def autoencoder_train(folder, batch_size, epoch_size, model_name):
"""
Autoencoding, inherently UNET, is a data compression algorithm where the compression and decompression functions are:
- data specific, ie, only compress data similar to what they have been trained on
- lossy, ie, decompressed output will be degraded
- learned automatically from examples.
Two practical applications of autoencoders are data removal and dimensionality reduction
There is an implementation from scikit-learn:
http://scikit-learn.org/stable/modules/generated/sklearn.manifold.TSNE.html
:param folder: image folder for training
:param batch_size: training batch size
:param epoch_size: training epoch size
:param model_name: IR2, InceptionResNetV2; NL, NASNetLarge; NM, NASNetLarge
:return: None
"""
image_wh = system_config['image_wh']
image_size = (image_wh, image_wh)
image_shape = (image_wh, image_wh, 1)
train_list, valid_list = create_tv_list(folder)
print(f'Train size: {len(train_list)}, valid size: {len(valid_list)}')
train_df = pd.DataFrame(train_list, columns=['fname', 'class'])
valid_df = pd.DataFrame(valid_list, columns=['fname', 'class'])
model = None
if 'NM' in model_name:
model_name = 'NM'
model = NASNetMobile(include_top=True,
weights=None,
input_tensor=None,
input_shape=image_shape,
pooling='max',
classes=6)
elif 'NL' in model_name:
model_name = 'NL'
model = NASNetLarge(include_top=True,
weights=None,
input_tensor=None,
input_shape=image_shape,
pooling='max',
classes=6)
elif 'XC' in model_name:
model_name = 'XC'
model = Xception(include_top=True,
weights=None,
input_tensor=None,
input_shape=image_shape,
pooling='max',
classes=6)
elif 'D21' in model_name:
model_name = 'D21'
model = DenseNet201(include_top=True,
weights=None,
input_tensor=None,
input_shape=image_shape,
pooling='max',
classes=6)
elif 'IV3' in model_name:
model_name = 'IV3'
model = InceptionV3(include_top=True,
weights=None,
input_tensor=None,
input_shape=image_shape,
pooling='max',
classes=6)
elif 'SC' in model_name:
model_name = 'SC'
model = simple_cnn(input_shape=image_shape, classes=6)
else:
model_name = 'IR2'
model = InceptionResNetV2(include_top=True,
weights=None,
input_tensor=None,
input_shape=image_shape,
pooling='max',
classes=6)
model.compile(loss='categorical_crossentropy',
optimizer=Adam(lr=lr_schedule(0)),
metrics=['accuracy'])
model.summary()
# Image generator does data augmentation:
datagen = data_generator()
train_gen = datagen.flow_from_dataframe(dataframe=train_df,
directory=folder,
x_col="fname",
y_col="class",
class_mode="categorical",
target_size=image_size,
color_mode='grayscale',
batch_size=batch_size,
shuffle=False)
valid_gen = datagen.flow_from_dataframe(dataframe=valid_df,
directory=folder,
x_col="fname",
y_col="class",
class_mode="categorical",
target_size=image_size,
color_mode='grayscale',
batch_size=batch_size,
shuffle=False)
# Prepare model model saving directory.
save_dir = Path(os.path.dirname(
os.path.realpath(__file__))).joinpath('models')
if not save_dir.is_dir():
save_dir.mkdir(exist_ok=True)
filepath = f'{str(save_dir)}/{MODEL_NAMES[model_name]}'
print(f'{filepath}\n')
# Prepare callbacks for model saving and for learning rate adjustment.
checkpoint = ModelCheckpoint(filepath=filepath,
monitor='val_acc',
verbose=1,
save_best_only=True)
lr_scheduler = LearningRateScheduler(lr_schedule)
lr_reducer = ReduceLROnPlateau(factor=np.sqrt(0.1),
cooldown=0,
patience=5,
min_lr=0.5e-6)
callbacks = [checkpoint, lr_reducer, lr_scheduler]
# Fit the model on the batches generated by datagen.flow().
steps_per_epoch = int(len(train_list) / batch_size)
history = model.fit_generator(generator=train_gen,
steps_per_epoch=steps_per_epoch,
validation_data=valid_gen,
validation_steps=steps_per_epoch,
epochs=epoch_size,
use_multiprocessing=False,
verbose=1,
workers=4,
callbacks=callbacks)
# Score trained model.
scores = model.evaluate_generator(generator=valid_gen,
steps=steps_per_epoch,
verbose=1)
print('Test loss:', scores[0])
print('Test accuracy:', scores[1])
# Save score in configuration file
system_config[f'{model_name}_Accuracy'] = scores[1]
save_config()
return history
def aishufan_train(folder, batch_size, epoch_size, model_name):
"""
Train network with the parameters specified.
:param folder: image folder for training
:param batch_size: training batch size
:param epoch_size: training epoch size
:param model_name: IR2, InceptionResNetV2; NL, NASNetLarge; NM, NASNetLarge
:return: None
"""
image_wh = system_config['image_wh']
image_size = (image_wh, image_wh)
image_shape= (image_wh, image_wh, 3)
train_list, valid_list = create_tv_list(folder)
print(f'Train size: {len(train_list)}, valid size: {len(valid_list)}')
train_df = pd.DataFrame(train_list, columns=['fname', 'class'])
valid_df = pd.DataFrame(valid_list, columns=['fname', 'class'])
model = None
if 'NM' in model_name:
model_name = 'NM'
model = NASNetMobile(include_top=True,
weights=None,
input_tensor=None,
input_shape=image_shape,
pooling='max',
classes=2)
elif 'XC' in model_name:
model_name = 'XC'
model = Xception(include_top=True,
weights=None,
input_tensor=None,
input_shape=image_shape,
pooling='max',
classes=2)
elif 'D21' in model_name:
model_name = 'D21'
model = DenseNet201(include_top=True,
weights=None,
input_tensor=None,
input_shape=image_shape,
pooling='max',
classes=2)
elif 'IV3' in model_name:
model_name = 'IV3'
model = InceptionV3(include_top=True,
weights=None,
input_tensor=None,
input_shape=image_shape,
pooling='max',
classes=2)
else:
model_name = 'IR2'
model = InceptionResNetV2(include_top=True,
weights=None,
input_tensor=None,
input_shape=image_shape,
pooling='max',
classes=2)
model.compile(loss='categorical_crossentropy',
optimizer=Adam(lr=lr_schedule(0)),
metrics=['accuracy'])
model.summary()
# Image generator does data augmentation:
datagen = data_generator()
train_gen = datagen.flow_from_dataframe(
dataframe=train_df,
directory=folder,
x_col="fname",
y_col="class",
class_mode="categorical",
target_size=image_size,
color_mode='rgb',
batch_size=batch_size,
shuffle=False)
valid_gen = datagen.flow_from_dataframe(
dataframe=valid_df,
directory=folder,
x_col="fname",
y_col="class",
class_mode="categorical",
target_size=image_size,
color_mode='rgb',
batch_size=batch_size,
shuffle=False)
# Save class indices
system_config['class_indices'] = train_gen.class_indices
save_config()
# Prepare model model saving directory.
save_dir = Path(os.path.dirname(os.path.realpath(__file__))).joinpath('models')
if not save_dir.is_dir():
save_dir.mkdir(exist_ok=True)
filepath = f'{str(save_dir)}/{MODEL_NAMES[model_name]}'
print(f'{filepath}\n')
# Prepare callbacks for model saving and for learning rate adjustment.
checkpoint = ModelCheckpoint(filepath=filepath,
monitor='val_acc',
verbose=1,
save_best_only=True)
lr_scheduler = LearningRateScheduler(lr_schedule)
lr_reducer = ReduceLROnPlateau(factor=np.sqrt(0.1),
cooldown=0,
patience=5,
min_lr=0.5e-6)
callbacks = [checkpoint, lr_reducer, lr_scheduler]
# Fit the model on the batches generated by datagen.flow().
steps_per_epoch = int(len(train_list)/batch_size)
history = model.fit_generator(
generator=train_gen,
steps_per_epoch=steps_per_epoch,
validation_data=valid_gen,
validation_steps=steps_per_epoch,
epochs=epoch_size,
use_multiprocessing=False,
verbose=1,
workers=4,
callbacks=callbacks)
# Score trained model.
scores = model.evaluate_generator(generator=valid_gen, steps=steps_per_epoch, verbose=1)
print('Test loss:', scores[0])
print('Test accuracy:', scores[1])
# Save score in configuration file
system_config[f'{model_name}_Accuracy'] = scores[1]
save_config()
return history
# This callback will log model stats to Tensorboard.
tb_callback = TensorBoard()
# This callback will checkpoint the best model at every epoch.
mc_callback = ModelCheckpoint(filepath='current_best.hdf5',
verbose=1,
save_best_only=True)
# This is the train DataSequence.
train_sequence = DataSequence(train_pd, "./images", batch_size=batch_size)
train_steps = len(train_pd) // batch_size
# This is the validation DataSequence.
validation_sequence = DataSequence(test_pd, "./images", batch_size=batch_size)
validation_steps = len(test_pd) // batch_size
# These are the callbacks.
callbacks = [lr_callback, tb_callback, mc_callback]
# This line will train the model.
model.fit_generator(train_sequence,
validation_data=validation_sequence,
epochs=20,
use_multiprocessing=True,
workers=80,
steps_per_epoch=train_steps,
validation_steps=validation_steps,
callbacks=callbacks)
# Finally, we save the model.
model.save(MODEL_NAME)
callbacksList = [tensorboard]
if earlyStopFlag:
callbacksList.append(earlyStop)
if reduceLRFlag:
callbacksList.append(reduce_lr)
if modelCheckpointFlag:
callbacksList.append(modelCheckpoint)
print('passing class weights: {}'.format(
getClassWeights(trainGenerator.classes)))
history = model.fit_generator(
trainGenerator,
steps_per_epoch=trainSamplesNumber // batchSize * foldAugment,
epochs=epochs,
verbose=1,
callbacks=callbacksList,
validation_data=validationGenerator,
class_weight=getClassWeights(trainGenerator.classes),
shuffle=True,
validation_steps=validateSamplesNumber // batchSize)
score = model.evaluate_generator(testGenerator, testSamplesNumber)
print('Test loss:', score[0])
print('Test accuracy:', score[1])
# serialize model to JSON
model_json = model.to_json()
with open(modelFile, "w") as json_file:
json_file.write(model_json)
# serialize weights to HDF5
class Model_Nas:
def __init__(self, args, load=False):
self.ckpt = args.pre_train
self.model = "NasNet"
self.args = args
self.class_num = args.class_num
self.lr = args.lr
self.epoch = args.epoch
self.c = args.n_color
self.is_online = args.online
self.batch_size = args.batch_size
self.save_dir = args.save
self.sess = None
self.is_test = load
# self.mode = args.processing_mode
self.callbacks = []
self.init_callbacks()
def init_callbacks(self):
self.callbacks.append(
ModelCheckpoint(filepath=self.save_dir + self.model +
'_best_weights.h5',
verbose=1,
monitor='val_categorical_accuracy',
mode='auto',
save_best_only=True))
self.callbacks.append(
TensorBoard(
log_dir=self.args.save,
write_images=True,
write_graph=True,
))
self.callbacks.append(
EarlyStopping(
# patience=self.args.early_stopping
patience=1000))
# self.callbacks.append(
# ReduceLROnPlateau(
# monitor='val_loss',
# factor=0.5,
# patience=5,
# min_lr=1e-5
# )
# )
def custom_schedule(epochs):
if epochs <= 5:
lr = 1e-3
elif epochs <= 50:
lr = 5e-4
elif epochs <= 100:
lr = 2.5e-4
elif epochs <= 500:
lr = 1e-4
elif epochs <= 700:
lr = 5e-5
else:
lr = 1e-5
return lr
self.callbacks.append(LearningRateScheduler(custom_schedule))
def train(self, training_images, training_labels, validation_images,
validation_labels):
self.model = NASNetMobile(classes=2, include_top=True, weights=None)
self.model.trainable = True
self.model.compile(optimizer=Adam(lr=0.0001, beta_1=0.1),
loss='categorical_crossentropy',
metrics=['categorical_accuracy'])
train_datagen = ImageDataGenerator(rotation_range=40,
width_shift_range=0.2,
height_shift_range=0.2,
shear_range=0.2,
zoom_range=0.2,
horizontal_flip=True,
fill_mode='nearest')
val_datagen = ImageDataGenerator(rotation_range=40,
width_shift_range=0.2,
height_shift_range=0.2,
shear_range=0.2,
zoom_range=0.2,
horizontal_flip=True,
fill_mode='nearest')
steps = int(np.size(training_images, 0) // self.batch_size)
val_steps = int(np.size(validation_images, 0) // self.batch_size)
self.model.fit_generator(
generator=train_datagen.flow(x=training_images,
y=training_labels,
batch_size=self.batch_size),
epochs=self.args.epoch,
steps_per_epoch=steps,
validation_steps=val_steps,
verbose=1,
callbacks=self.callbacks,
validation_data=val_datagen.flow(x=validation_images,
y=validation_labels,
batch_size=self.batch_size))