batch_size = 64
epochs = 1
validation_split = 0.2
step_size_train = 150
NAME = "Smart_Fridge_BaseSet_imageAugmentation:{}_batchSize:{}_epochs:{}_valSplit:{}_stepSize:{}_trainedFromLayer:{}_time:{}".format(
imageAugmentation, batch_size, epochs, validation_split, step_size_train,
train_from_layer, int(time.time()))
tensorboard = TensorBoard(log_dir='logs/{}'.format(NAME))
train_datagen = ImageDataGenerator(rescale=1. / 255,
validation_split=validation_split)
if imageAugmentation == True:
train_datagen.height_shift_range = 0.2
train_datagen.width_shift_range = 0.2
train_datagen.zoom_range = zoom_range = [0.1, 0.3]
train_datagen.shear_range = 30
train_generator = train_datagen.flow_from_directory(train_data_dir,
target_size=(224, 224),
color_mode='rgb',
batch_size=batch_size,
class_mode='categorical',
subset='training',
shuffle=True)
validation_generator = train_datagen.flow_from_directory(
train_data_dir, # same directory as training data
target_size=(224, 224),
def run(x_train_fn,
x_test_fn,
y_train_fn,
y_test_fn,
img_size=[256, 256],
do_featurewise_norm=True,
rotation_range=0,
width_shift_range=.0,
height_shift_range=.0,
zoom_range=[1.0, 1.0],
horizontal_flip=False,
vertical_flip=False,
batch_size=32,
nb_epoch=100,
pos_cls_weight=1.0,
nb_init_filter=32,
init_filter_size=5,
init_conv_stride=2,
pool_size=2,
pool_stride=2,
weight_decay=.0001,
alpha=.0001,
l1_ratio=.0,
inp_dropout=.0,
hidden_dropout=.0,
init_lr=.01,
lr_patience=20,
es_patience=40,
resume_from=None,
best_model='./modelState/roi_clf.h5',
final_model="NOSAVE"):
'''Train a deep learning model for ROI classifications
'''
# =========== Load training data =============== #
X_train = np.load(x_train_fn)
X_test = np.load(x_test_fn)
X_train = resize_img_dat(X_train, img_size)
X_test = resize_img_dat(X_test, img_size)
y_train = np.load(y_train_fn)
y_test = np.load(y_test_fn)
# ============ Train & validation set =============== #
if do_featurewise_norm:
imgen = ImageDataGenerator(featurewise_center=True,
featurewise_std_normalization=True)
imgen.fit(X_train)
else:
imgen = ImageDataGenerator(samplewise_center=True,
samplewise_std_normalization=True)
imgen.rotation_range = rotation_range
imgen.width_shift_range = width_shift_range
imgen.height_shift_range = height_shift_range
imgen.zoom_range = zoom_range
imgen.horizontal_flip = horizontal_flip
imgen.vertical_flip = vertical_flip
train_generator = imgen.flow(X_train,
y_train,
batch_size=batch_size,
shuffle=True,
seed=12345)
X_test -= imgen.mean
X_test /= imgen.std
validation_set = (X_test, y_test)
# ================= Model training ============== #
nb_worker = int(os.getenv('NUM_CPU_CORES', 4))
if resume_from is not None:
model = load_model(resume_from,
custom_objects={
'sensitivity': DMMetrics.sensitivity,
'specificity': DMMetrics.specificity
})
else:
model = ResNetBuilder.build_resnet_50(
(1, img_size[0], img_size[1]), 1, nb_init_filter, init_filter_size,
init_conv_stride, pool_size, pool_stride, weight_decay, alpha,
l1_ratio, inp_dropout, hidden_dropout)
sgd = SGD(lr=init_lr, momentum=0.9, decay=0.0, nesterov=True)
model.compile(optimizer=sgd,
loss='binary_crossentropy',
metrics=[DMMetrics.sensitivity, DMMetrics.specificity])
reduce_lr = ReduceLROnPlateau(monitor='val_loss',
factor=0.1,
patience=lr_patience,
verbose=1)
early_stopping = EarlyStopping(monitor='val_loss',
patience=es_patience,
verbose=1)
auc_checkpointer = DMAucModelCheckpoint(best_model,
validation_set,
batch_size=batch_size)
hist = model.fit_generator(
train_generator,
samples_per_epoch=len(X_train),
nb_epoch=nb_epoch,
class_weight={
0: 1.0,
1: pos_cls_weight
},
validation_data=validation_set,
callbacks=[reduce_lr, early_stopping, auc_checkpointer],
nb_worker=nb_worker,
pickle_safe=True, # turn on pickle_safe to avoid a strange error.
verbose=2)
# Training report.
min_loss_locs, = np.where(
hist.history['val_loss'] == min(hist.history['val_loss']))
best_val_loss = hist.history['val_loss'][min_loss_locs[0]]
best_val_sensitivity = hist.history['val_sensitivity'][min_loss_locs[0]]
best_val_specificity = hist.history['val_specificity'][min_loss_locs[0]]
print "\n==== Training summary ===="
print "Minimum val loss achieved at epoch:", min_loss_locs[0] + 1
print "Best val loss:", best_val_loss
print "Best val sensitivity:", best_val_sensitivity
print "Best val specificity:", best_val_specificity
if final_model != "NOSAVE":
model.save(final_model)
return hist
