def load_teacher_model(model_path, training_data):
model = build_cnn(training_data)
model.load_weights(model_path)
# remove softmax
model.layers.pop()
model = Model(model.input, model.layers[-1].output)
# now model outputs logits
return model
args = parser.parse_args()
bin_dir = os.path.dirname(os.path.realpath(__file__)) + '/bin'
if not os.path.exists(bin_dir):
os.makedirs(bin_dir)
training_data = load_data(args.file,
width=args.width,
height=args.height,
max_=args.max,
verbose=True) # args.verbose)
if args.model == 'cnn':
label = '10cnn_%d_%d_%d' % (32, 128, args.epochs)
model = build_cnn(training_data,
width=args.width,
height=args.height,
verbose=True) # args.verbose)
train(model, label, training_data, epochs=args.epochs)
elif args.model == 'mlp':
label = '10mlp_%d_%d' % (32, args.epochs)
model = build_mlp(training_data,
width=args.width,
height=args.height,
verbose=True) # args.verbose)
train(model, label, training_data, epochs=args.epochs)
elif args.model == 'student':
model = build_mlp(training_data,
width=args.width,
height=args.height,
verbose=True) # args.verbose)
temp = 2.0
# Split the data
(trainAttrX, testAttrX, trainImagesX,
testImagesX) = train_test_split(df, images, test_size=0.25, random_state=42)
# Normalization
max_price = trainAttrX["price"].max()
trainY = trainAttrX["price"] / max_price
testY = testAttrX["price"] / max_price
(trainAttrX, testAttrX) = datasets.process_attributes(df, trainAttrX,
testAttrX)
# Create network
mlp = models.build_mlp(trainAttrX.shape[1], False)
cnn = models.build_cnn(64, 64, 3, regress=False)
concatenated = concatenate([mlp.output, cnn.output])
x = Dense(4, activation="relu")(concatenated)
x = Dense(1, activation="linear")(x)
model = Model(inputs=[mlp.input, cnn.input], outputs=x)
optimizer = Adam(lr=1e-3, decay=1e-3 / 200)
model.compile(loss="mean_absolute_percentage_error", optimizer=optimizer)
model.fit([trainAttrX, trainImagesX],
trainY,
validation_data=([testAttrX, testImagesX], testY),
epochs=200,
def exp( args ):
'''
alg -- elu, qelu, relu, qrelu
lrate -- learning rate
dropout -- whether to dropout
batch_size -- number of samples in a mini batch
epochs -- how many epochs to train
'''
counter = Counter()
callbacks = [ counter ]
if args.arch == 'mlp':
model = models.build_mlp( args, counter )
elif args.arch == 'cnn':
model = models.build_cnn( args, counter )
elif args.arch == 'resnet':
model = models.build_resnet( args, counter )
elif args.arch == 'siamese':
model = models.build_siamese( args, counter )
elif args.arch == 'ae':
model = models.build_ae( args, counter )
else:
raise RuntimeError( 'unknown architecture' )
if args.arch == 'resnet':
if args.alg[0] == 'n':
optimizer = add_gradient_noise( keras.optimizers.Adam )( lr=lr_schedule(0), noise_eta=args.init_v, counter=counter )
else:
optimizer = keras.optimizers.Adam( lr=lr_schedule(0) )
lr_scheduler = LearningRateScheduler( lr_schedule )
lr_reducer = ReduceLROnPlateau( factor=np.sqrt(0.1),
cooldown=0,
patience=5,
min_lr=0.5e-6 )
callbacks += [ lr_reducer, lr_scheduler ]
elif args.alg[0] == 'n':
optimizer = add_gradient_noise( keras.optimizers.SGD )( lr=args.lrate, decay=1e-6, momentum=0.5, nesterov=False, noise_eta=args.init_v, counter=counter )
else:
optimizer = keras.optimizers.SGD( lr=args.lrate, decay=1e-6, momentum=0.5, nesterov=False )
if args.arch == 'siamese':
model.compile( loss=contrastive_loss, optimizer=optimizer, metrics=[siamese_accuracy] )
tr_pairs, tr_y = data.train_generator
te_pairs, te_y = data.test_generator
history = model.fit( [tr_pairs[:, 0], tr_pairs[:, 1]], tr_y,
batch_size=args.batch_size,
epochs=args.epochs,
callbacks=callbacks,
validation_data=( [te_pairs[:, 0], te_pairs[:, 1]], te_y),
verbose=args.verbose )
return history.history['loss'], history.history['val_siamese_accuracy']
elif args.arch == 'ae':
model.compile( loss=keras.losses.mse,
optimizer=optimizer )
history = model.fit_generator(
data.train_generator,
steps_per_epoch=(data.x_train.shape[0]//args.batch_size+1),
epochs=args.epochs,
callbacks=callbacks,
validation_data=data.test_generator,
validation_steps=data.x_test.shape[0]//args.test_batch_size,
verbose=args.verbose )
return history.history['loss'], history.history['val_loss']
else:
model.compile( loss=keras.losses.categorical_crossentropy,
optimizer=optimizer,
metrics=['accuracy'] )
history = model.fit_generator(
data.train_generator,
steps_per_epoch=(data.x_train.shape[0]//args.batch_size+1),
epochs=args.epochs,
callbacks=callbacks,
validation_data=data.test_generator,
validation_steps=data.x_test.shape[0]//args.test_batch_size,
verbose=args.verbose )
return history.history['loss'], history.history['val_acc']
# MONITOR = 'loss'
TRAIN_DATA = './data/image.npy'
TRAIN_FOLDER = './model_chk/auto_en_' + ENCODER + utils.get_time_mark()
CHECK_POINT_NAME = TRAIN_FOLDER + '/model.{epoch:04d}-{' + MONITOR + ':.4f}.h5'
LOG_NAME = TRAIN_FOLDER + '/log.csv'
if not os.path.exists(TRAIN_FOLDER):
os.makedirs(TRAIN_FOLDER)
train = np.load(TRAIN_DATA) / 255.
train = utils.shuffle(train)
if ENCODER == 'cnn':
train = train.reshape(-1, 28, 28, 1)
model = models.build_cnn()
else:
model = models.build_dnn()
model.summary()
callback = [
TensorBoard(TRAIN_FOLDER),
CSVLogger(LOG_NAME, append=True),
ReduceLROnPlateau(MONITOR,
factor=0.1,
patience=int(PATIENCE / 2),
verbose=1),
EarlyStopping(MONITOR, patience=PATIENCE),
ModelCheckpoint(CHECK_POINT_NAME, monitor=MONITOR, period=1)
]
from models import build_cnn
from train import fit_model
from dataset import load_dataset
from evaluate import evaluate_model
from options import opts
if __name__ == '__main__':
# read dataset
train_data, val_data, test_data, train_label, val_label, test_label = load_dataset(opts)
# training
if opts['run'] == 'train' or opts['run'] == 'both':
# build model
model = build_cnn(opts)
# fit model
fit_model(model, train_data, train_label, val_data, val_label, opts)
# testing
if opts['run'] == 'test' or opts['run'] == 'both':
if opts['run'] == 'test':
model = build_cnn(opts)
# test
evaluate_model(model, test_data, test_label, opts)