def tp_xception_multi_period_score_fusion(dataset, config):
result = []
for i in range(iteration):
pre_list = []
y_test_list = []
for period in ['R1', 'R3', 'R4', 'R5', 'R6']:
model = BaseModel.Combined_Model(parallels=4, config=config)
model.load_weights('./{}_pdm_iteration-{}-{}.hdf5'.format(
dataset, i, period))
x_list = np.loadtxt('{}-{}_file_list.txt'.format(dataset, period))
img_x_list, shape_x, texture_x, vein_x, y_list = utils.data_loader_for_combined_model(
file_list=x_list,
dataset=dataset,
config=config,
isVenation=True)
img_x = np.array(img_x_list)
y = np.array(y_list)
id_map = np.loadtxt(dataset + '_id.txt')
for index, d in enumerate(y):
for label in id_map:
if d == label[0]:
y[index] = label[1]
y_one_hot = to_categorical(y)
test_index = np.load(
'{}_iteration_{}_img_{}_tp_xception_test_index.npy'.format(
dataset, i, period))
img_x_test = img_x[test_index]
shape_x_test = shape_x[test_index]
texture_x_test = texture_x[test_index]
vein_x_test = vein_x[test_index]
y_test = y_one_hot[test_index]
y_test_list.append(y_test)
x_test_list = create_input_list(shape_x_test,
texture_x_test,
vein_x_test,
img_x_test,
config,
isVenation=True)
pre = model.predict(x_test_list)
pre_list.append(pre)
pre_final_arr = np.array(pre_list)
pre_final = np.sum(pre_final_arr, axis=2)
pre_final_label = [np.argmax(d) for d in pre_final]
for i in range(1, len(y_test_list) + 1):
if y_test_list[i - 1] != y_test_list[i]:
print("The test label of different period should be the same")
return -1
y_test_label = [np.argmax(d) for d in y_test_list[0]]
performance = get_performance(pre_final_label, y_test_label)
result.append(performance)
plot_result(result)
def tp_xception_model_training_and_test(img_x_list, shape_x, texture_x, vein_x,
isVenation, y_list, config):
img_x = np.array(img_x_list)
shape_x = np.array(shape_x)
texture_x = np.array(texture_x)
if isVenation:
vein_x = np.array(vein_x)
y = np.array(y_list)
dataset = config['dataset']
id_map = np.loadtxt(dataset + '_id.txt')
for index, d in enumerate(y):
for label in id_map:
if d == label[0]:
y[index] = int(label[1])
y_one_hot = to_categorical(y)
result = []
lr_adjust = ReduceLROnPlateau(monitor='val_loss',
factor=0.5,
patience=5,
min_lr=1e-6)
dataset = config['dataset']
if dataset == 'soybean':
period = config['period']
else:
period = dataset
for i in range(iteration):
index = np.arange(len(y))
# print(index)
X_train_index, X_test_index, y_train, y_test = train_test_split(
index,
y_one_hot,
test_size=0.3,
random_state=i,
shuffle=True,
stratify=y_one_hot)
print(len(X_train_index))
print(len(X_test_index))
np.save(
'{}_iteration_{}_img_{}_tp_xception_train_index.npy'.format(
dataset, i, period), X_train_index)
np.save(
'{}_iteration_{}_img_{}_tp_xception_test_index.npy'.format(
dataset, i, period), X_test_index)
shape_x_train = shape_x[X_train_index]
texture_x_train = texture_x[X_train_index]
img_x_train = img_x[X_train_index]
shape_x_train_list = [
shape_x_train[:, i, :, :] for i in range(config["shape_views"])
]
texture_x_train_list = [
texture_x_train[:, i, :, :] for i in range(config["texture_views"])
]
if isVenation:
vein_x_train = vein_x[X_train_index]
vein_x_train[:, 0, :, 1] = (vein_x_train[:, 0, :, 1] - np.mean(
vein_x_train[:, 0, :, 1])) / np.std(vein_x_train[:, 0, :, 1])
vein_x_train[:, 1, :, 1] = (vein_x_train[:, 1, :, 1] - np.mean(
vein_x_train[:, 1, :, 1])) / np.std(vein_x_train[:, 1, :, 1])
vein_x_train_list = [
vein_x_train[:, i, :, :] for i in range(config["vein_views"])
]
x_train_list = []
for index, d in enumerate(texture_x_train_list):
texture_x_train_list[index] = np.reshape(
d, [d.shape[0], d.shape[1], d.shape[2], 1])
if isVenation:
for index, d in enumerate(vein_x_train_list):
vein_x_train_list[index] = np.reshape(
d, [d.shape[0], d.shape[1], d.shape[2], 1])
x_train_list.extend(shape_x_train_list)
x_train_list.extend(texture_x_train_list)
if isVenation:
x_train_list.extend(vein_x_train_list)
x_train_list.append(img_x_train)
y_train_label = [np.argmax(d) for d in y_train]
my_class_weights = list(
class_weight.compute_class_weight('balanced',
np.unique(y_train_label),
y_train_label))
class_weight_dict = dict(
zip([x for x in np.unique(y_train_label)], my_class_weights))
save_best_weight = ModelCheckpoint(
'./{}_pdm_iteration-{}-{}.hdf5'.format(dataset, i, period),
monitor='val_loss',
verbose=1,
save_best_only=True,
mode='auto',
save_weights_only=True)
model = BaseModel.Combined_Model(parallels=1, config=config)
lr_reduce = LearningRateScheduler(lr_reducer)
model.fit(x_train_list,
y_train,
batch_size=16,
epochs=100,
validation_split=0.1,
class_weight=class_weight_dict,
callbacks=[save_best_weight, lr_reduce, lr_adjust])
shape_x_test = shape_x[X_test_index]
texture_x_test = texture_x[X_test_index]
img_x_test = img_x[X_test_index]
shape_x_test_list = [
shape_x_test[:, i, :, :] for i in range(config["shape_views"])
]
texture_x_test_list = [
texture_x_test[:, i, :, :] for i in range(config["texture_views"])
]
if isVenation:
vein_x_test = vein_x[X_test_index]
vein_x_test[:, 0, :, 1] = (vein_x_test[:, 0, :, 1] - np.mean(
vein_x_train[:, 0, :, 1])) / np.std(vein_x_train[:, 0, :, 1])
vein_x_test[:, 1, :, 1] = (vein_x_test[:, 1, :, 1] - np.mean(
vein_x_train[:, 1, :, 1])) / np.std(vein_x_train[:, 1, :, 1])
vein_x_test_list = [
vein_x_test[:, i, :, :] for i in range(config["vein_views"])
]
x_test_list = []
for index, d in enumerate(texture_x_test_list):
texture_x_test_list[index] = np.reshape(
d, [d.shape[0], d.shape[1], d.shape[2], 1])
x_test_list.extend(shape_x_test_list)
x_test_list.extend(texture_x_test_list)
if isVenation:
for index, d in enumerate(vein_x_test_list):
vein_x_test_list[index] = np.reshape(
d, [d.shape[0], d.shape[1], d.shape[2], 1])
x_test_list.extend(vein_x_test_list)
x_test_list.append(img_x_test)
K.clear_session()
model2 = BaseModel.Combined_Model(parallels=1, config=config)
model2.load_weights('./{}_pdm_iteration-{}-{}.hdf5'.format(
dataset, i, period))
score = model2.evaluate(x_test_list, y_test, batch_size=128)
print(score)
y_test_label = np.array([np.argmax(d) for d in y_test])
y_pre = model2.predict(x_test_list)
y_pre_label = np.array([np.argmax(d) for d in y_pre])
performance = get_performance(y_pre_label, y_test_label)
performance['test_acc'] = score[1]
performance['test_loss'] = score[0]
result.append(performance)
K.clear_session()
print("precision_score: {}".format(performance['precision']))
print("recall_score: {}".format(performance['recall']))
print("f1_score: {}".format(performance['f1_score']))
json_str = json.dumps(performance, indent=4)
with open(
'./{}_pd-rgb-combined-iteration-{}-{}.json'.format(
dataset, i, period), 'w') as json_file:
json_file.write(json_str)
plot_result(result)