I’m trying to apply transfer learning using MobileNet on a custom dataset. My code works fine until I add tf.keras.metrics.Precision(name=”precision”) to the model’s metrics. After adding this metric, I encounter the following error during model.fit:

ValueError: Shapes (None, 4) and (None, 1) are incompatible

The relevant code block:

import tensorflow as tf
from tensorflow.keras.utils import image_dataset_from_directory as get_dataset

img_size = (224, 224)
preprocessing_layer = tf.keras.applications.mobilenet.preprocess_input
img_shape = img_size + (3,)

base_model = tf.keras.applications.MobileNet(
    input_shape=img_shape,
    include_top=False,
    weights='imagenet',
)

batch_size = 32
train_set = get_dataset(
    'some path',
    shuffle=True,
    batch_size=batch_size,
    image_size=img_size,
)

class_names = train_set.class_names
num_classes = len(class_names)

base_model.trainable = False
global_average_layer = tf.keras.layers.GlobalAveragePooling2D()

inputs = tf.keras.Input(shape=img_shape)
k = preprocessing_layer(inputs)
k = base_model(k, training=False)
k = global_average_layer(k)
k = tf.keras.layers.Dropout(0.2)(k)

prediction_layer = tf.keras.layers.Dense(
    num_classes,
    activation="softmax"
)
outputs = prediction_layer(k)

model = tf.keras.Model(inputs, outputs)

model.compile(
    optimizer=tf.keras.optimizers.Adam(learning_rate=0.0001),
    loss=tf.keras.losses.SparseCategoricalCrossentropy(),
    metrics=[
        tf.keras.metrics.SparseCategoricalAccuracy(name='accuracy'),
        tf.keras.metrics.Precision(name="precision"),
    ]
)

AUTOTUNE = tf.data.AUTOTUNE

train_set = train_set.prefetch(buffer_size=AUTOTUNE)

epochs = 1
history = model.fit(
    train_set,
    epochs=epochs,
)

The dataset is organized into subdirectories where each subdirectory name is the class label. There are four class labels.

The dataset loading and preprocessing seem to work correctly since the model trains without the tf.keras.metrics.Precision metric.

The issue arises specifically when adding the precision metric.

epochs is intentionally set to 1 for experimental purposes.