Concepts 🚀 Project 🧮 Labs

Load a .pkl model file (joblib + pickle)

The first step of any benchmarking engine: load the model the user uploaded. joblib.load() and pickle.loads() both work.

script.py v3.1

import joblib
import pickle
import io
import numpy as np
from sklearn.linear_model import LinearRegression
from sklearn.datasets import make_regression

# ── Step 1: Train a dummy model so we have something to save ──
X, y = make_regression(n_samples=100, n_features=3, noise=5, random_state=42)
model = LinearRegression().fit(X, y)
print('Trained:', type(model).__name__)

# ── Step 2: Save it as a .pkl file ──
joblib.dump(model, 'model.pkl')
print('Saved -> model.pkl')

# ── Step 3: Load it back TWO ways ──

# Way 1: joblib.load (preferred for sklearn models)
model_v1 = joblib.load('model.pkl')
print('Loaded via joblib:', type(model_v1).__name__)

# Way 2: pickle.loads (works on bytes — what you get from UploadFile.read())
with open('model.pkl', 'rb') as f:
    file_bytes = f.read()
model_v2 = pickle.loads(file_bytes)
print('Loaded via pickle.loads:', type(model_v2).__name__)

# ── Step 4: Verify both loaded models predict the same thing ──
X_new = np.array([[1.5, -0.5, 2.0]])
pred1 = model_v1.predict(X_new)[0]
pred2 = model_v2.predict(X_new)[0]
print(f'joblib prediction:  {pred1:.4f}')
print(f'pickle prediction:  {pred2:.4f}')
print(f'Same? {abs(pred1 - pred2) < 1e-10}')

Python output click Run to execute

Click "Run" to execute the code on the left.
The output will appear here.

🧠 How it works

Two ways to load a .pkl model: joblib.load('model.pkl') — loads from a file path. Faster for numpy arrays. pickle.loads(bytes) — loads from a bytes object. This is what you use when the model comes from an HTTP upload: `file_bytes = await upload_file.read()` then `model = pickle.loads(file_bytes)`. Both produce the same object — the trained model with all its coefficients. WARNING: pickle is unsafe for untrusted files — it can execute arbitrary code. In production, run pickle.loads in a sandbox.

✏️ Try these changes

Change make_regression n_features from 3 to 5
→ model trains with 5 features instead of 3
Change the save path from 'model.pkl' to 'my_model.pkl'
→ saves to a different file
Add a print(model.coef_) after loading
→ shows the learned coefficients
Replace LinearRegression with from sklearn.ensemble import RandomForestRegressor
→ different model type — same load code works

Next lab →

Run predict() on a loaded model