def main():
config = configparser.RawConfigParser()
config.read('config.txt')
experiment_name = config.get('train', 'name')
if not os.path.exists('./logs/' + experiment_name):
os.system('mkdir ./logs/' + experiment_name)
epochs_num = int(config.get('train', 'epochs_num'))
batch_size = int(config.get('train', 'batch_size'))
# Load datasets.
datasets = config.get('train', 'datasets')
datasets_path = config.get(datasets, 'h5py_save_path')
height = int(config.get(datasets, 'height'))
width = int(config.get(datasets, 'width'))
pad_height = int(config.get(datasets, 'pad_height'))
pad_width = int(config.get(datasets, 'pad_width'))
x_train, y_train, masks = Generator(datasets_path, 'train', height, width,
pad_height, pad_width)()
visualize(group_images(x_train, 4),
'./logs/' + experiment_name + '/train_images.png').show()
visualize(group_images(y_train, 4),
'./logs/' + experiment_name + '/train_labels.png').show()
visualize(group_images(masks, 4),
'./logs/' + experiment_name + '/train_masks.png').show()
y_train = to_categorical(y_train)
# Build model and save.
unet = Unet((pad_height, pad_width, 1), 5)
unet.summary()
unet_json = unet.to_json()
open('./logs/' + experiment_name + '/architecture.json',
'w').write(unet_json)
plot_model(unet, to_file='./logs/' + experiment_name + '/model.png')
# Training.
checkpointer = ModelCheckpoint(filepath='./logs/' + experiment_name +
'/weights.h5',
verbose=1,
monitor='val_loss',
mode='auto',
save_best_only=True)
unet.fit(
x_train,
y_train,
epochs=epochs_num,
batch_size=batch_size,
verbose=1,
shuffle=True,
validation_split=0.1,
#class_weight=(0.5, 1.3),
callbacks=[checkpointer])
print(np.max(x_train), np.min(x_train), x_train.shape, x_train.dtype)
print(np.max(y_train), np.min(y_train), y_train.shape, y_train.dtype)
# Build model and save.
unet = Unet((sub_height, sub_width, 1))
unet.summary()
unet_json = unet.to_json()
open('./logs/' + experiment_name + '_architecture.json', 'w').write(unet_json)
plot_model(unet, to_file='./logs/' + experiment_name + '_model.png')
# Training.
checkpointer = ModelCheckpoint(filepath='./logs/' + experiment_name +
'_best_weights.h5',
verbose=1,
monitor='val_loss',
mode='auto',
save_best_only=True)
unet.fit(x_train,
y_train,
epochs=epochs_num,
batch_size=batch_size,
verbose=1,
shuffle=True,
validation_split=0.1,
callbacks=[checkpointer])
unet.save_weights('./logs/' + experiment_name + '_last_weights.h5',
overwrite=True)
sys.path.insert(1, './src')
from unet import Unet
from kerasaug import DataGenerator
if not os.path.exists(str("./" + dest)): os.mkdir(str("./" + dest))
os.chdir(str("./" + dest))
strategy = tf.distribute.experimental.MultiWorkerMirroredStrategy()
tf_config_json = os.environ.get("TF_CONFIG", "{}")
tf_config = json.loads(tf_config_json)
with strategy.scope():
model = Unet(num_class=1)
trainset = DataGenerator("../dataset/SNEMI3D_mito/train", batch_size=batch_size)
validset = DataGenerator("../dataset/SNEMI3D_mito/valid", batch_size=batch_size)
history = model.fit(trainset, steps_per_epoch=steps_per_epoch, epochs=epochs, validation_data=validset, validation_steps=validation_steps, verbose=1)
model.save('./seg_mito_keras.h5')
plt.plot(history.history['accuracy'], label='training accuracy')
plt.plot(history.history['val_accuracy'], label='validation accuracy')
plt.title('training and validation accuracy')
plt.xlabel("epoch")
plt.ylabel("accuracy")
plt.legend()
plt.savefig('./seg_mito_keras_acc.png')
plt.plot(history.history['loss'], label='training loss')
plt.plot(history.history['val_loss'], label='validation loss')
plt.title('Training and Validation loss')
