def fit_multiple(self, data_df, input_shape, batch_size=32, epochs=1):
# 分割数据集
train_indexs, val_indexs = get_kfold_train_val_idx(data_df)
# steps
steps = len(train_indexs) / batch_size
val_steps = len(val_indexs) / batch_size
# generator
train_gen = image_generator()
val_gen = image_generator()
history = self.model.fit_generator(
generator=self.__generator_multiple_batch_data(
train_gen, data_df, train_indexs, input_shape, batch_size),
steps_per_epoch=steps,
epochs=epochs,
validation_data=self.__generator_multiple_batch_data(
val_gen, data_df, val_indexs, input_shape, batch_size),
validation_steps=val_steps,
callbacks=self.__register_callbacks())
self.__save_history(history)
return history
def train(self, x_train, x_val, batch_size=128, lr=LEARN_RATE, epochs=1):
self.model.compile(optimizer=Adam(lr=lr), loss='mse', metrics=['mse'])
# @TODO: complete image_generator from path (img paths in csv file)
# Image generators from csv file and img dir
train_generator = image_generator(x_train,
batch_size,
self.img_shape,
outputShape=[3],
is_training=True)
val_generator = image_generator(x_val,
batch_size,
self.img_shape,
outputShape=[3],
is_training=False)
# Stop when error is below threshold 0.01
stop_callback = EarlyStopping(monitor='val_loss',
patience=20,
min_delta=0.01)
# For backup model
check_callback = ModelCheckpoint('psyncModel.ckpt',
monitor='val_loss',
save_best_only=True)
# For TensorBoard Visualization
vis_callback = TensorBoard(log_dir='./logs/%d' % int(time.time()),
histogram_freq=0,
write_graph=True,
write_images=True)
# Start training
self.model.fit_generator(
train_generator,
callbacks=[stop_callback, check_callback, vis_callback],
nb_epoch=1,
steps_per_epoch=len(x_train) % batch_size,
max_q_size=24,
validation_data=val_generator,
nb_val_samples=len(x_val),
nb_worker=4,
pickle_safe=True)
self.model.load_weights('psyncModel.ckpt')
self.model.save("Test")
def fit_disk(self, input_shape, batch_size=32, epochs=1):
# generator
train_gen = image_generator()
val_gen = image_generator()
train_generator = train_gen.flow_from_directory(
directory='data/train',
target_size=(input_shape[0], input_shape[1]),
batch_size=batch_size)
val_generator = val_gen.flow_from_directory(
directory='data/validation',
target_size=(input_shape[0], input_shape[1]),
batch_size=batch_size)
history = self.model.fit_generator(
generator=train_generator,
validation_data=val_generator,
epochs=epochs,
callbacks=self.__register_callbacks())
self.__save_history(history)
return history
from tensorflow.keras.optimizers import Adam
from data_generator import image_generator
from config import *
import os
import tensorflow as tf
from keras.backend import set_session
from PIL import Image
os.environ["CUDA_VISIBLE_DEVICES"] = "0"
train_step_per_epoch = len(
os.listdir(image_source_dir + 'train/')) / batch_size + 1
test_step_per_epoch = len(
os.listdir(image_source_dir + 'test/')) / batch_size + 1
train_image_generator = image_generator(image_source_dir + 'train/',
image_source_dir + 'train/',
batch_size=batch_size,
shuffle=True)
test_image_generator = image_generator(image_source_dir + 'test/',
image_source_dir + 'test/',
batch_size=batch_size,
shuffle=False)
opt1 = Adam(lr=lr)
opt2 = Adam(lr=lr)
discriminator = get_discriminator()
print(discriminator.summary())
generator = get_generator_unet()
generator.compile(optimizer=opt2,
loss='mae',
metrics=['mean_absolute_percentage_error'])
print(generator.summary())