def __init__(self, img_paths, labels, batch_size, img_size, use, preprocess_input):
assert len(img_paths) == len(labels), "len(img_paths) must equal to len(lables)"
assert img_size[0] == img_size[1], "img_size[0] must equal to img_size[1]"
self.x_y = np.hstack((np.array(img_paths).reshape(len(img_paths), 1), np.array(labels)))
self.batch_size = batch_size
self.img_size = img_size
self.use = use
self.preprocess_input = preprocess_input
self.eraser = get_random_eraser( s_h=0.3,pixel_level=True)
def multi_erase(x):
eraser = get_random_eraser(p=1.0,
s_l=0.02,
s_h=0.4,
r_1=0.3,
r_2=1 / 0.3,
v_l=0,
v_h=255,
pixel_level=True)
x = eraser(x)
eraser = get_random_eraser(p=0.8,
s_l=0.02,
s_h=0.4,
r_1=0.3,
r_2=1 / 0.3,
v_l=0,
v_h=255,
pixel_level=True)
x = eraser(x)
return x
def __init__(self, prefix, labels, batch_size):
self.name = prefix + 'dataset'
# Load preprocessed data
self.X_train = np.load(prefix + 'X_train.npy')
self.y_train = keras.utils.to_categorical(
np.load(prefix + 'y_train.npy'))
self.X_valid = np.load(prefix + 'X_valid.npy')
self.y_valid = keras.utils.to_categorical(
np.load(prefix + 'y_valid.npy'))
self.X_test = np.load(prefix + 'X_test.npy')
self.y_test = keras.utils.to_categorical(np.load(prefix +
'y_test.npy'))
# Make label from/to class converter
self.labels = labels
self.label2int = {l: i for i, l in enumerate(labels)}
self.int2label = {i: l for i, l in enumerate(labels)}
self.num_classes = len(self.labels)
# Normalize
max_amplitude = np.max(
np.abs(np.vstack([self.X_train, self.X_valid, self.X_test])))
self.X_train = self.X_train / max_amplitude
self.X_valid = self.X_valid / max_amplitude
self.X_test = self.X_test / max_amplitude
# Add dimension [:, features, timesetep] -> [:, features, timestep, 1]
self.X_train = self.X_train[..., np.newaxis]
self.X_valid = self.X_valid[..., np.newaxis]
self.X_test = self.X_test[..., np.newaxis]
# Make data generators
datagen = ImageDataGenerator(
featurewise_center=True,
featurewise_std_normalization=True,
rotation_range=0,
width_shift_range=0.4,
height_shift_range=0,
horizontal_flip=True,
preprocessing_function=get_random_eraser(
v_l=np.min(self.X_train),
v_h=np.max(self.X_train)) # RANDOM ERASER
)
datagen.fit(np.r_[self.X_train, self.X_valid, self.X_test])
test_datagen = ImageDataGenerator(
featurewise_center=datagen.featurewise_center,
featurewise_std_normalization=datagen.featurewise_std_normalization
)
test_datagen.mean, test_datagen.std = datagen.mean, datagen.std
self.datagen = datagen
self.test_datagen = test_datagen
self.batch_size = batch_size
self.reset_generators()
def __init__(self, label_file, input_size=224):
'''
img_dir: 图片路径：img_dir + img_name.jpg构成图片的完整路径
'''
# 所有图片的绝对路径
with open(label_file, 'r') as f:
#label_file的格式， （label_file image_label)
self.imgs = list(map(lambda line: line.strip().split(' '), f))
# 相关预处理的初始化
# self.transforms=transform
self.img_aug = True
self.transform = get_train_transform(size=cfg.INPUT_SIZE)
self.eraser = get_random_eraser(s_h=0.1, pixel_level=True)
self.input_size = cfg.INPUT_SIZE
def get_datagen(dataset, batch_size, use_random_eraser=True, use_mixup=True):
datagen = ImageDataGenerator(
featurewise_center=True, # set input mean to 0 over the dataset
featurewise_std_normalization=True, # divide inputs by std of the dataset
rotation_range=0, # randomly rotate images in the range (degrees, 0 to 180)
width_shift_range=0.6, # randomly shift images horizontally (fraction of total width)
height_shift_range=0, # randomly shift images vertically (fraction of total height)
horizontal_flip=True, # randomly flip images
preprocessing_function=\
get_random_eraser(v_l=np.min(dataset.X_train), v_h=np.max(dataset.X_train)) \
if use_random_eraser else None
)
test_datagen = ImageDataGenerator(
featurewise_center=True, # set input mean to 0 over the dataset
featurewise_std_normalization=
True, # divide inputs by std of the dataset
)
datagen.fit(np.r_[dataset.X_train, dataset.X_valid, dataset.X_test])
test_datagen.mean, test_datagen.std = datagen.mean, datagen.std
train_flow = datagen.flow(dataset.X_train,
dataset.y_train,
batch_size=batch_size)
if use_mixup:
train_flow = MixupGenerator(dataset.X_train,
dataset.y_train,
alpha=1.0,
batch_size=batch_size,
datagen=datagen)()
valid_flow = test_datagen.flow(dataset.X_valid,
dataset.y_valid,
shuffle=False)
y_test_just_for_api = keras.utils.to_categorical(
np.ones(len(dataset.X_test)))
test_flow = test_datagen.flow(dataset.X_test,
y_test_just_for_api,
shuffle=False)
return train_flow, valid_flow, test_flow, datagen, test_datagen
def prepare_generator(df, x_col, y_col, width, height, batch_size, test_size,
classes):
x_train, x_val, y_train, y_val = train_test_split(df[x_col],
df[y_col],
test_size=test_size,
shuffle=True,
random_state=SEED,
stratify=df[y_col])
cutout = get_random_eraser(p=0.9, v_l=0, v_h=1, pixel_level=True)
train_gen = ImageDataGenerator(rotation_range=45,
width_shift_range=.15,
height_shift_range=.15,
horizontal_flip=True,
vertical_flip=True,
zoom_range=0.5,
preprocessing_function=cutout,
rescale=1. / 255)
train_generator = train_gen.flow_from_dataframe(
pd.concat([x_train, y_train], axis=1),
x_col=x_col,
y_col=y_col,
target_size=(width, height),
batch_size=batch_size,
class_mode='categorical',
subset='training',
classes=classes)
valid_gen = ImageDataGenerator(rescale=1. / 255)
valid_generator = valid_gen.flow_from_dataframe(pd.concat([x_val, y_val],
axis=1),
x_col=x_col,
y_col=y_col,
target_size=(width,
height),
batch_size=batch_size,
class_mode='categorical',
subset='training',
classes=classes)
return train_generator, valid_generator
def getdata(train_path, val_path, test_path):
# create a data generator
image_size = config.IMAGE_SIZE
datagen_batch_size = config.batch_size
datagen = ImageDataGenerator(horizontal_flip=True,
width_shift_range=0.2,
height_shift_range=0.2,
zoom_range=[0.5, 1.0],
shear_range=0.2,
preprocessing_function=get_random_eraser(
v_l=0, v_h=255))
# test data shouldn't be augmented
test_datagen = ImageDataGenerator()
train_it = datagen.flow_from_directory(train_path,
class_mode='categorical',
batch_size=datagen_batch_size,
target_size=(image_size,
image_size))
# load and iterate validation dataset
val_it = datagen.flow_from_directory(val_path,
class_mode='categorical',
batch_size=datagen_batch_size,
target_size=(image_size, image_size))
# load and iterate test dataset
test_it = test_datagen.flow_from_directory(test_path,
class_mode='categorical',
batch_size=datagen_batch_size,
target_size=(image_size,
image_size))
return train_it, val_it, test_it
p = {
# your parameter boundaries come here
#'image_size': [144, 196, 240],
#'model': ['inception_resnet_v2', 'nasnet'],
#'learning_rate': [0.1, 0.01],
#'decay_factor': [1, 2, 5, 10, 100],
#'momentum': [0.9, 0.95, 0.99],
'deep_layers': [2, 3, 4],
'freeze_layers': [339, 399]
}
eraser = get_random_eraser(p=0.8,
s_l=0.02,
s_h=0.4,
r_1=0.3,
r_2=1 / 0.3,
v_l=0,
v_h=255,
pixel_level=True)
def preprocess(x):
x = eraser(x)
x = preprocess_input(x)
return x
def input_model(x_train, y_train, x_val, y_val, params):
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
session = tf.Session(config=config)
batch_size_test = 32
epochs = 100
aeons = 1000
alpha = 1
train_data = np.copy(X) #np.copy(X[tch[:, j]])
val_data = np.copy(Xv) #np.copy(Xv[tchv[:, j]])
tst_data = np.copy(X2) #np.copy(X2[tchv[:, j]])
# create data generator for mixup and random erasing for every batch
datagen = keras.preprocessing.image.ImageDataGenerator(
#featurewise_center=True,
#featurewise_std_normalization=True,
width_shift_range=7,
height_shift_range=3,
preprocessing_function=get_random_eraser(v_l=np.min(train_data),
v_h=np.max(train_data)))
test_datagen = keras.preprocessing.image.ImageDataGenerator(
#featurewise_center=True,
#featurewise_std_normalization=True
)
datagen.fit(np.r_[train_data])
test_datagen.fit(np.r_[train_data])
training_generator = MixupGenerator(train_data,
tch,
batch_size=batch_size,
alpha=alpha,
datagen=datagen)
# compile model
input, model_output, z_mean, z_log_var = vae_def(feat_length=60, nceps=40)
vae = keras.Model(inputs=[input], outputs=[model_output])
def main():
args = get_args()
input_path = args.input
batch_size = args.batch_size
nb_epochs = args.nb_epochs
depth = args.depth
k = args.width
validation_split = args.validation_split
use_augmentation = args.aug
logging.debug("Loading data...")
image, gender, age, _, image_size, _ = load_data(input_path)
X_data = image
y_data_g = np_utils.to_categorical(gender, 2)
y_data_a = np_utils.to_categorical(age, 101)
model = WideResNet(image_size, depth=depth, k=k)()
sgd = SGD(lr=0.1, momentum=0.9, nesterov=True)
model.compile(optimizer=sgd, loss=["categorical_crossentropy", "categorical_crossentropy"],
metrics=['accuracy'])
logging.debug("Model summary...")
model.count_params()
model.summary()
logging.debug("Saving model...")
mk_dir("models")
with open(os.path.join("models", "WRN_{}_{}.json".format(depth, k)), "w") as f:
f.write(model.to_json())
mk_dir("checkpoints")
callbacks = [LearningRateScheduler(schedule=Schedule(nb_epochs)),
ModelCheckpoint("checkpoints/weights.{epoch:02d}-{val_loss:.2f}.hdf5",
monitor="val_loss",
verbose=1,
save_best_only=True,
mode="auto")
]
logging.debug("Running training...")
data_num = len(X_data)
indexes = np.arange(data_num)
np.random.shuffle(indexes)
X_data = X_data[indexes]
y_data_g = y_data_g[indexes]
y_data_a = y_data_a[indexes]
train_num = int(data_num * (1 - validation_split))
X_train = X_data[:train_num]
X_test = X_data[train_num:]
y_train_g = y_data_g[:train_num]
y_test_g = y_data_g[train_num:]
y_train_a = y_data_a[:train_num]
y_test_a = y_data_a[train_num:]
if use_augmentation:
datagen = ImageDataGenerator(
width_shift_range=0.1,
height_shift_range=0.1,
horizontal_flip=True,
preprocessing_function=get_random_eraser(v_l=0, v_h=255))
training_generator = MixupGenerator(X_train, [y_train_g, y_train_a], batch_size=batch_size, alpha=0.2,
datagen=datagen)()
hist = model.fit_generator(generator=training_generator,
steps_per_epoch=train_num // batch_size,
validation_data=(X_test, [y_test_g, y_test_a]),
epochs=nb_epochs, verbose=1,
callbacks=callbacks)
else:
hist = model.fit(X_train, [y_train_g, y_train_a], batch_size=batch_size, epochs=nb_epochs, callbacks=callbacks,
validation_data=(X_test, [y_test_g, y_test_a]))
logging.debug("Saving weights...")
model.save_weights(os.path.join("models", "WRN_{}_{}.h5".format(depth, k)), overwrite=True)
pd.DataFrame(hist.history).to_hdf(os.path.join("models", "history_{}_{}.h5".format(depth, k)), "history")
def train_model_with_2_classes(model,train_X,train_Y,test_X,test_Y):
train_X = (train_X-0.1307)/0.3081
test_X = (test_X-0.1307)/0.3081
id_list=[]
out_dir = "models/"
if not os.path.exists(out_dir):
os.makedirs(out_dir)
for i in range(2,3):
for j in range(8,10):
id_list.append([i,j])
two_classes_id=(train_Y==i)+ (train_Y==j)
trainX=train_X[two_classes_id]
trainY=train_Y[two_classes_id]
two_classes_id=(test_Y==i)+( test_Y==j)
testX=test_X[two_classes_id]
testY=test_Y[two_classes_id]
Y_train = np_utils.to_categorical(trainY, nb_classes)
Y_test = np_utils.to_categorical(testY, nb_classes)
# Load model
weights_file = "models/shake_shake"+str(i)+'_'+str(j)+".h5"
if os.path.exists(weights_file):
model.load_weights(weights_file)
print(weights_file,"Model loaded.\r\n")
lr_reducer = ReduceLROnPlateau(monitor='val_loss', factor=0.5, # 当标准评估停止提升时，降低学习速率。
cooldown=0, patience=20, min_lr=1e-8)
model_checkpoint = ModelCheckpoint(weights_file, monitor="val_acc", save_best_only=True,
save_weights_only=True, mode='auto')
earlyStopping = EarlyStopping(monitor='acc', patience=10, verbose=1, mode='auto')
callbacks = [lr_reducer, model_checkpoint,earlyStopping]
train_data = ImageDataGenerator(featurewise_center=True,
featurewise_std_normalization=True,
# preprocessing_function=random_crop_image,
preprocessing_function=get_random_eraser(v_l=0, v_h=1),
rotation_range=10,
width_shift_range=5. / 28,
height_shift_range=5. / 28,
horizontal_flip=True)
validation_data = ImageDataGenerator(featurewise_center=True, featurewise_std_normalization=True)
for data in (train_data, validation_data):
data.fit(trainX)
model.fit_generator(train_data.flow(trainX, Y_train, batch_size=batch_size),
steps_per_epoch=len(trainX) // batch_size,
epochs=nb_epoch,
callbacks=callbacks,
validation_data=validation_data.flow(testX, Y_test, batch_size=batch_size),
validation_steps=testX.shape[0] // batch_size, verbose=1)
# model.fit(trainX, Y_train, batch_size=batch_size, epochs=1, verbose=1, validation_data=(testX, Y_test),
# callbacks=callbacks)
yPreds = model.predict(testX)
yPred = np.argmax(yPreds, axis=1)
yTrue = testY
accuracy = metrics.accuracy_score(yTrue, yPred) * 100
error = 100 - accuracy
print(i,'pk',j)
print("Accuracy : ", accuracy)
print("Error : ", error)
def main():
start = time.time()
ap = argparse.ArgumentParser()
ap.add_argument("-e",
"--epochs",
required=True,
type=int,
help="Number of epochs",
default=25)
ap.add_argument("-m",
"--model_name",
required=True,
type=str,
help="Imagenet model to train",
default="xception")
ap.add_argument("-b",
"--batch_size",
required=True,
type=int,
help="Batch size",
default=8)
ap.add_argument("-im_size",
"--image_size",
required=True,
type=int,
help="Batch size",
default=224)
args = ap.parse_args()
# Training dataset loading
train_data = np.load("train_data.npy")
train_label = np.load("train_label.npy")
encoder = LabelEncoder()
encoder.fit(train_label)
encoded_y = encoder.transform(train_label)
Y = utils.to_categorical(encoded_y)
print("Dataset Loaded...")
# Train and validation split
trainX, valX, trainY, valY = train_test_split(train_data,
Y,
test_size=0.1,
shuffle=True,
random_state=42,
stratify=Y)
print(trainX.shape, valX.shape, trainY.shape, valY.shape)
# Train nad validation image data generator
trainAug = ImageDataGenerator(
rescale=1.0 / 255.0,
preprocessing_function=get_random_eraser(p=0.5,
s_l=0.02,
s_h=0.4,
r_1=0.3,
r_2=1 / 0.3,
v_l=0,
v_h=255,
pixel_level=False),
rotation_range=30,
zoom_range=0.15,
width_shift_range=0.2,
height_shift_range=0.2,
shear_range=0.15,
horizontal_flip=True,
fill_mode="nearest",
)
valAug = ImageDataGenerator(rescale=1.0 / 255.0)
model = cnn_model(args.model_name, img_size=args.image_size)
# Number of trainable and non-trainable parameters
trainable_count = int(
np.sum([K.count_params(p) for p in set(model.trainable_weights)]))
non_trainable_count = int(
np.sum([K.count_params(p) for p in set(model.non_trainable_weights)]))
print("Total params: {:,}".format(trainable_count + non_trainable_count))
print("Trainable params: {:,}".format(trainable_count))
print("Non-trainable params: {:,}".format(non_trainable_count))
if not exists("./trained_wts"):
makedirs("./trained_wts")
if not exists("./training_logs"):
makedirs("./training_logs")
if not exists("./plots"):
makedirs("./plots")
# Keras backend
model_checkpoint = ModelCheckpoint(
"trained_wts/" + args.model_name + ".hdf5",
monitor="val_loss",
verbose=1,
save_best_only=True,
save_weights_only=True,
)
stopping = EarlyStopping(monitor="val_loss", patience=10, verbose=0)
clr = CyclicLR(mode=CLR_METHOD,
base_lr=MIN_LR,
max_lr=MAX_LR,
step_size=STEP_SIZE * (trainX.shape[0] // args.batch_size))
print("Training is going to start in 3... 2... 1... ")
# Model Training
H = model.fit_generator(
trainAug.flow(trainX, trainY, batch_size=args.batch_size),
steps_per_epoch=len(trainX) // args.batch_size,
validation_data=valAug.flow(valX, valY),
validation_steps=len(valX) // args.batch_size,
epochs=args.epochs,
callbacks=[model_checkpoint],
)
# plot the training loss and accuracy
plt.style.use("ggplot")
plt.figure()
N = args.epochs
plt.plot(np.arange(0, N), H.history["loss"], label="train_loss")
plt.plot(np.arange(0, N), H.history["val_loss"], label="val_loss")
plt.plot(np.arange(0, N), H.history["accuracy"], label="train_acc")
plt.plot(np.arange(0, N), H.history["val_accuracy"], label="val_acc")
plt.title("Training Loss and Accuracy")
plt.xlabel("Epoch #")
plt.ylabel("Loss/Accuracy")
plt.legend(loc="lower left")
plt.savefig("plots/training_plot.png")
N = np.arange(0, len(clr.history["lr"]))
plt.figure()
plt.plot(N, clr.history["lr"])
plt.title("Cyclical Learning Rate (CLR)")
plt.xlabel("Training Iterations")
plt.ylabel("Learning Rate")
plt.savefig("plots/cyclic_lr.png")
end = time.time()
dur = end - start
if dur < 60:
print("Execution Time:", dur, "seconds")
elif dur > 60 and dur < 3600:
dur = dur / 60
print("Execution Time:", dur, "minutes")
else:
dur = dur / (60 * 60)
print("Execution Time:", dur, "hours")
train_datagen = ImageDataGenerator(
rotation_range=0,
width_shift_range=0.1,
height_shift_range=0.1,
brightness_range=(0.9, 1.1),
shear_range=0.0,
zoom_range=0.2,
channel_shift_range=0.0,
fill_mode='reflect',
horizontal_flip=True,
vertical_flip=False,
rescale=1 / 255,
preprocessing_function=get_random_eraser(p=0.8,
s_l=0.02,
s_h=0.4,
r_1=0.3,
r_2=1 / 0.3,
v_l=0,
v_h=255,
pixel_level=True))
valid_datagen = ImageDataGenerator(rescale=1 / 255)
train = train_datagen.flow_from_directory(TRAIN_DIR,
target_size=IMAGE_SIZE,
color_mode='rgb',
batch_size=BATCH_SIZE,
interpolation='bicubic')
valid = valid_datagen.flow_from_directory(VAL_DIR,
target_size=IMAGE_SIZE,
color_mode='rgb',
batch_size=BATCH_SIZE,
interpolation='bicubic')
class_weights = compute_class_weight('balanced', np.arange(0, N_CLASSES),
def train_tf(image_size=64,
batch_size=128,
lr=0.001,
min_lr=0.00001,
epoch=30,
logging=True,
save_model=True,
save_result_to_csv=True,
lr_reduce_patience=5,
lr_reduce_factor=0.7,
n_fold=5,
aug_config=None,
create_model=dense_net_121_model,
three_channel=False):
gpus = tf.config.experimental.list_physical_devices('GPU')
if gpus:
try:
# Currently, memory growth needs to be the same across GPUs
for gpu in gpus:
tf.config.experimental.set_memory_growth(gpu, True)
logical_gpus = tf.config.experimental.list_logical_devices('GPU')
print(len(gpus), "Physical GPUs,", len(logical_gpus),
"Logical GPUs")
except RuntimeError as e:
# Memory growth must be set before GPUs have been initialized
print(e)
else:
raise NotImplementedError("can't train with gpu")
if aug_config is None:
aug_config = dict()
augmentations.IMAGE_SIZE = image_size
# need to set this to train in rtx gpu
train = pd.read_csv("data/train.csv")
train["image_path"] = train["image_id"].apply(
lambda x: f"data/image_{image_size}/{x}.png")
train.drop(["grapheme", "image_id"], axis=1, inplace=True)
x = train["image_path"].values
y_root = pd.get_dummies(train['grapheme_root']).values
y_vowel = pd.get_dummies(train['vowel_diacritic']).values
y_consonant = pd.get_dummies(train['consonant_diacritic']).values
print(f"overall dataset: "
f"root - {len(np.unique(train['grapheme_root'].values))} "
f"vowel - {len(np.unique(train['vowel_diacritic'].values))} "
f"con - {len(np.unique(train['consonant_diacritic'].values))}")
transformers = []
if AFFINE in aug_config:
seq_augmenter = iaa.Sequential([
iaa.Sometimes(1.0, iaa.Affine(**aug_config[AFFINE])),
])
transformers.append(lambda img: seq_augmenter.augment_image(img))
if GRID_MASK in aug_config:
transformers.append(
lambda img: grid_mask(img, **aug_config[GRID_MASK]))
if RANDOM_CUTOUT in aug_config:
transformers.append(
lambda img: get_random_eraser(**aug_config[RANDOM_CUTOUT])(img))
if AUGMIX in aug_config:
transformers.append(lambda img: augmentations.augment_and_mix(
img, **aug_config[AUGMIX]))
"""
32, 0.79
60, 0.58
61, 0.68
62, 0.73
84, 0.80
37, 0.86
45, 0.86
110, 0.87
122, 0.85
"""
skf = StratifiedKFold(n_splits=n_fold, shuffle=True)
for train_idx, test_idx in skf.split(x, train['grapheme_root'].values):
x_train, x_test = x[train_idx], x[test_idx]
y_train_root, y_test_root = y_root[train_idx], y_root[test_idx]
y_train_consonant, y_test_consonant = y_consonant[
train_idx], y_consonant[test_idx]
y_train_vowel, y_test_vowel = y_vowel[train_idx], y_vowel[test_idx]
root_truth = np.argmax(y_test_root, axis=1)
vowel_truth = np.argmax(y_test_vowel, axis=1)
con_truth = np.argmax(y_test_consonant, axis=1)
print(f"train set: "
f"root - {len(np.unique(np.argmax(y_train_root, axis=1)))} "
f"vowel - {len(np.unique(np.argmax(y_train_vowel, axis=1)))} "
f"con - {len(np.unique(np.argmax(y_train_consonant, axis=1)))}")
print(f"test set: "
f"root - {len(np.unique(root_truth))} "
f"vowel - {len(np.unique(vowel_truth))} "
f"con - {len(np.unique(con_truth))}")
c = Counter(vowel_truth)
vowel_test_percentage = [(i, c[i] / len(vowel_truth) * 100.0)
for i, count in c.most_common()]
print(f"test vowel percentage: {vowel_test_percentage}")
c = Counter(con_truth)
con_test_percentage = [(i, c[i] / len(con_truth) * 100.0)
for i, count in c.most_common()]
print(f"test con percentage: {con_test_percentage}")
input_shape = (image_size, image_size,
3) if three_channel else (image_size, image_size, 3)
train_gen = BengaliImageMixUpGenerator(
x_train,
image_size,
root=y_train_root,
vowel=y_train_vowel,
consonant=y_train_consonant,
batch_size=batch_size,
mixup=MIXUP in aug_config,
alpha=aug_config[MIXUP]["alpha"] if MIXUP in aug_config else 0.2,
transformers=transformers,
three_channel=three_channel)
test_gen = BengaliImageGenerator(x_test,
image_size,
root=y_test_root,
vowel=y_test_vowel,
consonant=y_test_consonant,
batch_size=batch_size,
three_channel=three_channel)
model = create_model(input_shape)
optimizer = Adam(learning_rate=lr)
model.compile(optimizer=optimizer,
loss='categorical_crossentropy',
metrics=['accuracy'])
callbacks = [
ReduceLROnPlateau(monitor='val_root_accuracy',
patience=lr_reduce_patience,
verbose=1,
factor=lr_reduce_factor,
min_lr=min_lr),
OnEpochEnd(train_gen),
]
aug_keys = list(aug_config.keys())
aug_keys.sort(reverse=True)
if len(aug_keys) == 0:
aug_keys.append("base")
if logging:
current_timestamp = datetime.now().strftime("%Y%m%d-%H%M%S")
logdir = f"logs/scalars/{image_size}/{create_model.__name__}/{'_'.join(aug_keys)}/{current_timestamp}"
callbacks.append(TensorBoard(log_dir=logdir))
if save_model:
Path(f"model/{create_model.__name__}").mkdir(parents=True,
exist_ok=True)
callbacks.append(
ModelCheckpoint(
f"model/{create_model.__name__}/tf_model_{image_size}_{'_'.join(aug_keys)}.h5",
monitor='val_root_accuracy',
verbose=1,
save_best_only=True,
mode='max'))
model.fit(train_gen,
epochs=epoch,
callbacks=callbacks,
validation_data=test_gen)
prediction = model.predict(test_gen)
scores = []
root_prediction = np.argmax(prediction[0], axis=1)
scores.append(
recall_score(root_truth, root_prediction, average='macro'))
# print(classification_report(root_truth, root_prediction))
vowel_pred = np.argmax(prediction[1], axis=1)
scores.append(recall_score(vowel_truth, vowel_pred, average='macro'))
# print(classification_report(vowel_truth, vowel_pred))
con_pred = np.argmax(prediction[2], axis=1)
scores.append(recall_score(con_truth, con_pred, average='macro'))
# print(classification_report(con_truth, con_pred))
cv_score = np.average(scores, weights=[2, 1, 1])
print(cv_score)
if save_result_to_csv:
info = {
"model": create_model.__name__,
"image_size": image_size,
"batch_size": batch_size,
"starting_lr": lr,
"epoch": epoch,
"lr_reduce_patience": lr_reduce_patience,
"lr_reduce_factor": lr_reduce_factor,
"min_lr": min_lr,
"augmentation": json.dumps(aug_config),
"cv_score": cv_score,
"public_cv": "",
}
with open("train_result.csv", 'a+') as write_obj:
dict_writer = DictWriter(write_obj,
fieldnames=list(info.keys()))
dict_writer.writerow(info)
break
def main():
args = get_args()
input_path = args.input
batch_size = args.batch_size
nb_epochs = args.nb_epochs
max_age = args.max_age + 1
depth = args.depth
k = args.width
transfer_learning = args.transfer_learning
validation_split = args.validation_split
use_augmentation = args.aug
initial_weights = '/home/paula/THINKSMARTER_/Model/demographics-model-prediction/pretrained_models/weights.18-4.06.hdf5'
# weight_file = '/home/paula/THINKSMARTER_/Model/age-gender-estimation-adapted/checkpoints/weights.09-4.32.hdf5'
_weight_decay = 0.0005
_use_bias = False
_weight_init = "he_normal"
logging.debug("Loading data...")
image, gender, age, _, image_size, _ = load_data(input_path)
X_data = image
y_data_g = np_utils.to_categorical(gender, 2)
y_data_a = np_utils.to_categorical(age, max_age)
if transfer_learning:
model = WideResNet(image_size, depth=depth, k=k, units_age=101)()
model.load_weights(initial_weights)
inputs = model.input
flatten = model.layers[-3].output # capa flatten
dense1 = Dense(units=2,
kernel_initializer=_weight_init,
use_bias=_use_bias,
kernel_regularizer=l2(_weight_decay),
activation="softmax")(flatten)
dense2 = Dense(units=117,
kernel_initializer=_weight_init,
use_bias=_use_bias,
kernel_regularizer=l2(_weight_decay),
activation="softmax")(flatten)
model = Model(inputs=inputs, outputs=[dense1, dense2])
# ---------------------------------
# IDEA: fine tuning (nomes entreno les dos ultimes capes)
# for layer in model.layers[:-2]:
# layer.trainable = False
else:
model = WideResNet(image_size, depth=depth, k=k, units_age=max_age)()
sgd = SGD(lr=0.1, momentum=0.9, nesterov=True)
model.compile(
optimizer=sgd,
loss=["categorical_crossentropy", "categorical_crossentropy"],
metrics=['accuracy'])
logging.debug("Model summary...")
model.count_params()
model.summary()
if args.plot_model:
plot_model(model,
to_file='experiments_pictures/model_plot.png',
show_shapes=True,
show_layer_names=True)
logging.debug("Saving model...")
mk_dir("models")
with open(os.path.join("models", "WRN_{}_{}.json".format(depth, k)),
"w") as f:
f.write(model.to_json())
mk_dir("checkpoints")
# tensorBoard = TensorBoard(log_dir='events', histogram_freq=0, batch_size=32, write_graph=True, write_grads=False, write_images=True, embeddings_freq=0, embeddings_layer_names=None, embeddings_metadata=None, embeddings_data=None)
callbacks = [
LearningRateScheduler(schedule=Schedule(nb_epochs)),
ModelCheckpoint("checkpoints/weights.{epoch:02d}-{val_loss:.2f}.hdf5",
monitor="val_loss",
verbose=1,
save_best_only=True,
mode="auto")
]
logging.debug("Running training...")
data_num = len(X_data)
indexes = np.arange(data_num)
np.random.shuffle(indexes)
X_data = X_data[indexes]
y_data_g = y_data_g[indexes]
y_data_a = y_data_a[indexes]
train_num = int(data_num * (1 - validation_split))
X_train = X_data[:train_num]
X_test = X_data[train_num:]
y_train_g = y_data_g[:train_num]
y_test_g = y_data_g[train_num:]
y_train_a = y_data_a[:train_num]
y_test_a = y_data_a[train_num:]
if use_augmentation:
datagen = ImageDataGenerator(width_shift_range=0.1,
height_shift_range=0.1,
horizontal_flip=True,
preprocessing_function=get_random_eraser(
v_l=0, v_h=255))
training_generator = MixupGenerator(X_train, [y_train_g, y_train_a],
batch_size=batch_size,
alpha=0.2,
datagen=datagen)()
hist = model.fit_generator(generator=training_generator,
steps_per_epoch=train_num // batch_size,
validation_data=(X_test,
[y_test_g, y_test_a]),
epochs=nb_epochs,
verbose=1,
callbacks=callbacks)
else:
hist = model.fit(X_train, [y_train_g, y_train_a],
batch_size=batch_size,
epochs=nb_epochs,
callbacks=callbacks,
validation_data=(X_test, [y_test_g, y_test_a]))
logging.debug("Saving weights...")
model.save_weights(os.path.join("models", "WRN_{}_{}.h5".format(depth, k)),
overwrite=True)
pd.DataFrame(hist.history).to_hdf(
os.path.join("models", "history_{}_{}.h5".format(depth, k)), "history")
with open('history_tmp.txt', 'w') as f:
for key in hist.history:
print(key, file=f)
f.write('\n')
json.dump(hist.history, f)
def train():
mnist = read_data_sets('./data/fashion', reshape=False, validation_size=0,
source_url='http://fashion-mnist.s3-website.eu-central-1.amazonaws.com/')
x_train = mnist.train.images
y_train = mnist.train.labels
x_test = mnist.test.images
y_test = mnist.test.labels
y_train = to_categorical(y_train, 10)
y_test = to_categorical(y_test, 10)
x_train = x_train.reshape((-1, 28, 28, 1))
x_test = x_test.reshape((-1, 28, 28, 1))
main_input = Input((28, 28, 1))
aux_input = Input((10,))
final_output, side_output = build_model(inputs=main_input,labels=aux_input, n=16, k=8,dropout=0.2)
model = Model(inputs=[main_input, aux_input], outputs=[final_output, side_output])
model.summary()
# optim = optimizers.SGD(lr=initial_learning_rate, momentum=0.9)
optim = optimizers.Adam(lr=3e-4)
model.compile(optimizer=optim,
loss={'main_out':losses.categorical_crossentropy,'centerlosslayer':zero_loss},loss_weights=[1,0.01], metrics=['accuracy'])
train_data = ImageDataGenerator(featurewise_center=True,
featurewise_std_normalization=True,
# preprocessing_function=random_crop_image,
preprocessing_function=get_random_eraser(v_l=0, v_h=1),
rotation_range=10,
width_shift_range=5. / 28,
height_shift_range=5. / 28,
horizontal_flip=True)
validation_data = ImageDataGenerator(featurewise_center=True, featurewise_std_normalization=True)
for data in (train_data, validation_data):
data.fit(x_train) # 実用を考えると、x_validationでのfeaturewiseのfitは無理だと思う……。
best_weights_filepath = './model/best_weights.hdf5'
earlyStopping = keras.callbacks.EarlyStopping(monitor='val_loss', patience=10, verbose=1, mode='auto')
saveBestModel = keras.callbacks.ModelCheckpoint(best_weights_filepath, monitor='val_loss', verbose=1,
save_best_only=True, mode='auto')
lr_reducer = keras.callbacks.ReduceLROnPlateau(monitor='val_loss', factor=np.sqrt(0.1), #当标准评估停止提升时，降低学习速率。
cooldown=0, patience=10, min_lr=1e-6)
dummy = np.zeros((batch_size, 1))
def gen_flow_for_two_inputs(gen,X1, y, dummy):
genX1 = gen.flow(X1, y, batch_size=batch_size, seed=666)
# genX2 = gen.flow(X1, y1, batch_size=batch_size, seed=666)
while True:
X1i = genX1.__next__()
# X2i = genX2.__next__()
yield [X1i[0], X1i[1]], [X1i[1],dummy]
hist = model.fit_generator(gen_flow_for_two_inputs(train_data, x_train, y_train,dummy),
# batch_size=batch_size,
steps_per_epoch=int(50000 / batch_size),
epochs=epochs,
verbose=1,
validation_data=gen_flow_for_two_inputs(validation_data,x_test, y_test,dummy),
validation_steps=int(10000 / batch_size),
callbacks=[earlyStopping,saveBestModel,lr_reducer]
# validation_data=([x_test, y_test_onehot], [y_test_onehot, dummy2])
)
def main():
args = get_args()
input_path = args.input
batch_size = args.batch_size
nb_epochs = args.nb_epochs
depth = args.depth
k = args.width
validation_split = args.validation_split
use_augmentation = args.aug
output_path = Path(__file__).resolve().parent.joinpath(args.output_path)
output_path.mkdir(parents=True, exist_ok=True)
logging.debug("Loading data...")
image, gender, age, _, image_size, _ = load_data(input_path)
X_data = image
y_data_g = np_utils.to_categorical(gender, 2)
y_data_a = np_utils.to_categorical(age, 101)
model = WideResNet(image_size, depth=depth, k=k)()
sgd = SGD(lr=0.1, momentum=0.9, nesterov=True)
model.compile(
optimizer=sgd,
loss=["categorical_crossentropy", "categorical_crossentropy"],
metrics=['accuracy'])
logging.debug("Model summary...")
model.count_params()
model.summary()
callbacks = [
LearningRateScheduler(schedule=Schedule(nb_epochs)),
ModelCheckpoint(str(output_path) +
"/weights.{epoch:02d}-{val_loss:.2f}.hdf5",
monitor="val_loss",
verbose=1,
save_best_only=True,
mode="auto")
]
logging.debug("Running training...")
data_num = len(X_data)
indexes = np.arange(data_num)
np.random.shuffle(indexes)
X_data = X_data[indexes]
y_data_g = y_data_g[indexes]
y_data_a = y_data_a[indexes]
train_num = int(data_num * (1 - validation_split))
X_train = X_data[:train_num]
X_test = X_data[train_num:]
y_train_g = y_data_g[:train_num]
y_test_g = y_data_g[train_num:]
y_train_a = y_data_a[:train_num]
y_test_a = y_data_a[train_num:]
if use_augmentation:
datagen = ImageDataGenerator(width_shift_range=0.1,
height_shift_range=0.1,
horizontal_flip=True,
preprocessing_function=get_random_eraser(
v_l=0, v_h=255))
training_generator = MixupGenerator(X_train, [y_train_g, y_train_a],
batch_size=batch_size,
alpha=0.2,
datagen=datagen)()
hist = model.fit_generator(generator=training_generator,
steps_per_epoch=train_num // batch_size,
validation_data=(X_test,
[y_test_g, y_test_a]),
epochs=nb_epochs,
verbose=1,
callbacks=callbacks)
else:
hist = model.fit(X_train, [y_train_g, y_train_a],
batch_size=batch_size,
epochs=nb_epochs,
callbacks=callbacks,
validation_data=(X_test, [y_test_g, y_test_a]))
logging.debug("Saving history...")
pd.DataFrame(hist.history).to_hdf(
output_path.joinpath("history_{}_{}.h5".format(depth, k)), "history")
def eraser(self, image_array):
return get_random_eraser(image_array, v_l=0, v_h=255)
featurewise_center=False, # set input mean to 0 over the dataset
samplewise_center=False, # set each sample mean to 0
featurewise_std_normalization=
False, # divide inputs by std of the dataset
samplewise_std_normalization=False, # divide each input by its std
zca_whitening=False, # apply ZCA whitening
rotation_range=
0, # randomly rotate images in the range (degrees, 0 to 180)
width_shift_range=
0.1, # randomly shift images horizontally (fraction of total width)
height_shift_range=
0.1, # randomly shift images vertically (fraction of total height)
horizontal_flip=True, # randomly flip images
vertical_flip=False, # randomly flip images
preprocessing_function=get_random_eraser(v_l=0,
v_h=1,
pixel_level=pixel_level))
# Compute quantities required for featurewise normalization
# (std, mean, and principal components if ZCA whitening is applied).
datagen.fit(x_train)
# Fit the model on the batches generated by datagen.flow().
model.fit_generator(datagen.flow(x_train, y_train, batch_size=batch_size),
steps_per_epoch=x_train.shape[0] // batch_size,
validation_data=(x_test, y_test),
epochs=epochs,
verbose=1,
workers=4,
callbacks=callbacks)
import sys
import numpy as np
from uuid import uuid1
import os
from custom_charset import markReplaceDict
from keys import alphabet as charset
import time
from random_eraser import get_random_eraser
reload(sys)
sys.setdefaultencoding('utf8')
erase = get_random_eraser(p=1,
s_l=0.01,
s_h=0.02,
r_1=0.3,
r_2=1 / 0.3,
v_l=0,
v_h=255,
pixel_level=False)
def getRandomDateList(num=20):
a1 = (2000, 1, 1, 0, 0, 0, 0, 0, 0) # 设置开始日期时间元组（2000-01-01 00：00：00）
a2 = (2050, 12, 30, 23, 59, 59, 0, 0, 0) # 设置结束日期时间元组（2050-12-31 23：59：59）
start = time.mktime(a1) # 生成开始时间戳
end = time.mktime(a2) # 生成结束时间戳
ret = []
# 随机生成10个日期字符串
for i in range(num):
t = random.randint(start, end) # 在开始和结束时间戳中随机取出一个
def main():
args = get_args()
input_path = args.input
batch_size = args.batch_size
nb_epochs = args.nb_epochs
depth = args.depth
k = args.width
validation_split = args.validation_split
use_augmentation = args.aug
logging.debug("Loading data...")
image, gender, age, _, image_size, _ = load_data(input_path)
X_data = image
y_data_a = np_utils.to_categorical(age, 101)
#custom parameters
nb_class = 2
hidden_dim = 512
vgg_model = VGGFace(include_top=False, input_shape=(224, 224, 3))
last_layer = vgg_model.get_layer('pool5').output
x = Flatten(name='flatten')(last_layer)
x = Dense(hidden_dim, activation='relu', name='fc6')(x)
x = Dense(hidden_dim, activation='relu', name='fc7')(x)
out = Dense(nb_class, activation='softmax', name='fc8')(x)
model = Model(vgg_model.input, out)
sgd = SGD(lr=0.1, momentum=0.9, nesterov=True)
model.compile(optimizer=sgd,
loss=["categorical_crossentropy"],
metrics=['accuracy'])
logging.debug("Model summary...")
model.count_params()
model.summary()
logging.debug("Saving model...")
mk_dir("models")
with open(os.path.join("models", "vgg_{}_{}.json".format(depth, k)),
"w") as f:
f.write(model.to_json())
mk_dir("checkpoints")
callbacks = [
LearningRateScheduler(schedule=Schedule(nb_epochs)),
ModelCheckpoint("checkpoints/weights.{epoch:02d}-{val_loss:.2f}.hdf5",
monitor="val_loss",
verbose=1,
save_best_only=True,
mode="auto")
]
logging.debug("Running training...")
data_num = len(X_data)
indexes = np.arange(data_num)
np.random.shuffle(indexes)
X_data = X_data[indexes]
y_data_g = y_data_g[indexes]
y_data_a = y_data_a[indexes]
train_num = int(data_num * (1 - validation_split))
X_train = X_data[:train_num]
X_test = X_data[train_num:]
y_train_g = y_data_g[:train_num]
y_test_g = y_data_g[train_num:]
y_train_a = y_data_a[:train_num]
y_test_a = y_data_a[train_num:]
if use_augmentation:
datagen = ImageDataGenerator(width_shift_range=0.1,
height_shift_range=0.1,
horizontal_flip=True,
preprocessing_function=get_random_eraser(
v_l=0, v_h=255))
training_generator = MixupGenerator(X_train, [y_train_g, y_train_a],
batch_size=batch_size,
alpha=0.2,
datagen=datagen)()
hist = model.fit_generator(generator=training_generator,
steps_per_epoch=train_num // batch_size,
validation_data=(X_test,
[y_test_g, y_test_a]),
epochs=nb_epochs,
verbose=1,
callbacks=callbacks)
else:
hist = model.fit(X_train, [y_train_g, y_train_a],
batch_size=batch_size,
epochs=nb_epochs,
callbacks=callbacks,
validation_data=(X_test, [y_test_g, y_test_a]))
logging.debug("Saving weights...")
model.save_weights(os.path.join("models", "vgg_{}_{}.h5".format(depth, k)),
overwrite=True)
pd.DataFrame(hist.history).to_hdf(
os.path.join("models", "history_{}_{}.h5".format(depth, k)), "history")
def main():
logging.debug("Reading Configuration...")
args = get_args()
Config = configparser.ConfigParser()
Config.read(args.config_path)
def ConfigSectionMap(section):
dict1 = {}
options = Config.options(section)
for option in options:
try:
dict1[option] = Config.get(section, option)
if dict1[option] == -1:
DebugPrint("skip: %s" % option)
except:
print("exception on %s!" % option)
dict1[option] = None
return dict1
#Loading Fixed parameters
input_path = ConfigSectionMap("Fixed")['input_path']
batch_size = int(ConfigSectionMap("Fixed")['batch_size'])
nb_epochs = int(ConfigSectionMap("Fixed")['nb_epochs'])
validation_split = float(ConfigSectionMap("Fixed")['validation_split'])
dim = int(ConfigSectionMap("Fixed")['dimension'])
use_augmentation = bool(ConfigSectionMap("Fixed")['use_augmentation'])
metrics_type = ConfigSectionMap("Fixed")['metrics']
history = ConfigSectionMap("Fixed")['history_save_path']
checkpoints = ConfigSectionMap("Fixed")['checkpoint_save_path']
logs_dir = ConfigSectionMap("Fixed")['log_save_path']
#Loading parameters that vary over different configurations
config_number = args.config_number
distribution = ConfigSectionMap(config_number)['distribution']
feature_extractor = ConfigSectionMap(config_number)['feature_extractor']
sigma = float(ConfigSectionMap(config_number)['sigma'])
optimizer_type = ConfigSectionMap(config_number)['optimizer']
loss_type = ConfigSectionMap(config_number)['loss']
logging.debug("Loading data...")
image, _, age, _, image_size, _ = load_data(input_path)
X_data = image
#Alter age according to distribution type
if distribution == "GaussBins":
age = age[:,np.newaxis]
lin_y = np.linspace(0,100,dim)[:,np.newaxis]
y_data_a = (1/(sigma*np.sqrt(2*np.pi)))*np.exp(-np.square((age-lin_y.T)/(np.sqrt(2)*sigma)))
elif distribution == "Cls":
y_data_a = np_utils.to_categorical(age, dim)
print(y_data_a.shape)
data_num = len(X_data)
#Randomly shuffle data
indexes = np.arange(data_num)
np.random.shuffle(indexes)
X_data = X_data[indexes]
y_data_a = y_data_a[indexes]
#Split in test train set
train_num = int(data_num * (1 - validation_split))
X_train = X_data[:train_num]
X_test = X_data[train_num:]
y_train_a = y_data_a[:train_num]
y_test_a = y_data_a[train_num:]
#Choose network
if feature_extractor == "WideResNet":
model = WideResNet(image_size, depth=16, k=8)()
#Choose optimizer
if optimizer_type == "sgd":
optimizer = SGD(lr=0.1, momentum=0.9, nesterov=True)
#Choose loss
if loss_type == "kullback_leibler_divergence":
loss = "kullback_leibler_divergence"
elif loss_type == "Wasserstein":
loss = Wasserstein
elif loss_type == "wass1":
loss = Wass1
elif loss_type == "loss1":
loss = custom_loss1
elif loss_type == "loss2":
loss = custom_loss2
elif loss_type == "loss3":
loss = "mean_squared_error"
elif loss_type == "loss4":
loss = "mean_absolute_error"
elif loss_type == "categorical_crossentropy":
loss = "categorical_crossentropy"
#Choose metric
if metrics_type == "mean_absolute_error":
metric = mean_absolute_error
#Final compilation
model.compile(optimizer=optimizer, loss=[loss], metrics=[metric])
logging.debug("Model summary...")
model.count_params()
# model.summary()
#Callbacks
json_log = open(os.path.join(logs_dir,"{}_{}_{:.5}_{}_{}.log".format(distribution,feature_extractor,loss_type,optimizer_type,sigma)),
mode='wt',
buffering=1)
logging_callback = LambdaCallback(
on_train_begin=lambda logs: json_log.write(
json.dumps({'distribution': distribution, 'feature_extractor': feature_extractor,
'loss_type': loss_type, 'optimizer_type': optimizer_type, 'sigma': sigma}) + '\n'),
on_epoch_end=lambda epoch, logs: json_log.write(
json.dumps({'epoch': epoch, 'val_mean_absolute_error': logs['val_mean_absolute_error'], 'val_loss': logs['val_loss'], 'mean_absolute_error': logs['mean_absolute_error'], 'loss': logs['loss']}) + '\n'),
on_train_end=lambda logs: json_log.close()
)
callbacks = [LearningRateScheduler(schedule=Schedule(nb_epochs)),
ModelCheckpoint(os.path.join(checkpoints,"weights.{}_{}_{:.5}_{}_{}.hdf5".format(distribution,feature_extractor,loss_type,optimizer_type,sigma)),
monitor="val_mean_absolute_error",
verbose=1,
save_best_only=True,
mode="auto"),
logging_callback,
TensorBoard(log_dir=os.path.join(logs_dir,"{}_{}_{:.5}_{}_{}/".format(distribution,feature_extractor,loss_type,optimizer_type,sigma)),
histogram_freq=0, batch_size=batch_size, write_graph=False, write_grads=False, write_images=False) ]
logging.debug("Running training...")
if use_augmentation:
datagen = ImageDataGenerator(
width_shift_range=0.1,
height_shift_range=0.1,
horizontal_flip=True,
preprocessing_function=get_random_eraser(v_l=0, v_h=255))
training_generator = MixupGenerator(X_train, y_train_a, batch_size=batch_size, alpha=0.2,
datagen=datagen)()
hist = model.fit_generator(generator=training_generator,
steps_per_epoch=train_num // batch_size,
validation_data=(X_test, y_test_a),
epochs=nb_epochs, verbose=1,
callbacks=callbacks)
else:
hist = model.fit(X_train, y_train_a, batch_size=batch_size, epochs=nb_epochs, verbose=1, callbacks=callbacks,
validation_data=(X_test, y_test_a))
logging.debug("Saving history and graphs...")
pd.DataFrame(hist.history).to_hdf(os.path.join(history, "history.{}_{}_{:.5}_{}_{}.hdf5".format(distribution,feature_extractor,loss_type,optimizer_type,sigma)), "history")
shutil.copyfile(pretrained_path, model_final)
parser = argparse.ArgumentParser()
# parser.add_argument('dataset_root')
# parser.add_argument('classes')
# parser.add_argument('result_root')
parser.add_argument('--epochs_pre', type=int, default=2)
parser.add_argument('--epochs_fine', type=int, default=2)
parser.add_argument('--batch_size_pre', type=int, default=16)
parser.add_argument('--batch_size_fine', type=int, default=16)
parser.add_argument('--lr_pre', type=float, default=1e-3)
parser.add_argument('--lr_fine', type=float, default=1e-4)
parser.add_argument('--snapshot_period_pre', type=int, default=1)
parser.add_argument('--snapshot_period_fine', type=int, default=1)
# parser.add_argument('--split', type=float, default=0.8)
eraser = get_random_eraser()
def categorical_focal_loss(gamma=2., alpha=.25):
"""
Softmax version of focal loss.
m
FL = ∑ -alpha * (1 - p_o,c)^gamma * y_o,c * log(p_o,c)
c=1
where m = number of classes, c = class and o = observation
Parameters:
alpha -- the same as weighing factor in balanced cross entropy
gamma -- focusing parameter for modulating factor (1-p)
Default value:
gamma -- 2.0 as mentioned in the paper
alpha -- 0.25 as mentioned in the paper
for img in datagen.flow(x_train[1:2], batch_size=1):
imgs.append(img[0])
if len(imgs) >= max_img_num:
break
show_imgs(imgs)
!git clone https://github.com/yu4u/cutout-random-erasing ./random_eraser
import sys
sys.path.append("/content/random_eraser/")
from random_eraser import get_random_eraser
datagen = ImageDataGenerator(
preprocessing_function = get_random_eraser(p=0.5, s_l=0.02, s_h=0.2, r_1=0.3, r_2=1/0.3, v_l=0, v_h=0))
max_img_num = 12
imgs = []
for img in datagen.flow(x_train[1:2], batch_size=1):
imgs.append(img[0])
if len(imgs) >= max_img_num:
break
show_imgs(imgs)
!git clone https://github.com/yu4u/cutout-random-erasing ./random_eraser
import sys
sys.path.append("/content/random_eraser/")
callbacks = [
LearningRateScheduler(schedule=Schedule(nb_epochs, 0.1)), #Callbacks
ModelCheckpoint(str(output_path) +
"/weights.{epoch:02d}-{val_loss:.2f}.hdf5",
monitor="val_loss",
verbose=1,
save_best_only=True,
mode="auto"),
TensorBoard(log_dir='TensorBoard/imdb-64'.format(namecall))
]
if aug:
datagen = ImageDataGenerator(width_shift_range=0.1,
height_shift_range=0.1,
horizontal_flip=True,
preprocessing_function=get_random_eraser(
v_l=0, v_h=255))
training_generator = MixupGenerator(x_train, [y_train_g, y_train_a],
batch_size=32,
alpha=0.2,
datagen=datagen)()
hist = model.fit_generator(generator=training_generator,
steps_per_epoch=train_num // batch_size,
validation_data=(x_test, [y_test_g, y_test_a]),
epochs=nb_epochs,
verbose=1,
callbacks=callbacks)
else:
hist = model.fit(x_train, [y_train_g, y_train_a],
batch_size=32,
def train():
x, y_a, y_g, y_r = load_data()
print(x.shape)
print(y_a.shape)
print(y_g.shape)
print(y_r.shape)
train_index = int(len(x) * (1 - test_split))
x_train = x[:train_index]
y_train_a = y_a[:train_index]
y_train_g = y_g[:train_index]
y_train_r = y_r[:train_index]
x_test = x[train_index:]
y_test_a = y_a[train_index:]
y_test_g = y_g[train_index:]
y_test_r = y_r[train_index:]
model = Face(train=True)
opt = Adam(lr=initial_lr)
#opt = SGD(lr=initial_lr, momentum=0.9, nesterov=True)
model.compile(optimizer=opt,
loss=[
'categorical_crossentropy', 'categorical_crossentropy',
'categorical_crossentropy'
],
metrics=['accuracy'])
callbacks = [
LearningRateScheduler(schedule=Schedule(nb_epochs, initial_lr)),
ModelCheckpoint(
weights_output_path + "/face_weights.{epoch:02d}"
"-val_loss-{val_loss:.2f}-val_age_loss-{val_predications_age_loss:.2f}"
"-val_gender_loss-{val_predications_gender_loss:.2f}-val_race_loss-{val_predications_race_loss:.2f}.utk.h5",
monitor="val_loss",
verbose=1,
save_best_only=True,
mode="auto"),
TensorBoard(log_dir='logs\{0}-{1}'.format(model.name, time.time()),
histogram_freq=1,
batch_size=batch_size,
write_graph=True,
write_grads=False,
write_images=True,
embeddings_freq=0,
embeddings_layer_names=None,
embeddings_metadata=None,
embeddings_data=None,
update_freq=500)
]
# if use data augmentation
if not data_augmentation:
history = model.fit(x_train, [y_train_a, y_train_g, y_train_r],
batch_size=batch_size,
epochs=nb_epochs,
callbacks=callbacks,
validation_data=(x_test,
[y_test_a, y_test_g, y_test_r]))
else:
datagen = ImageDataGenerator(width_shift_range=0.1,
height_shift_range=0.1,
horizontal_flip=True,
preprocessing_function=get_random_eraser(
v_l=0, v_h=255))
training_generator = MixupGenerator(x_train,
[y_train_a, y_train_g, y_train_r],
batch_size=batch_size,
alpha=0.2,
datagen=datagen)()
history = model.fit_generator(
generator=training_generator,
steps_per_epoch=len(x_train) // batch_size,
validation_data=(x_test, [y_test_a, y_test_g, y_test_r]),
epochs=nb_epochs,
verbose=1,
callbacks=callbacks)
checkpoint = ModelCheckpoint(join(refine_path,
'best_semi_%d_{val_acc:.5f}.h5' % i),
monitor='val_acc',
verbose=1,
save_best_only=True)
early = EarlyStopping(monitor="val_acc", mode="max", patience=60)
print("#" * 50)
print("Fold: ", i)
datagen = ImageDataGenerator(
featurewise_center=True, # set input mean to 0 over the dataset
width_shift_range=0.2,
height_shift_range=0.2,
horizontal_flip=True,
preprocessing_function=get_random_eraser(
v_l=np.min(X_train), v_h=np.max(X_train)) # Trainset's boundaries.
)
mygenerator = MixupGenerator(X_train,
Y_train,
alpha=1.0,
batch_size=128,
datagen=datagen)
history = model.fit_generator(mygenerator(),
steps_per_epoch=X_train.shape[0] // 128,
epochs=10000,
validation_data=(X_test, Y_test),
callbacks=[checkpoint, early])
def fine_tune(MODEL,
model_name,
optimizer,
lr,
epoch,
patience,
batch_size,
weights,
X=x_train,
test=None):
# Fine-tune the model
from random_eraser import get_random_eraser
datagen = ImageDataGenerator(horizontal_flip=True,
preprocessing_function=get_random_eraser(
v_h=60, pixel_level=True))
val_datagen = ImageDataGenerator()
inputs = Input(input_shape)
x = inputs
cnn_model = MODEL(include_top=False,
input_shape=input_shape,
weights=None)
x = cnn_model(x)
x = GlobalAveragePooling2D()(x)
x = Dropout(0.5)(x)
x = Dense(128, activation='relu', name='sim')(x)
x = Dropout(0.5)(x)
x = Dense(n_class, activation='softmax', name='predictions')(x)
model = Model(inputs=inputs, outputs=x)
# Loading weights
try:
model.load_weights(model_name + '.h5')
print('Load ' + model_name + '.h5 successfully.')
except:
if weights == 'imagenet':
print('Start computing ' + model_name +
' bottleneck feature: ')
features = get_features(MODEL, X)
# Training models
inputs = Input(features.shape[1:])
x = inputs
x = Dropout(0.5)(x)
x = Dense(128, activation='relu', name='sim')(x)
x = Dropout(0.5)(x)
x = Dense(n_class, activation='softmax', name='predictions')(x)
model_fc = Model(inputs, x)
model_fc.compile(optimizer='adam',
loss='categorical_crossentropy',
metrics=['accuracy'])
h = model_fc.fit(features,
y_train,
batch_size=128,
epochs=5,
validation_split=0.1)
model_fc.save('fc_' + model_name + '.h5')
model.load_weights('fc_' + model_name + '.h5', by_name=True)
print("Optimizer=" + optimizer + " lr=" + str(lr) + " \n")
if optimizer == "Adam":
model.compile(loss='categorical_crossentropy',
optimizer='adam',
metrics=['accuracy'])
elif optimizer == "SGD":
model.compile(loss='categorical_crossentropy',
optimizer=SGD(lr=lr, momentum=0.9, nesterov=True),
metrics=['accuracy'])
if not test:
datagen.fit(x_train)
val_datagen.fit(x_val)
class LossHistory(keras.callbacks.Callback):
def on_train_begin(self, logs={}):
self.losses = []
def on_epoch_end(self, batch, logs={}):
self.losses.append(
(logs.get('loss'), logs.get("val_loss")))
history = LossHistory()
early_stopping = EarlyStopping(monitor='val_loss',
patience=patience,
verbose=1,
mode='auto')
checkpointer = ModelCheckpoint(filepath=model_name + '.h5',
verbose=0,
save_best_only=True)
reduce_lr = ReduceLROnPlateau(factor=0.5, patience=3, verbose=1)
if optimizer == "Adam":
callbacks = [history, early_stopping, checkpointer]
else:
callbacks = [history, early_stopping, checkpointer, reduce_lr]
h = model.fit_generator(datagen.flow(x_train,
y_train,
batch_size=batch_size),
steps_per_epoch=len(x_train) / batch_size,
validation_data=val_datagen.flow(
x_val, y_val, batch_size=batch_size),
validation_steps=len(x_val) / batch_size,
epochs=epoch,
callbacks=callbacks)
return h
else:
print('Evalute on test set')
val_datagen.fit(x_test)
score = model.evaluate_generator(
val_datagen.flow(x_test, y_test, batch_size=batch_size),
len(x_test) / batch_size)
print(score)
return score
X_train, y_train, X_val, y_val = split_data(X, y, int(val_set_num))
# X_train, X_val = normalize(X_train, X_val)
X_train = np.concatenate((X_train, X_semi))
y_train = np.concatenate((y_train, y_semi))
tmpidx = list(range(len(X_train)))
random.shuffle(tmpidx)
X_train = X_train[tmpidx]
y_train = y_train[tmpidx]
datagen = ImageDataGenerator(width_shift_range=0.05,
height_shift_range=0.05,
shear_range=0.084375,
preprocessing_function=get_random_eraser(v_l=np.min(X_train),v_h=np.max(X_train)))
training_generator = MixupGenerator(X_train, y_train, batch_size=32, alpha=0.5, datagen=datagen)()
filename = os.path.join(base_path, 'phase3_mfcc4_resnet18_3/model' + val_set_num)
if not os.path.exists(filename):
os.makedirs(filename)
callback = ModelCheckpoint(filename + '/weights.{epoch:04d}-{val_loss:.4f}-{val_acc:.4f}.h5', monitor='val_acc', save_best_only=True, verbose=1)
early = EarlyStopping(monitor='val_acc', mode='max', patience=30, verbose=1)
# model = resnet.ResnetBuilder.build_resnet_152((1, 20, 690), 41)
modelfile = os.path.join(base_path, 'cnn_model_18/model' + val_set_num)
model = load_model(modelfile)
# model.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])