def objective(trial):
arch = get_arch_type(trial)
if arch == 'b0':
model = EfficientNetB0(weights=None, classes=n_classes, input_shape=in_shape)
elif arch == 'b1':
model = EfficientNetB1(weights=None, classes=n_classes, input_shape=in_shape)
elif arch == 'b2':
model = EfficientNetB2(weights=None, classes=n_classes, input_shape=in_shape)
elif arch == 'b3':
model = EfficientNetB3(weights=None, classes=n_classes, input_shape=in_shape)
elif arch == 'b4':
model = EfficientNetB4(weights=None, classes=n_classes, input_shape=in_shape)
elif arch == 'b5':
model = EfficientNetB5(weights=None, classes=n_classes, input_shape=in_shape)
elif arch == 'b6':
model = EfficientNetB6(weights=None, classes=n_classes, input_shape=in_shape)
elif arch == 'b7':
model = EfficientNetB7(weights=None, classes=n_classes, input_shape=in_shape)
loss = tfk.losses.CategoricalCrossentropy()
opt = get_optimizer(trial)
alpha = get_alpha(trial)
hist = training_mixup(model, train_ds, test_ds, loss, opt, n_epochs, batch_size, n_classes, alpha, output_best_weights=False)
test_acc = np.array(hist['test_acc'])
return 1 - test_acc.max()
def objective(trial):
model = DenseNet121(weights=None, classes=n_classes, input_shape=in_shape)
loss = tfk.losses.CategoricalCrossentropy()
opt = get_optimizer(trial)
alpha = get_alpha(trial)
hist = training_mixup(model, train_ds, test_ds, loss, opt, n_epochs, batch_size, n_classes, alpha, output_best_weights=False)
test_acc = np.array(hist['test_acc'])
return 1 - test_acc.max()
def objective(trial):
model = InceptionResNetV2(classes=n_classes, input_shape=in_shape)
loss = tfk.losses.CategoricalCrossentropy()
opt = get_optimizer(trial)
alpha = get_alpha(trial)
train_weights = model.layers[1].trainable_weights
hist = training_mixup(model, train_ds, test_ds, loss, opt, n_epochs, batch_size, n_classes, alpha, output_best_weights=False, train_weights=train_weights)
test_acc = np.array(hist['test_acc'])
return 1 - test_acc.max()
with open('test_ds.pkl', 'wb') as fp:
pickle.dump((x_test, y_test), fp, protocol=4)
del x_train, y_train, x_test, y_test
models_and_params = {
'vgg16': (models.wrapper.VGG16(weights=None, classes=3, input_shape=in_shape), 6.33e-5, 0.129, False),
'vgg16_pretrained': (models.wrapper_T.VGG16(classes=3, input_shape=in_shape), 5.66e-5, 0.121, True),
'vgg16_ft': (models.wrapper_T.VGG16(classes=3, input_shape=in_shape), 1.51e-5, 0.241, False),
}
# Loss
loss = tfk.losses.CategoricalCrossentropy()
# Validating
for model_name in models_and_params:
model, lr, alpha, pretrained = models_and_params[model_name]
opt = tfk.optimizers.Adam(lr)
weight_name = 'best_param_{}'.format(model_name)
if not pretrained:
hist = training_mixup(model, train_ds, test_ds, loss, opt, n_epochs, batch_size, n_classes, alpha=alpha,
output_best_weights=True, weight_name=weight_name)
else:
train_weights = model.layers[1].trainable_variables
hist = training_mixup(model, train_ds, test_ds, loss, opt, n_epochs, batch_size, n_classes, alpha=alpha,
output_best_weights=True, weight_name=weight_name, train_weights=train_weights)
plot_learning_curve('history_{}.png'.format(model_name), history=hist, epochs=n_epochs)
pd.DataFrame(hist).to_csv('history_{}.csv'.format(model_name))
in_shape = x_train.shape[1:]
del x_train, y_train, x_test, y_test
# Model
# model = SimpleCNN(in_shape, n_out=3)
# model = VGG16(weights=None, classes=3, input_shape=in_shape)
# model = ResNet50(weights=None, classes=3, input_shape=in_shape)
# model = InceptionV3(weights=None, classes=3, input_shape=in_shape)
# model = InceptionV3(classes=3, input_shape=in_shape)
model = ResNet50(classes=n_classes, input_shape=in_shape)
# Loss
loss = tfk.losses.CategoricalCrossentropy()
# Optimizer
opt = tfk.optimizers.Adam(lr)
# Training
hist = training_mixup(model,
train_ds,
test_ds,
loss,
opt,
n_epochs,
batch_size,
n_classes,
alpha=0.2)
pd.DataFrame(hist).to_csv('history.csv')