||
"""Prédictions via similarité joueurs"""
|
|
||||||
team_embedding = [player_db.get_embedding(p.id) for p in roster]
|
|
||||||
similar_teams = find_similar_teams(team_embedding, top_k=5)
|
|
||||||
return weighted_prediction_from_similar(similar_teams)
|
|
||||||
```
|
|
||||||
|
|
||||||
#pagebreak()
|
|
||||||
|
|
||||||
= Atelier 2 : Expériences et Monitoring
|
|
||||||
|
|
||||||
== Tracking des Expériences avec MLflow
|
|
||||||
|
|
||||||
=== Configuration et Logging Structuré
|
|
||||||
|
|
||||||
```python
|
|
||||||
mlflow.set_tracking_uri("http://mlflow-server:5000")
|
|
||||||
mlflow.set_experiment("csgo-match-prediction")
|
|
||||||
|
|
||||||
def train_and_log_experiment(config):
|
|
||||||
with mlflow.start_run(run_name=f"csgo-v{config.version}"):
|
|
||||||
|
|
||||||
# Hyperparamètres
|
|
||||||
mlflow.log_params({
|
|
||||||
"model_type": config.model_type,
|
|
||||||
"learning_rate": config.lr,
|
|
||||||
"batch_size": config.batch_size,
|
|
||||||
"data_version": config.data_version
|
|
||||||
})
|
|
||||||
|
|
||||||
# Métriques par époque
|
|
||||||
for epoch in range(config.epochs):
|
|
||||||
train_loss = train_one_epoch(model, train_loader)
|
|
||||||
val_metrics = evaluate_model(model, val_loader)
|
|
||||||
|
|
||||||
mlflow.log_metrics({
|
|
||||||
"train_loss": train_loss,
|
|
||||||
"val_accuracy": val_metrics['accuracy'],
|
|
||||||
"betting_roi": val_metrics['roi'],
|
|
||||||
"upset_detection": val_metrics['upset_rate']
|
|
||||||
}, step=epoch)
|
|
||||||
|
|
||||||
# Artefacts finaux
|
|
||||||
mlflow.pytorch.log_model(model, "model")
|
|
||||||
mlflow.log_artifacts("evaluation_plots/")
|
|
||||||
```
|
|
||||||
|
|
||||||
=== Métriques Trackées
|
|
||||||
|
|
||||||
#table(
|
|
||||||
columns: (2fr, 3fr),
|
|
||||||
stroke: 0.5pt,
|
|
||||||
[*Catégorie*], [*Métriques*],
|
|
||||||
[*Performance ML*], [
|
|
||||||
• Accuracy, Precision, Recall, F1-Score \
|
|
||||||
• ROC-AUC, Calibration Error \
|
|
||||||
• Performance par segment (tier tournoi)
|
|
||||||
],
|
|
||||||
[*Business*], [
|
|
||||||
• ROI betting, Profit/Loss \
|
|
||||||
• Sharpe Ratio, Upset Detection Rate \
|
|
||||||
• User Engagement, Revenue Impact
|
|
||||||
],
|
|
||||||
[*Computational*], [
|
|
||||||
• Training Time, Inference Latency \
|
|
||||||
• Model Size, Memory Usage \
|
|
||||||
• API Response Time
|
|
||||||
]
|
|
||||||
)
|
|
||||||
|
|
||||||
== Stratégie de Monitoring Complète
|
|
||||||
|
|
||||||
=== Métriques de Surveillance Multi-Niveaux
|
|
||||||
|
|
||||||
*Surveillance de la qualité des données :*
|
|
||||||
```python
|
|
||||||
class DataMonitoring:
|
|
||||||
def monitor_data_quality(self, new_batch):
|
|
||||||
metrics = {}
|
|
||||||
|
|
||||||
# Volume et couverture
|
|
||||||
metrics['daily_match_count'] = len(new_batch)
|
|
||||||
metrics['team_coverage'] = new_batch['team_name'].nunique()
|
|
||||||
|
|
||||||
# Qualité
|
|
||||||
metrics['missing_rate'] = new_batch.isnull().mean().mean()
|
|
||||||
metrics['duplicates'] = new_batch.duplicated().sum()
|
|
||||||
|
|
||||||
# Drift distribution
|
|
||||||
for col in ['team_ranking', 'match_duration']:
|
|
||||||
drift = calculate_drift_score(new_batch[col], baseline[col])
|
|
||||||
metrics[f'{col}_drift'] = drift
|
|
||||||
|
|
||||||
return metrics
|
|
||||||
```
|
|
||||||
|
|
||||||
*Model Performance Monitoring :*
|
|
||||||
```python
|
|
||||||
def monitor_model_performance(predictions, actuals):
|
|
||||||
"""Monitoring performance temps-réel"""
|
|
||||||
rolling_metrics = {}
|
|
||||||
|
|
||||||
# Fenêtres glissantes
|
|
||||||
for window in [1, 7, 30]: # jours
|
|
||||||
recent = get_recent_data(window)
|
|
||||||
rolling_metrics[f'accuracy_{window}d'] = accuracy_score(
|
|
||||||
recent['actual'], recent['predicted']
|
|
||||||
)
|
|
||||||
rolling_metrics[f'roi_{window}d'] = calculate_roi(
|
|
||||||
recent['predictions'], recent['outcomes']
|
|
||||||
)
|
|
||||||
|
|
||||||
return rolling_metrics
|
|
||||||
```
|
|
||||||
|
|
||||||
=== Système d'Alertes Intelligent
|
|
||||||
|
|
||||||
#table(
|
|
||||||
columns: (1fr, 2fr, 2fr),
|
|
||||||
stroke: 0.5pt,
|
|
||||||
[*Sévérité*], [*Seuils*], [*Actions*],
|
|
||||||
[*CRITIQUE*], [
|
|
||||||
• Accuracy 7j \< 60% \
|
|
||||||
• ROI 7j \< 100% \
|
|
||||||
• API errors \> 5%
|
|
||||||
], [
|
|
||||||
• PagerDuty + Slack \
|
|
||||||
• Email équipe oncall \
|
|
||||||
• Rollback automatique
|
|
||||||
],
|
|
||||||
[*WARNING*], [
|
|
||||||
• Accuracy trending ↓ \
|
|
||||||
• Concept drift p\<0.05 \
|
|
||||||
• Latency \> 300ms
|
|
||||||
], [
|
|
||||||
• Slack \#alerts \
|
|
||||||
• Email ML team \
|
|
||||||
• Investigation requise
|
|
||||||
],
|
|
||||||
[*INFO*], [
|
|
||||||
• Nouveaux tournaments \
|
|
||||||
• Performance updates \
|
|
||||||
• System health
|
|
||||||
], [
|
|
||||||
• Slack \#monitoring \
|
|
||||||
• Dashboard updates
|
|
||||||
]
|
|
||||||
)
|
|
||||||
|
|
||||||
=== Dashboards et Rapports
|
|
||||||
|
|
||||||
*Dashboard Temps-Réel (Grafana) :*
|
|
||||||
|
|
||||||
- *Model Performance* : Accuracy, ROI, Calibration trends
|
|
||||||
- *Data Pipeline Health* : Volume, freshness, quality scores
|
|
||||||
- *API Performance* : Latency P95, request rate, error rate
|
|
||||||
- *Business Metrics* : Revenue impact, user engagement
|
|
||||||
|
|
||||||
*Rapports Hebdomadaires Automatisés :*
|
|
||||||
|
|
||||||
```python
|
|
||||||
class WeeklyReportGenerator:
|
|
||||||
def generate_performance_report(self, week_start, week_end):
|
|
||||||
sections = [
|
|
||||||
self.executive_summary(), # KPIs clés
|
|
||||||
self.model_performance(), # Analyse détaillée
|
|
||||||
self.business_impact(), # Valeur générée
|
|
||||||
self.technical_health(), # Infrastructure
|
|
||||||
self.recommendations() # Actions recommandées
|
|
||||||
]
|
|
||||||
return self.compile_html_report(sections)
|
|
||||||
```
|
|
||||||
|
|
||||||
== Architecture de Monitoring Production
|
|
||||||
|
|
||||||
=== Alerting Multi-Canal
|
|
||||||
|
|
||||||
```python
|
|
||||||
class AlertManager:
|
|
||||||
def __init__(self):
|
|
||||||
self.channels = {
|
|
||||||
'slack': SlackNotifier(SLACK_WEBHOOK),
|
|
||||||
'email': EmailNotifier(EMAIL_CONFIG),
|
|
||||||
'pagerduty': PagerDutyNotifier(PAGERDUTY_KEY)
|
|
||||||
}
|
|
||||||
|
|
||||||
def send_alert(self, alert):
|
|
||||||
if alert['severity'] == 'CRITICAL':
|
|
||||||
// Alertes critiques sur tous les canaux
|
|
||||||
self.channels['pagerduty'].send(alert)
|
|
||||||
self.channels['slack'].send_critical(alert)
|
|
||||||
self.channels['email'].send_oncall(alert)
|
|
||||||
elif alert['severity'] == 'WARNING':
|
|
||||||
// Warnings vers Slack et email
|
|
||||||
self.channels['slack'].send_warning(alert)
|
|
||||||
self.channels['email'].send_team(alert)
|
|
||||||
```
|
|
||||||
|
|
||||||
=== Runbooks d'Incident
|
|
||||||
|
|
||||||
*Alerte Critique : Accuracy < 60%*
|
|
||||||
|
|
||||||
1. *Actions Immédiates (0-15min)*
|
|
||||||
- Vérifier qualité des données récentes
|
|
||||||
- Identifier changements meta/tournois
|
|
||||||
- Rollback si accuracy < 55%
|
|
||||||
|
|
||||||
2. *Investigation (15-60min)*
|
|
||||||
- Analyse drift sur données récentes
|
|
||||||
- Comparaison prédictions vs résultats
|
|
||||||
- Validation pipeline features
|
|
||||||
|
|
||||||
3. *Résolution (1-4h)*
|
|
||||||
- Retraining d'urgence si drift détecté
|
|
||||||
- Fix pipeline si problème data quality
|
|
||||||
- Rollback si problème infrastructure
|
|
||||||
|
|
||||||
= Conclusion
|
|
||||||
|
|
||||||
L'architecture MLOps développée pour ce projet CS:GO présente plusieurs caractéristiques importantes :
|
|
||||||
|
|
||||||
*Architecture de production robuste :*
|
|
||||||
- Apprentissage multi-tâches permettant des prédictions variées selon les besoins métier
|
|
||||||
- Service en temps réel respectant les contraintes de latence
|
|
||||||
- Gestion de la dérive conceptuelle liée à l'évolution du meta-jeu
|
|
||||||
- Surveillance complète des données, modèles et métriques business
|
|
||||||
|
|
||||||
*Mesure de la valeur métier :*
|
|
||||||
- Suivi du retour sur investissement pour les applications de paris et fantasy leagues
|
|
||||||
- Métriques d'engagement utilisateur pour optimiser la rétention
|
|
||||||
- Impact sur le chiffre d'affaires pour justifier les investissements
|
|
||||||
|
|
||||||
*Fiabilité opérationnelle :*
|
|
||||||
- Retour en arrière automatique en cas de dégradation des performances
|
|
||||||
- Système d'alertes multi-canaux pour une réaction rapide
|
|
||||||
- Procédures documentées pour la résolution d'incidents
|
|
||||||
- Plan de continuité d'activité pour les événements critiques
|
|
||||||
|
|
||||||
Ce travail démontre l'application des principes MLOps modernes à un domaine spécialisé, en mettant l'accent sur la création de valeur métier et la fiabilité opérationnelle.
|
|
||||||
|
|
||||||
#align(center)[
|
|
||||||
#line(length: 50%)
|
|
||||||
#v(0.3cm)
|
|
||||||
*Équipe MLOps - Projet CS:GO Intelligence Platform*
|
|
||||||
]
|
|
||||||
File diff suppressed because one or more lines are too long
|
Before Width: | Height: | Size: 40 KiB |
Loading…
x
Reference in New Issue
Block a user