Why Software Testers Should Learn TensorFlow in 2026

import tensorflow as tf
import numpy as np

model = tf.keras.models.load_model("spam_model.keras")
samples = np.array(["win a free phone now", "meeting moved to 3pm"])
scores = model.predict(tf.constant(samples))
print(scores)   # probabilities, e.g. [[0.97], [0.04]]
# assumes the saved model includes a text vectorisation layer

Questions a tester asks about an ML feature:
- What does 'correct' mean, and what accuracy is acceptable?
- Which mistakes are costly (false positives or negatives)?
- Does it work equally well for all user groups?
- What happens with empty, very long or unusual input?
- How will we know if quality drops after release?

import tensorflow as tf

precision = tf.keras.metrics.Precision()
recall = tf.keras.metrics.Recall()

y_true = [1, 0, 1, 1, 0, 1]
y_pred = [1, 0, 0, 1, 1, 1]   # model output after threshold 0.5
precision.update_state(y_true, y_pred)
recall.update_state(y_true, y_pred)
print("precision:", precision.result().numpy())   # 0.75
print("recall:", recall.result().numpy())         # 0.75

import pandas as pd
from sklearn.metrics import recall_score

df = pd.read_csv("test_predictions.csv")   # columns: group, actual, predicted
for group, part in df.groupby("group"):
    r = recall_score(part["actual"], part["predicted"])
    print(f"{group:10s} recall={r:.2f} rows={len(part)}")

# test_sentiment_behaviour.py - run with pytest
import numpy as np
import tensorflow as tf

model = tf.keras.models.load_model("sentiment_model.keras")

def score(text):
    return float(model.predict(tf.constant([text]), verbose=0)[0][0])

def test_invariance_to_names():
    a = score("Rahul said the service was excellent")
    b = score("Priya said the service was excellent")
    assert abs(a - b) < 0.1

def test_directional_negative_word():
    assert score("the food was good but service terrible") < score("the food was good")

def test_empty_input_does_not_crash():
    assert 0.0 <= score("") <= 1.0

# Add to the pipeline that publishes a new model
- name: Model behaviour tests
  run: pytest tests/test_sentiment_behaviour.py -q
# Block release if any behaviour test fails

ML Feature Test Plan: ______________
Model version: ____   Data version: ____
1. Data checks: schema, ranges, label quality, no train/test leak
2. Metrics: accuracy / precision / recall vs agreed thresholds
3. Slices: metrics per user group, region, device
4. Behaviour: invariance, directional, minimum functionality
5. Robustness: empty, long, unusual input
6. Integration: API response time and error handling
7. Monitoring plan after release

import pandas as pd

train = pd.read_csv("train.csv")
test = pd.read_csv("test.csv")
assert train["label"].isin([0, 1]).all(), "unexpected labels"
assert train["text"].str.len().gt(0).all(), "empty texts found"
overlap = set(train["text"]) & set(test["text"])
assert not overlap, f"{len(overlap)} test rows also in training data"

[ ] All metric thresholds met on the latest test set
[ ] No group below its slice threshold
[ ] Behaviour tests pass in CI
[ ] Model card / notes updated (data, limits, version)
[ ] Rollback to previous model version tested
[ ] Monitoring dashboard and alerts ready