পাঠ ৪২ · ৪৫-এর মধ্যে · মডিউল ৫

Home / AI Courses / Machine Learning / Pipeline & Deployment

Pipeline ও Model Deployment

Pipelines & deployment — production ML

৭ মিনিট পড়া মাঝারি · Intermediate sklearn / FastAPI

এই পাঠে যা শিখবেন

Pipeline-এর core idea — leak prevention
ColumnTransformer — heterogeneous data
Cross-validation pipeline-এর ভেতরে — proper CV
FastAPI দিয়ে REST endpoint
Model versioning, monitoring, drift

১ · Why pipeline?

Naive ML code:

scaler = StandardScaler().fit(X_train)
X_train_s = scaler.transform(X_train)
X_test_s = scaler.transform(X_test)
model = LogisticRegression().fit(X_train_s, y_train)
preds = model.predict(X_test_s)

কাজ করে — কিন্তু dangerous। CV-এ যদি accidentally fit_transform(X) use করেন (full data), test info leak। 100 line-এর এর project-এ এই bug সাধারণ।

Pipeline: preprocessing + model একসাথে। CV-এ each fold-এ pipeline পুরো refit। Leak প্রায় impossible।

Pipeline = leak-proof + reproducible

Pipeline একটি single object যা fit() ও predict() support করে। Train data দিয়ে fit, test data দিয়ে predict — পুরো preprocessing-চেইন automatic apply। CV-এ প্রতিটি fold-এ separate fit। Production-এ একই pipeline serialize ও deploy।

২ · Basic Pipeline

Python · sklearn

from sklearn.pipeline import Pipeline
from sklearn.preprocessing import StandardScaler
from sklearn.linear_model import LogisticRegression
from sklearn.datasets import make_classification
from sklearn.model_selection import cross_val_score

X, y = make_classification(n_samples=500, n_features=20, random_state=0)

pipe = Pipeline([
    ("scale", StandardScaler()),
    ("clf",   LogisticRegression(max_iter=1000)),
])

# CV proper — each fold scaler fit fresh
scores = cross_val_score(pipe, X, y, cv=5, scoring="roc_auc")
print(f"CV AUC: {scores.mean():.3f} ± {scores.std():.3f}")

pipe.fit(X, y)
print(f"Full data score: {pipe.score(X, y):.3f}")

Pipeline একটি object — CV-এ প্রতি fold-এর জন্য fresh fit। Step-গুলো named (scale, clf) — hyperparameter access সহজ (scale__with_mean, clf__C)।

৩ · ColumnTransformer — heterogeneous data

বাস্তব data — numeric, categorical, text একসাথে। Different transformation দরকার:

Python · sklearn

import pandas as pd
import numpy as np
from sklearn.compose import ColumnTransformer
from sklearn.preprocessing import StandardScaler, OneHotEncoder
from sklearn.impute import SimpleImputer
from sklearn.pipeline import Pipeline
from sklearn.ensemble import RandomForestClassifier

# Synthetic Bangladesh loan dataset
np.random.seed(0)
n = 1000
df = pd.DataFrame({
    "age":         np.random.randint(20, 65, n),
    "income":      np.random.lognormal(10, 0.6, n),
    "education":   np.random.choice(["primary", "secondary", "graduate"], n),
    "occupation":  np.random.choice(["farmer", "service", "business", "other"], n),
    "credit_history": np.random.choice([0, 1, np.nan], n, p=[0.3, 0.6, 0.1]),
})
y = (df["income"] > df["income"].median()).astype(int).values

numeric_features = ["age", "income"]
categorical_features = ["education", "occupation"]

numeric_pipe = Pipeline([
    ("impute", SimpleImputer(strategy="median")),
    ("scale", StandardScaler()),
])
categorical_pipe = Pipeline([
    ("impute", SimpleImputer(strategy="most_frequent")),
    ("onehot", OneHotEncoder(handle_unknown="ignore")),
])

preprocessor = ColumnTransformer([
    ("num", numeric_pipe, numeric_features),
    ("cat", categorical_pipe, categorical_features),
])

full_pipe = Pipeline([
    ("preprocess", preprocessor),
    ("clf", RandomForestClassifier(n_estimators=100, random_state=0)),
])

full_pipe.fit(df, y)
print(f"Pipeline trained — score: {full_pipe.score(df, y):.3f}")
print(f"Number of features after preprocessing: {full_pipe[:-1].transform(df).shape[1]}")

ColumnTransformer numeric column-এ scale, categorical-এ one-hot — leak-free। নতুন data point arrive করলে — same pipeline-এ predict, manual preprocessing লাগবে না।

৪ · Pipeline + HPO + CV

L41-এর hyperparameter tuning Pipeline-এর সাথে natural। Hyperparameter name step_name__param format-এ:

Python · sklearn

from sklearn.model_selection import GridSearchCV

param_grid = {
    "clf__n_estimators": [50, 100, 200],
    "clf__max_depth":    [5, 10, None],
}

gs = GridSearchCV(full_pipe, param_grid, cv=3, n_jobs=-1, scoring="roc_auc")
gs.fit(df, y)
print(f"Best params: {gs.best_params_}")
print(f"Best CV score: {gs.best_score_:.4f}")

৫ · Model serialization

Python · joblib

import joblib

# Save trained pipeline
joblib.dump(full_pipe, "loan_model_v1.joblib")

# Later — production load
loaded = joblib.load("loan_model_v1.joblib")
new_applicant = pd.DataFrame([{
    "age": 35, "income": 50000,
    "education": "graduate", "occupation": "service",
    "credit_history": 1
}])
prob = loaded.predict_proba(new_applicant)[0, 1]
print(f"Default probability: {prob:.3f}")

joblib NumPy array efficient serialization-এ better than pickle। File size XGBoost ~১-১০০ MB। Save করলে — Python version, sklearn version compatible রাখুন।

ML lifecycle — training pipeline + serving + monitoring + retrain loop। Production system এই cycle-এই চলে।

৬ · FastAPI deployment

Python · FastAPI (api.py)

# pip install fastapi uvicorn joblib pandas
# api.py
from fastapi import FastAPI
from pydantic import BaseModel
import joblib
import pandas as pd

app = FastAPI(title="Loan Default API")
model = joblib.load("loan_model_v1.joblib")

class Applicant(BaseModel):
    age: int
    income: float
    education: str
    occupation: str
    credit_history: float | None = None

@app.get("/health")
def health():
    return {"status": "ok", "model": "v1"}

@app.post("/predict")
def predict(app_data: Applicant):
    df = pd.DataFrame([app_data.dict()])
    prob = float(model.predict_proba(df)[0, 1])
    decision = "deny" if prob > 0.5 else "approve"
    return {"default_probability": round(prob, 3),
            "decision": decision}

# Run: uvicorn api:app --host 0.0.0.0 --port 8000
# Test: curl -X POST http://localhost:8000/predict \
#   -H "Content-Type: application/json" \
#   -d '{"age":35,"income":50000,"education":"graduate","occupation":"service","credit_history":1}'

FastAPI + Pydantic — type-safe, auto-doc (Swagger UI at /docs)। Production-এ Docker-এ wrap, Kubernetes-এ deploy, gunicorn worker দিয়ে scale।

৭ · Deployment patterns

Real-time REST: single request, <১০০ms latency। FastAPI, Flask।
Batch: cron job, রাতে ১০ লাখ user score। Spark, Airflow।
Streaming: Kafka consumer, transaction fraud realtime।
Edge: mobile app, IoT — TFLite, ONNX, quantized model।
Serverless: AWS Lambda, Cloud Functions — autoscale, pay-per-use।

৮ · Monitoring ও drift

Production model টিকে থাকে না — world বদলায়:

Data drift: input distribution shift। COVID-এ shopping pattern বদলে গেছে।
Concept drift: input-output relationship বদলে। Fraud pattern evolve।
Label drift: target distribution shift।

Detection:

KS test, PSI (Population Stability Index) — distribution shift।
Performance monitor — accuracy/AUC track over time।
Statistical alarm — threshold breach।

Tooling:

MLflow: experiment tracking + model registry।
Evidently: drift detection।
Weights & Biases: experiment + monitoring।
Prometheus + Grafana: metrics dashboard।

৯ · Versioning

Model file — semantic versioning (v1.2.3)।
Training data hash।
Code git commit।
Hyperparameter snapshot।
Environment requirements.txt।
MLflow ও DVC — full lineage tracking।

Production ML-এ pipeline-এর ভেতরে imputation, encoding, scaling, model — সব। আলাদা step-এ করে save করলে inference-এ মিস হবে। "Pipeline as single object" — production rule #1।

ভাবনার প্রশ্ন

প্রতিটি প্রশ্ন নিজে কিছুক্ষণ ভাবুন — তারপর "→ উত্তর" চাপুন।

প্র ০১ Data leak — সবচেয়ে common ML bug। Subtle examples ও prevention strategies?

ML-এর silent killer। Result fool the analyst, fail in production।

Common leak types:

(১) Preprocessing leak:

StandardScaler full data fit।
Test mean ও std contributing।
CV optimistic।
Solution: Pipeline।

(২) Target encoding leak:

Categorical → mean target।
Full data target use।
Solution: K-fold target encoding।

(৩) Feature engineering leak:

"Customer's average past order" — uses future।
Solution: time-based feature, only past data।

(৪) Time leak:

Random CV on time series।
Future predict past।
Solution: TimeSeriesSplit।

(৫) Group leak:

Same patient train ও test।
GroupKFold use।

(৬) Label leak:

Feature directly contains label info।
"Loan_status_recent" predict default।
Solution: domain check।

(৭) Imputation leak:

Mean impute full data।
Test influences train mean।
Solution: Pipeline imputer।

Detection:

CV vs test gap — >5% suspect।
Feature importance — leakage feature dominate।
Time-based split, hold-out test।
Code review — preprocessing order।

Prevention checklist:

Pipeline always।
Time-aware split if temporal।
Group-aware if patient/user repeat।
Feature creation — only past data।
Hold-out test ZERO leakage।
Reviewer — separate person।

Bangladesh examples:

Loan default — applicant's most recent payment-প্রসঙ্গে — leak (label-related)।
Customer churn — "last activity date" feature — leak if > cutoff।
Medical diagnosis — same patient revisit — group leak।

মূল উপলব্ধি: Leak silent। Pipeline + time/group split + sanity check — defense।

প্র ০২ Pickle vs joblib vs ONNX vs TorchScript — model serialization-এর comparison?

Production deployment-এ format choice critical।

Pickle:

Python standard library।
Any Python object।
Pros: simple, ubiquitous।
Cons: Python-version-dependent, security (arbitrary code execute)।
Use: quick local prototype।

Joblib:

Pickle wrapper, NumPy array efficient।
Compression।
sklearn's recommendation।
Same Python-version concerns।

ONNX:

Open Neural Network Exchange।
Framework-agnostic (sklearn, PyTorch, TF, XGBoost)।
Inference runtime — fast (C++ optimized)।
Cross-language (C#, Java, Rust deploy)।
Hardware-specific (CPU, GPU, NPU)।
Cons: not all sklearn supported।

TorchScript:

PyTorch-specific।
Trace বা script PyTorch model।
C++ deploy via libtorch।
Mobile (LibTorch mobile)।

TensorFlow SavedModel:

TF ecosystem।
Serving framework (TF Serving)।
TFLite mobile।

PMML:

XML-based, classical ML।
Enterprise (SAS, IBM)।
Limited modern model।

Comparison:

sklearn pipeline: joblib (default), ONNX (production scaling)।
PyTorch: TorchScript বা ONNX।
TF/Keras: SavedModel বা ONNX।
XGBoost: native binary, ONNX, joblib।

Production decision:

(১) Latency critical:

ONNX runtime (C++ optimized)।
~10-100× faster than Python predict।

(২) Cross-platform:

ONNX — Java, C#, JS।
Mobile — TFLite।

(৩) Python-only:

Joblib simple, sufficient।

(৪) Versioning:

MLflow standard format-agnostic।
Save metadata + serialized।

Common pitfalls:

sklearn version mismatch — joblib break।
Custom transformer — pickle path issue।
ONNX export — unsupported op।
Security — never load untrusted pickle।

Modern stack:

Train: PyTorch/sklearn।
Export: ONNX।
Deploy: ONNX Runtime / Triton।
Monitor: MLflow / Weights & Biases।

Bangladesh case:

Local startup — joblib + FastAPI sufficient।
Bank — ONNX for compliance, audit trail।
Mobile app — TFLite for offline।

মূল উপলব্ধি: Format depends on deployment target। Joblib quick + Python-only। ONNX production-grade portable। MLflow integrate everything।

প্র ০৩ Drift detection — কীভাবে set up করবেন? False alarm avoid?

Production ML-এর continuous health check।

Drift types:

Data drift: input distribution change।
Concept drift: input-output relationship change।
Label drift: target distribution change।

Detection methods:

(১) Statistical tests:

KS test — continuous distribution।
Chi-square — categorical।
PSI (Population Stability Index) — banking standard।
Wasserstein distance।

(২) Performance metrics:

AUC, F1 monitor।
Lag — true label arrive late।
Proxy metric (e.g., click-through rate)।

(৩) Prediction monitoring:

Output distribution shift।
"Model approving more loans?" — flag।
Confidence score distribution।

(৪) Feature attribution:

SHAP value distribution।
Important feature importance change।
Concept drift signal।

Implementation:

Reference window — training data baseline।
Detection window — recent production।
Compute statistic per feature।
Threshold (PSI > 0.2 = warning, > 0.25 = action)।
Alert system।

False alarm avoid:

Multiple metric agreement।
Sufficient sample size (> 1000)।
Smooth window (rolling 7-day)।
Domain priors — seasonality expected।
Effect size metric (statistical significance + practical)।

Tools:

Evidently — Python-based, dashboards।
WhyLabs — managed service।
NannyML — performance monitoring।
AWS SageMaker Model Monitor।
Custom Prometheus metrics।

Action triggers:

Mild drift — investigate।
Moderate — retrain।
Severe — rollback to known good।
Continuous — automated retrain pipeline।

Bangladesh fintech case:

COVID-19 — credit behavior shift।
Eid season — spending spike।
Political event — payment delay।
Seasonal expected vs anomaly distinguish।

Retraining strategy:

Schedule (monthly/quarterly)।
Trigger-based (drift detect)।
Champion-challenger A/B।
Shadow mode evaluate।

Operational:

On-call rotation।
Runbook।
Escalation।
Postmortem।

Pitfalls:

Alarm fatigue — too sensitive।
Sampling bias — production data filtered।
Definition drift — feature-engineering change masks model drift।

মূল উপলব্ধি: Drift inevitable। Multi-signal detection + threshold + action pipeline। Bangladesh-এ — economic event-এ দ্রুত adapt-able system competitive advantage।

প্র ০৪ Bangladesh-এ একটি মাঝারি startup-এ ML deployment — সীমিত resources-এ pragmatic stack?

Real-world resource-constrained scenario।

Constraints:

2-3 ML engineer।
Limited cloud budget।
Python expertise।
Need fast iteration।

(১) Training environment:

Local laptop (CPU/Mac M2)।
Google Colab Pro — GPU on demand।
Kaggle Notebooks — free TPU।
One GCP/AWS instance — heavy training।

(২) Code & version control:

GitHub free/Pro।
DVC for data version।
poetry/pip for environment।

(৩) Experiment tracking:

MLflow — open-source, self-host।
Weights & Biases — free tier।
Neptune — alternative।

(৪) Pipeline:

sklearn Pipeline (single object)।
Pydantic schema input validation।
FastAPI REST।
Joblib serialize।

(৫) Deployment:

Docker container।
Cloud Run (GCP) / App Runner (AWS) / Render — pay-per-use।
Cloudflare Workers AI — edge inference।
VPS (DigitalOcean) — cheap fallback।

(৬) Database:

PostgreSQL (Supabase free)।
SQLite — small scale।
Redis — cache predictions।

(৭) Monitoring:

Custom Prometheus metrics।
Grafana cloud free tier।
Sentry error tracking।
Custom Evidently dashboard।

(৮) CI/CD:

GitHub Actions free।
Test on push।
Auto-deploy on main।
Slack notification।

(৯) Authentication:

API key for clients।
Rate limiting।
HTTPS via Cloudflare।

(১০) Cost optimization:

Spot instances।
Caching predictions (Redis)।
Batch inference for non-realtime।
Model size: distillation, quantization।

Sample architecture:

Frontend (React) → Cloudflare → FastAPI (Cloud Run) → Postgres।
Async batch (Celery + Redis) → ML jobs।
Model registry (MLflow self-hosted)।
Monitoring (Grafana cloud)।

Avoid:

Kubernetes — overkill জন্য startup।
SageMaker — vendor lock + cost।
Custom infra — engineering time waste।

Scaling roadmap:

0-100 user/day: single VM, joblib।
1K user/day: Cloud Run, Postgres।
100K user/day: Kubernetes, ONNX Runtime।
1M+ user/day: managed ML (SageMaker/Vertex AI)।

Bangladesh-specific:

Hosting in Singapore region (lower latency)।
BDT pricing visibility।
bKash/SSL Commerz integration।
Compliance — data privacy law।

Team workflow:

Weekly sprint।
Monthly model review।
Quarterly architecture review।
On-call rotation (২ engineer)।

মূল উপলব্ধি: Startup ML — simplicity > sophistication। FastAPI + Cloud Run + Postgres + MLflow — 80%+ use case cover। Scale পরে, দরকার হলে।

অনুশীলন

হিসাব করুন: Pipeline-এ scaler ও clf step। Hyperparameter tune করতে চান scaler__with_mean ও clf__C। Grid কেমন?
```
param_grid = {
    "scaler__with_mean": [True, False],
    "clf__C": [0.1, 1, 10],
}
# 2 × 3 = 6 combinations
GridSearchCV(pipe, param_grid, cv=5)
```
Pipeline-এ চেষ্টা: CSV থেকে data, mixed numeric+categorical, full pipeline।

উপরের ColumnTransformer example ব্যবহার করুন। সাথে CV score, hold-out test।
ভাবুন: Bangladesh-এর একটি e-commerce platform recommendation deploy করছে। Pipeline-এ কী include? Drift কীভাবে?
- Pipeline: user feature + product feature + interaction history → embedding → ranking model।
- Components: imputation, scaling, embedding lookup, NN model, post-processing।
- Versioning: daily train, weekly deploy।
- Drift: CTR shift detect, seasonal pattern (Eid), new product cold start।
- Monitor: top-k accuracy, diversity, novelty।
- A/B test: new vs old recommender।
- Latency: ONNX runtime, embedding cache।
- Privacy: aggregated logs, user opt-out।

এই পাঠে যা শিখবেন

১ · Why pipeline?

২ · Basic Pipeline

৩ · ColumnTransformer — heterogeneous data

৪ · Pipeline + HPO + CV

৫ · Model serialization

৬ · FastAPI deployment

৭ · Deployment patterns

৮ · Monitoring ও drift

৯ · Versioning

ভাবনার প্রশ্ন

অনুশীলন

আরও পড়ুন · ABCL TECH-এ আপনার পরবর্তী পদক্ষেপ