Inception ResNetV2 Model for ODIR challenge

This release contains the training log, results, and also the model weight in h5 format.

Training Conditions:

• Data Augmentation.
• Transfer learning using ImageNet weights.
• All layers are trained.
• Optimizer = SGD lr=0.001, decay=1e-6, momentum=0.9, nesterov=False

The results of this model are as follows:

Inception ResNetV2

Training:

• loss: 0.3823
• accuracy: 0.8906
• precision: 0.5723
• recall: 0.4950
• AUC: 0.8347

Validation:

• loss : 0.3409378457069397
• accuracy : 0.890625
• precision : 0.57225436
• recall : 0.495
• auc : 0.8346987

Final Score:

• Kappa score: 0.46929492039423804
• F-1 score: 0.890625
• AUC value: 0.8381830357142857
• Final Score: 0.7327009853695078

Changes in the model:

# Metrics
defined_metrics = [
    tf.keras.metrics.BinaryAccuracy(name='accuracy'),
    tf.keras.metrics.Precision(name='precision'),
    tf.keras.metrics.Recall(name='recall'),
    tf.keras.metrics.AUC(name='auc'),
]

# Added a new dense layer
x = GlobalAveragePooling2D()(x)
x = Dense(1024, activation='relu')(x)
predictions = Dense(num_classes, activation='sigmoid')(x)

sgd = SGD(lr=0.001, decay=1e-6, momentum=0.9, nesterov=True)
x.compile(optimizer=sgd, loss='binary_crossentropy', metrics=defined_metrics)

VGG-19 Model for ODIR challenge

This release contains the training log, results, and also the model weight in h5 format.

Training Conditions:

• Data Augmentation.
• Transfer learning using ImageNet weights.
• All layers are trained.
• Optimizer = SGD lr=0.001, decay=1e-6, momentum=0.9, nesterov=False

The results of this model are as follows:

VGG-19

Training:

• loss: 0.2320
• accuracy: 0.8988
• precision: 0.7241
• recall: 0.5074
• AUC: 0.9240

Validation:

• loss : 0.32848785161972044
• accuracy : 0.8746875
• precision : 0.49753696
• recall : 0.2525
• AUC : 0.78385717

Final Score:

• Kappa score: 0.2738795835219556
• F-1 score: 0.8746875
• AUC value: 0.7862857142857143
• Final Score: 0.6449509326025565

Changes in the model:

# Transfer learning, load previous weights
x.load_weights(r'C:\temp\vgg19_weights_tf_dim_ordering_tf_kernels.h5')

# Remove the last layer
 x.pop()

# Metrics
defined_metrics = [
    tf.keras.metrics.BinaryAccuracy(name='accuracy'),
    tf.keras.metrics.Precision(name='precision'),
    tf.keras.metrics.Recall(name='recall'),
    tf.keras.metrics.AUC(name='auc'),
]

# Added a new dense layer
x.add(layers.Dense(8, activation='sigmoid'))
sgd = SGD(lr=0.001, decay=1e-6, momentum=0.9, nesterov=False)
x.compile(optimizer=sgd, loss='binary_crossentropy', metrics=defined_metrics)

Xception Model for ODIR challenge

This release contains the training log, results, and also the model weight in h5 format.

Training Conditions:

• Data Augmentation.
• Transfer learning using ImageNet weights.
• All layers are trained.
• Optimizer = SGD lr=0.01, decay=1e-6, momentum=0.9, nesterov=True

The results of this model are as follows:

Xception:

Training:

• loss: 0.0201
• accuracy: 0.9934
• precision: 0.9838
• recall: 0.9713
• AUC: 0.9996

Validation:

• loss : 0.4635940623283386
• accuracy : 0.8875
• precision : 0.5531915
• recall : 0.52
• AUC : 0.8395691

Final Score:

• Kappa score: 0.47214076246334313
• F-1 score: 0.8875
• AUC value: 0.8611830357142858
• Final Score: 0.740274599392543

Changes in the model:

base_model = xception.Xception

base_model = base_model(weights='imagenet', include_top=False)
x = base_model.output
x = GlobalAveragePooling2D()(x)
x = Dense(1024, activation='relu')(x)
predictions = Dense(num_classes, activation='sigmoid')(x)
model = Model(inputs=base_model.input, outputs=predictions)

defined_metrics = [
    tf.keras.metrics.BinaryAccuracy(name='accuracy'),
    tf.keras.metrics.Precision(name='precision'),
    tf.keras.metrics.Recall(name='recall'),
    tf.keras.metrics.AUC(name='auc'),
]

sgd = SGD(lr=0.01, decay=1e-6, momentum=0.9, nesterov=True)
model.compile(loss='binary_crossentropy',
              optimizer=sgd,
              metrics=defined_metrics)

ResNet50 Model for ODIR challenge

This release contains the training log, results, and also the model weight in h5 format.

Training Conditions:

• Data Augmentation.
• Transfer learning using ImageNet weights.
• All layers are trained.
• Optimizer = SGD lr=0.0001, decay=1e-6, momentum=0.9, nesterov=True

The results of this model are as follows:

ResNet50:

Training:

• loss: 0.1591
• accuracy: 0.9344
• precision: 0.8432
• recall: 0.6823
• AUC: 0.9678

Validation:

• loss : 0.30101956367492677
• accuracy : 0.8834375
• precision : 0.543131
• recall : 0.425
• AUC : 0.8420562

Final Score:

• Kappa score: 0.41236707365104375
• F-1 score: 0.8834375
• AUC value: 0.843525
• Final Score: 0.7131098578836812

Changes in the model:

base_model = resnet50.ResNet50

base_model = base_model(weights='imagenet', include_top=False)
x = base_model.output
x = GlobalAveragePooling2D()(x)
x = Dense(1024, activation='relu')(x)
predictions = Dense(num_classes, activation='sigmoid')(x)
model = Model(inputs=base_model.input, outputs=predictions)

defined_metrics = [
    tf.keras.metrics.BinaryAccuracy(name='accuracy'),
    tf.keras.metrics.Precision(name='precision'),
    tf.keras.metrics.Recall(name='recall'),
    tf.keras.metrics.AUC(name='auc'),
]

sgd = SGD(lr=0.0001, decay=1e-6, momentum=0.9, nesterov=True)
model.compile(loss='binary_crossentropy',
              optimizer=sgd,
              metrics=defined_metrics)

VGG-16 Model for ODIR challenge

This release contains the training log, results and also the model weight in h5 format.

Training Conditions:

• Data Augmentation.
• Transfer learning using ImageNet weights.
• Only classifier layer is trained.
• Optimizer = SGD lr=0.001, decay=1e-6, momentum=0.9, nesterov=True
• No Dropout

The results of this model are as follows:

VGG-16

Training:

• loss: 0.7054
• accuracy: 0.8929
• precision: 0.6998
• recall: 0.4799
• AUC: 0.9168

Validation:

• loss : 0.31377036929130553
• accuracy : 0.8871875
• precision : 0.57768923
• recall : 0.3625
• AUC : 0.8140241

Final Score:

• Kappa score: 0.38631534211644714
• F-1 score: 0.8871875
• AUC value: 0.8176205357142858
• Final Score: 0.6970411259435777

Changes in the model:

# Transfer learning, load previous weights
x.load_weights(r'C:\temp\vgg16_weights_tf_dim_ordering_tf_kernels.h5')

# Remove the last layer
 x.pop()

# Metrics
defined_metrics = [
    tf.keras.metrics.BinaryAccuracy(name='accuracy'),
    tf.keras.metrics.Precision(name='precision'),
    tf.keras.metrics.Recall(name='recall'),
    tf.keras.metrics.AUC(name='auc'),
]

# Added a new dense layer
x.add(layers.Dense(8, activation='sigmoid'))
sgd = SGD(lr=0.001, decay=1e-6, momentum=0.9, nesterov=True)
x.compile(optimizer=sgd, loss='binary_crossentropy', metrics=defined_metrics)

Included in this release:

• Figures and test run output for comparison.
• Model weights of the trained model for inference testing.

Inception V3 Model for ODIR challenge

This release contains the training log, results and also the model weight in h5 format.

Training Conditions:

• Data Augmentation.
• Transfer learning using ImageNet weights.
• All layers are trained.
• Optimizer = SGD lr=0.01, decay=1e-6, momentum=0.9, nesterov=True

The results of this model are as follows:

Inception v3:

Training:

• loss: 0.0485
• accuracy: 0.9812
• precision: 0.9460
• recall: 0.9252
• AUC: 0.9969

Validation:

• loss : 0.37697273850440977
• accuracy : 0.8984375
• precision : 0.6021798
• recall : 0.5525
• AUC : 0.85563624

Final Score:

• Kappa score: 0.5186967789707515
• F-1 score: 0.8984375
• AUC value: 0.8838098214285715
• Final Score: 0.7669813667997744

Changes in the model:

base_model = inception_v3.InceptionV3

base_model = base_model(weights='imagenet', include_top=False)
x = base_model.output
x = GlobalAveragePooling2D()(x)
x = Dense(1024, activation='relu')(x)
predictions = Dense(num_classes, activation='sigmoid')(x)
model = Model(inputs=base_model.input, outputs=predictions)

defined_metrics = [
    tf.keras.metrics.BinaryAccuracy(name='accuracy'),
    tf.keras.metrics.Precision(name='precision'),
    tf.keras.metrics.Recall(name='recall'),
    tf.keras.metrics.AUC(name='auc'),
]

sgd = SGD(lr=0.01, decay=1e-6, momentum=0.9, nesterov=True)
model.compile(loss='binary_crossentropy',
              optimizer=sgd,
              metrics=defined_metrics)