return lr
INCV_lr_callback = LearningRateScheduler(INCV_lr_schedule)
## DONE
# Define optimizer and compile model
optimizer = optimizers.Adam(lr=INCV_lr_schedule(0),
beta_1=0.9,
beta_2=0.999,
epsilon=1e-08,
decay=0.0)
model = create_model(
input_shape=input_shape,
classes=n_classes,
name=INCV_name,
architecture='ResNet32'
) ## this is the INCV model, that is the model used in INCV for generating a clean dataset.
model.compile(optimizer=optimizer,
loss='categorical_crossentropy',
metrics=['accuracy'])
weights_initial = model.get_weights()
# Print model architecture
print('Architecture of INCV-model:')
model.summary()
## DONE
##################################################################################################################################
""" INCV iteration """
elif epoch > 80:
lr *= 1e-1
print('Learning rate: ', lr)
return lr
lr_callback = LearningRateScheduler(lr_schedule)
# Define optimizer and compile model
optimizer = optimizers.Adam(lr_schedule(0),
beta_1=0.9,
beta_2=0.999,
epsilon=1e-08,
decay=0.0)
model = create_model(input_shape=input_shape,
classes=n_classes,
name=network,
architecture=network)
model.summary()
parallel_model = multi_gpu_model(model, gpus=2)
parallel_model.compile(optimizer=optimizer,
loss='categorical_crossentropy',
metrics=['accuracy'])
##################################################################################################################################
""" Main training """
results = parallel_model.fit_generator(
datagen.flow(x_train[val1_idx, :],
y_train_noisy[val1_idx, :],
batch_size=batch_size),
epochs=epochs,
def INCV_lr_schedule(epoch):
# Learning Rate Schedule
lr = 1e-3
if epoch > 40:
lr *= 0.1
elif epoch > 30:
lr *= 0.25
elif epoch > 20:
lr *= 0.5
print('Learning rate: ', lr)
return lr
INCV_lr_callback = LearningRateScheduler(INCV_lr_schedule)
# Define optimizer and compile model
optimizer = optimizers.Adam(lr=INCV_lr_schedule(0), beta_1=0.9, beta_2=0.999, epsilon=1e-08, decay=0.0)
model = create_model(input_shape=input_shape, classes=n_classes, name=INCV_name, architecture='ResNet32')
model.compile(optimizer=optimizer, loss='categorical_crossentropy', metrics = ['accuracy'])
weights_initial = model.get_weights()
# Print model architecture
print('Architecture of INCV-model:')
model.summary()
##################################################################################################################################
""" INCV iteration """
train_idx = np.array([False for i in range(n_train)])
val_idx = np.array([True for i in range(n_train)])
INCV_save_best = True
for iter in range(1,INCV_iter+1):
print('INCV iteration %d including first half and second half. In total %d iterations.'%(iter,INCV_iter))
val_idx_int = np.array([i for i in range(n_train) if val_idx[i]]) # integer index
np.random.shuffle(val_idx_int)
def INCV_lr_schedule(epoch):
# Learning Rate Schedule
lr = 1e-3
if epoch > 40:
lr *= 0.1
elif epoch > 30:
lr *= 0.25
elif epoch > 20:
lr *= 0.5
print('Learning rate: ', lr)
return lr
INCV_lr_callback = LearningRateScheduler(INCV_lr_schedule)
# Define optimizer and compile model
optimizer = optimizers.Adam(lr=INCV_lr_schedule(0), beta_1=0.9, beta_2=0.999, epsilon=1e-08, decay=0.0)
model = create_model(input_shape=input_shape, classes=n_classes, name='INCV_ResNet32', architecture='ResNet32')
model.compile(optimizer=optimizer, loss='categorical_crossentropy', metrics = ['accuracy'])
weights_initial = model.get_weights()
# Print model architecture
print('Architecture of INCV-model:')
model.summary()
##################################################################################################################################
""" INCV iteration """
train_idx = np.array([False for i in range(n_train)])
val_idx = np.array([True for i in range(n_train)])
INCV_save_best = True
for iter in range(1,INCV_iter+1):
print('INCV iteration %d including first half and second half. In total %d iterations.'%(iter,INCV_iter))
val_idx_int = np.array([i for i in range(n_train) if val_idx[i]]) # integer index
np.random.shuffle(val_idx_int)