Confusion matrix on images in CNN keras

Question

I have trained my model(multiclass classification) of CNN using keras and now I want to evaluate the model on my test set of images.

What are the possible options for evaluating my model apart from the accuracy, precision and recall? I know how to get the precision and recall from a custom script. But I cannot find a way to get the confusion matrix for my 12 classes of images. Scikit-learn shows a way, but not for images. I am using model.fit_generator ()

Is there a way to create confusion matrix for all my classes or finding classification confidence on my classes? I am using Google Colab, though I can download the model and run it locally.

Any help would be appreciated.

Code:

train_data_path = 'dataset_cfps/train'
validation_data_path = 'dataset_cfps/validation'

#Parametres
img_width, img_height = 224, 224

vggface = VGGFace(model='resnet50', include_top=False, input_shape=(img_width, img_height, 3))

#vgg_model = VGGFace(include_top=False, input_shape=(224, 224, 3))

last_layer = vggface.get_layer('avg_pool').output
x = Flatten(name='flatten')(last_layer)
xx = Dense(256, activation = 'sigmoid')(x)
x1 = BatchNormalization()(xx)
x2 = Dropout(0.3)(x1)
y = Dense(256, activation = 'sigmoid')(x2)
yy = BatchNormalization()(y)
y1 = Dropout(0.6)(yy)
x3 = Dense(12, activation='sigmoid', name='classifier')(y1)

custom_vgg_model = Model(vggface.input, x3)


# Create the model
model = models.Sequential()

# Add the convolutional base model
model.add(custom_vgg_model)

model.summary()
#model = load_model('facenet_resnet_lr3_SGD_sameas1.h5')

def recall(y_true, y_pred):
     true_positives = K.sum(K.round(K.clip(y_true * y_pred, 0, 1)))
     possible_positives = K.sum(K.round(K.clip(y_true, 0, 1)))
     recall = true_positives / (possible_positives + K.epsilon())
     return recall

def precision(y_true, y_pred):
    true_positives = K.sum(K.round(K.clip(y_true * y_pred, 0, 1)))
    predicted_positives = K.sum(K.round(K.clip(y_pred, 0, 1)))
    precision = true_positives / (predicted_positives + K.epsilon())
    return precision

train_datagen = ImageDataGenerator(
      rescale=1./255,
      rotation_range=20,
      width_shift_range=0.2,
      height_shift_range=0.2,
      horizontal_flip=True,
      fill_mode='nearest')


validation_datagen = ImageDataGenerator(rescale=1./255)

# Change the batchsize according to your system RAM
train_batchsize = 32
val_batchsize = 32

train_generator = train_datagen.flow_from_directory(
        train_data_path,
        target_size=(img_width, img_height),
        batch_size=train_batchsize,
        class_mode='categorical')

validation_generator = validation_datagen.flow_from_directory(
        validation_data_path,
        target_size=(img_width, img_height),
        batch_size=val_batchsize,
        class_mode='categorical',
        shuffle=True)

# Compile the model
model.compile(loss='categorical_crossentropy',
              optimizer=optimizers.SGD(lr=1e-3),
              metrics=['acc', recall, precision])
# Train the model
history = model.fit_generator(
      train_generator,
      steps_per_epoch=train_generator.samples/train_generator.batch_size ,
      epochs=100,
      validation_data=validation_generator,
      validation_steps=validation_generator.samples/validation_generator.batch_size,
      verbose=1)

# Save the model
model.save('facenet_resnet_lr3_SGD_new_FC.h5')

Solution 1

Here's how to get the confusion matrix(or maybe statistics using scikit-learn) for all classes:

1.Predict classes

test_generator = ImageDataGenerator()
test_data_generator = test_generator.flow_from_directory(
    test_data_path, # Put your path here
     target_size=(img_width, img_height),
    batch_size=32,
    shuffle=False)
test_steps_per_epoch = numpy.math.ceil(test_data_generator.samples / test_data_generator.batch_size)

predictions = model.predict_generator(test_data_generator, steps=test_steps_per_epoch)
# Get most likely class
predicted_classes = numpy.argmax(predictions, axis=1)

2.Get ground-truth classes and class-labels

true_classes = test_data_generator.classes
class_labels = list(test_data_generator.class_indices.keys())   

3. Use scikit-learn to get statistics

report = metrics.classification_report(true_classes, predicted_classes, target_names=class_labels)
print(report)    

You can read more here

EDIT: If the above does not work, have a look at this video Create confusion matrix for predictions from Keras model. Probably look through the comments if you have an issue. Or Make predictions with a Keras CNN Image Classifier

Solution 2

Why would the scikit-learn function not do the job? You forward pass all your samples (images) in the train/test set, convert one-hot-encoding to label encoding (see link) and pass it into sklearn.metrics.confusion_matrix as y_pred. You proceed in a similar fashion with y_true (one-hot to label).

Sample code:

import sklearn.metrics as metrics

y_pred_ohe = KerasClassifier.predict(X)  # shape=(n_samples, 12)
y_pred_labels = np.argmax(y_pred_ohe, axis=1)  # only necessary if output has one-hot-encoding, shape=(n_samples)

confusion_matrix = metrics.confusion_matrix(y_true=y_true_labels, y_pred=y_pred_labels)  # shape=(12, 12)

Solution 3

Here cats and dogs are the class labels:

#Confusion Matrix and Classification Report
from sklearn.metrics import classification_report
from sklearn.metrics import confusion_matrix

Y_pred = model.predict_generator(validation_generator, nb_validation_samples // 
batch_size+1)
y_pred = np.argmax(Y_pred, axis=1)

print('Confusion Matrix')
print(confusion_matrix(validation_generator.classes, y_pred))

print('Classification Report')
target_names = ['Cats', 'Dogs']
print(classification_report(validation_generator.classes, y_pred, target_names=target_names))

Comments

[Abhishek Singh]

Could you please show some code so that I Can better understand and up vote/ accept your answer and close it?

[Abhishek Singh]

Hey @Jan K I have updated my code. Thank you for helping out. What can I add here?

[Nguai al]

In my case, how is that model.evaluate() returns accuracy rate of 75 percent but F1 score is only 25%? I would think that evaluate() and predict should be the same.