Project — Fashion Item Classifier

• Load Fashion-MNIST and map class indices to names.
• Build a CNN with conv + pool + dropout.
• Evaluate with a confusion matrix; find the hard pairs.
• Compare CNN vs dense net on a harder problem.

from tensorflow.keras.datasets import fashion_mnist
(Xtr, ytr), (Xte, yte) = fashion_mnist.load_data()

CLASSES = ["T-shirt", "Trouser", "Pullover", "Dress", "Coat",
           "Sandal", "Shirt", "Sneaker", "Bag", "Ankle boot"]
print(Xtr.shape, "labels 0-9 →", CLASSES)

Preprocess (add a channel dim for Conv2D)

Xtr = (Xtr.astype("float32") / 255.0)[..., None]   # (60000, 28, 28, 1)
Xte = (Xte.astype("float32") / 255.0)[..., None]

A solid CNN

from tensorflow import keras
from tensorflow.keras import layers

model = keras.Sequential([
    layers.Input(shape=(28, 28, 1)),
    layers.Conv2D(32, 3, activation="relu", padding="same"),
    layers.MaxPooling2D(),
    layers.Conv2D(64, 3, activation="relu", padding="same"),
    layers.MaxPooling2D(),
    layers.Flatten(),
    layers.Dense(128, activation="relu"),
    layers.Dropout(0.4),
    layers.Dense(10, activation="softmax"),
])
model.compile("adam", "sparse_categorical_crossentropy", metrics=["accuracy"])

What the conv layers learn

Conv2D(32, 3)   32 filters, each 3×3, slide across the image
                early layers learn edges; later layers learn textures/parts
MaxPooling2D    shrink the map, keep the strongest signals (translation tolerance)

# fashion.py — CNN clothing classifier
import numpy as np, matplotlib.pyplot as plt
from tensorflow import keras
from tensorflow.keras import layers
from tensorflow.keras.datasets import fashion_mnist
from tensorflow.keras.callbacks import EarlyStopping
from sklearn.metrics import classification_report, confusion_matrix, ConfusionMatrixDisplay

CLASSES = ["T-shirt", "Trouser", "Pullover", "Dress", "Coat",
           "Sandal", "Shirt", "Sneaker", "Bag", "Ankle boot"]

(Xtr, ytr), (Xte, yte) = fashion_mnist.load_data()
Xtr = (Xtr.astype("float32") / 255.0)[..., None]
Xte = (Xte.astype("float32") / 255.0)[..., None]

model = keras.Sequential([
    layers.Input(shape=(28, 28, 1)),
    layers.Conv2D(32, 3, activation="relu", padding="same"),
    layers.MaxPooling2D(),
    layers.Conv2D(64, 3, activation="relu", padding="same"),
    layers.MaxPooling2D(),
    layers.Flatten(),
    layers.Dense(128, activation="relu"),
    layers.Dropout(0.4),
    layers.Dense(10, activation="softmax"),
])
model.compile("adam", "sparse_categorical_crossentropy", metrics=["accuracy"])

es = EarlyStopping(monitor="val_loss", patience=3, restore_best_weights=True)
model.fit(Xtr, ytr, validation_split=0.1, epochs=20,
          batch_size=128, callbacks=[es], verbose=2)

acc = model.evaluate(Xte, yte, verbose=0)[1]
print(f"\ntest accuracy: {acc:.2%}")

preds = model.predict(Xte).argmax(axis=1)
print(classification_report(yte, preds, target_names=CLASSES))

ConfusionMatrixDisplay(confusion_matrix(yte, preds),
                       display_labels=CLASSES).plot(xticks_rotation=45)
plt.tight_layout(); plt.savefig("fashion_cm.png", dpi=130)

Sample output

test accuracy: 91.3%

              precision  recall  f1
T-shirt          0.84    0.86  0.85
...
Shirt            0.74    0.66  0.70   ← hardest class
Trouser          0.99    0.98  0.99   ← easiest

Read the diff

The per-class report is gold: "Trouser" is nearly perfect (distinctive shape), "Shirt" is the worst (it overlaps with T-shirt, Pullover, and Coat). The confusion matrix shows exactly which items bleed into each other. This is far more useful than the single 91% — it tells you where the model struggles.

Fashion-MNIST is the same format as digits but harder, so accuracy lands ~90% — accuracy is always relative to difficulty. A CNN (conv + pool) beats a dense net on images by respecting 2-D structure. The per-class report + confusion matrix reveal which classes overlap. Next: skip training entirely and use pre-trained models.