How TensorFlow Helps You Land an Automation Tester Job Faster

Test types for an ML feature
1. Smoke: model loads, returns output of correct shape and range
2. Behaviour: invariance and directional expectations
3. Edge cases: empty, extreme, unusual inputs
4. Data checks: schema, ranges, missing values
5. Quality gate: metric on a fixed test set >= agreed floor
6. Integration: API returns predictions correctly

Acceptance criteria (example: product image classifier)
- Returns one of the 10 known categories
- Probabilities sum to 1 (within 0.001)
- Accuracy on the frozen test set >= the agreed floor
- Small brightness change does not change the top category
- Responds within the agreed time on the test environment

# test_model_smoke.py
import numpy as np
import pytest
import tensorflow as tf

@pytest.fixture(scope="session")
def model():
    return tf.keras.models.load_model("models/classifier.keras")

def test_output_shape_and_probabilities(model):
    x = np.random.default_rng(0).random((8, 28, 28)).astype("float32")
    probs = model.predict(x, verbose=0)
    assert probs.shape == (8, 10)
    assert np.all((probs >= 0) & (probs <= 1))
    assert np.allclose(probs.sum(axis=1), 1.0, atol=1e-3)

def test_same_input_same_output(model):
    x = np.ones((1, 28, 28), dtype="float32") * 0.5
    assert np.allclose(model.predict(x, verbose=0), model.predict(x, verbose=0))

# test_model_behaviour.py
import numpy as np

def test_small_brightness_change_keeps_prediction(model, sample_images):
    # sample_images: fixture with labelled test images (float32, 0-1)
    base = model.predict(sample_images, verbose=0).argmax(axis=1)
    brighter = np.clip(sample_images * 1.05, 0, 1)
    shifted = model.predict(brighter, verbose=0).argmax(axis=1)
    agreement = (base == shifted).mean()
    assert agreement >= 0.95, f"only {agreement:.0%} predictions stable"

import pandas as pd

EXPECTED = {"age": "int64", "income": "float64", "city": "object"}

def test_input_schema_and_ranges():
    df = pd.read_csv("data/scoring_batch.csv")
    for col, dtype in EXPECTED.items():
        assert col in df.columns, f"missing column {col}"
        assert str(df[col].dtype) == dtype, f"{col} has type {df[col].dtype}"
    assert df["age"].between(18, 100).all()
    assert (df["income"] >= 0).all()
    assert df["city"].notna().mean() >= 0.99

import numpy as np

METRIC_FLOOR = 0.90   # agreed with the data science team

def test_new_model_meets_floor_and_does_not_regress(model, current_model, test_set):
    x, y = test_set
    new_acc = (model.predict(x, verbose=0).argmax(axis=1) == y).mean()
    old_acc = (current_model.predict(x, verbose=0).argmax(axis=1) == y).mean()
    assert new_acc >= METRIC_FLOOR, f"accuracy {new_acc:.3f} below floor"
    assert new_acc >= old_acc - 0.01, "new model is worse than current"

import numpy as np
import pytest
import tensorflow as tf

@pytest.fixture(scope="session")
def model():
    return tf.keras.models.load_model("models/classifier.keras")

@pytest.fixture(scope="session")
def current_model():
    return tf.keras.models.load_model("models/classifier_current.keras")

@pytest.fixture(scope="session")
def test_set():
    data = np.load("data/frozen_test_set.npz")
    return data["x"], data["y"]

@pytest.fixture(scope="session")
def sample_images(test_set):
    return test_set[0][:200]

# GitHub Actions step for ML tests
- name: ML model tests
  run: |
    pip install -r requirements-test.txt
    pytest tests/ml -q --junitxml=reports/ml-tests.xml

ML feature test plan
Feature: product image categorisation
Data checks: schema, ranges, missing values
Smoke: load, shape, probability range, determinism
Behaviour: brightness and small-crop invariance
Quality gate: accuracy >= floor, no regression vs current
Integration: /predict API returns category and score
Sign-off: QA lead + data science lead