Introduction

In TensorFlow, predicting probabilities means asking the model for class confidence scores rather than only the final label. The exact code depends on whether your model already outputs probabilities, such as from softmax or sigmoid, or whether it outputs raw logits that still need a probability transform.

Binary Classification: sigmoid Output

For binary classification, the final layer often has one unit with a sigmoid activation. In that setup, model.predict already returns probabilities between zero and one.

1import tensorflow as tf
2import numpy as np
3
4model = tf.keras.Sequential([
5    tf.keras.layers.Dense(16, activation="relu"),
6    tf.keras.layers.Dense(1, activation="sigmoid")
7])
8
9model.compile(optimizer="adam", loss="binary_crossentropy", metrics=["accuracy"])
10
11x = np.array([[0.2, 0.8], [0.9, 0.1]], dtype="float32")
12probs = model.predict(x, verbose=0)
13
14print(probs)

Each output value is the model’s estimated probability of the positive class.

Multiclass Classification: softmax Output

For multiclass classification, the final layer often uses softmax. In that case, the output vector is already a probability distribution over classes.

1import tensorflow as tf
2import numpy as np
3
4model = tf.keras.Sequential([
5    tf.keras.layers.Dense(32, activation="relu"),
6    tf.keras.layers.Dense(3, activation="softmax")
7])
8
9model.compile(optimizer="adam", loss="sparse_categorical_crossentropy", metrics=["accuracy"])
10
11x = np.array([[1.0, 0.5, 0.2], [0.1, 0.2, 0.9]], dtype="float32")
12probs = model.predict(x, verbose=0)
13
14print(probs)
15print(np.sum(probs, axis=1))

Each row contains class probabilities that sum to one.

If the Model Outputs Logits Instead

Some models intentionally leave off the final activation and return logits. This is common when training with from_logits=True for numerical stability.

Example model:

1model = tf.keras.Sequential([
2    tf.keras.layers.Dense(32, activation="relu"),
3    tf.keras.layers.Dense(3)
4])
5
6model.compile(
7    optimizer="adam",
8    loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True),
9    metrics=["accuracy"]
10)

Now model.predict returns logits, not probabilities. Convert them with tf.nn.softmax:

1logits = model.predict(x, verbose=0)
2probs = tf.nn.softmax(logits, axis=-1)
3
4print(probs.numpy())

For binary logits, use tf.nn.sigmoid instead.

Check the Output Shape Before Using It

A quick shape check avoids many mistakes. Binary classifiers often return shape (batch_size, 1), while multiclass models return (batch_size, num_classes).

predictions = model.predict(x, verbose=0)
print(predictions.shape)

That tells you whether you should flatten a single-column result, run argmax, or apply a probability transform first.

Getting the Predicted Class and Its Probability

A probability vector is often most useful when paired with the winning class.

1probs = model.predict(x, verbose=0)
2predicted_class = np.argmax(probs, axis=1)
3predicted_confidence = np.max(probs, axis=1)
4
5print(predicted_class)
6print(predicted_confidence)

This is the usual pattern for top-1 classification output in an application.

Why Probabilities Matter

Probabilities let you apply thresholds and reason about confidence. That matters when a wrong answer is costly or when low-confidence predictions should be reviewed manually.

For example, in binary classification:

probs = model.predict(x, verbose=0).reshape(-1)
preds = (probs >= 0.8).astype(int)

That uses a stricter threshold than the usual 0.5, which may improve precision at the expense of recall.

Common Pitfalls

The most common mistake is assuming model.predict always returns probabilities. If the final layer has no activation and the loss was configured with from_logits=True, the outputs are logits.

Another pitfall is applying softmax twice. If the model already ends with softmax, the predictions are already probabilities.

Developers also sometimes take argmax immediately and throw away the probability vector, which removes useful confidence information.

Finally, probability values are not always perfectly calibrated. A model can be accurate while still being overconfident or underconfident, so treat the scores as estimates rather than guarantees.

Summary

• 'model.predict returns probabilities only if the model output layer already uses sigmoid or softmax.'
• If the model returns logits, convert them with tf.nn.sigmoid or tf.nn.softmax.
• Binary classifiers usually output one probability for the positive class.
• Multiclass classifiers usually output a probability vector across classes.
• Keep the probability output when you need thresholds, confidence checks, or downstream ranking logic.