Binary image classification - Training Accuracy more than 1

Keras
,
#1

Hi,

I am training a binary classifier for X-Ray images (MURA dataset) using CNN.
I have created tensor dataset of (images,labels) .

On some runs the training accuracy is over 1. how could this be? I am using the standard accuracy metric from Keras, not a custom function.

Any help would be appreciated.

image
image1096×256 21.6 KB

#3

Can you share a snippet of your training code.

#4

Hi Atia,

This is it
def create_simple_cnn():
“”"
Creates a simple CNN model using the functional API for binary image classification.

Parameters:None

Returns:
- model: a tf.keras Model instance.
"""
input_layer = layers.Input(shape=(IMG_SIZE,IMG_SIZE,1))

x = augment_pipeline()(input_layer)

x = layers.Conv2D(8, (3,3), use_bias=False)(x)
x = layers.BatchNormalization(momentum=0.9)(x)
x = layers.Activation('relu')(x)

x = layers.Conv2D(16, (3,3), use_bias=False)(x)
x = layers.BatchNormalization(momentum=0.9)(x)
x = layers.Activation('relu')(x)

x = layers.Conv2D(32, (3,3), use_bias=False)(x)
x = layers.BatchNormalization(momentum=0.9)(x)
x = layers.Activation('relu')(x)

x = layers.MaxPooling2D((2, 2))(x)


x = layers.Conv2D(64, (3, 3), use_bias=False)(x)
x = layers.BatchNormalization(momentum=0.9)(x)
x = layers.Activation('relu')(x)

x = layers.Conv2D(128, (3, 3), use_bias=False)(x)
x = layers.BatchNormalization(momentum=0.9)(x)
x = layers.Activation('relu')(x)

x = layers.Conv2D(128, (3, 3), use_bias=False)(x)
x = layers.BatchNormalization(momentum=0.9)(x)
x = layers.Activation('relu')(x)
x = layers.MaxPooling2D((2, 2))(x)

x = layers.Conv2D(128, (3, 3), use_bias=False)(x)
x = layers.BatchNormalization(momentum=0.9)(x)
x = layers.Activation('relu')(x)

x = layers.Conv2D(128, (3, 3), use_bias=False)(x)
x = layers.BatchNormalization(momentum=0.9)(x)
x = layers.Activation('relu')(x)

x = layers.Conv2D(256, (3, 3), use_bias=False)(x)
x = layers.BatchNormalization(momentum=0.9)(x)
x = layers.Activation('relu')(x)
x = layers.MaxPooling2D((2, 2))(x)



x = layers.GlobalAveragePooling2D()(x)
x = layers.Dropout(0.2)(x)

output = layers.Dense(1,activation='sigmoid')(x)

model = models.Model(inputs=input_layer, outputs=output)

optimizer = tf.keras.optimizers.Adam(learning_rate=1e-4)
# Compiling the model
model.compile(optimizer=optimizer,
              loss='binary_crossentropy',
              metrics=['accuracy'])

return model

with strategy.scope():
model = create_simple_cnn()
history = model.fit(
train_dataset,
epochs=15,
callbacks=[cp_callback,reduce_lr],
validation_data=valid_dataset
)

#5

I run the same model with only one GPU, and the issue disappears.

1 Like