def train():
X = np.load('./data/train_X.npy')
y = np.load('./data/train_y.npy')
train_datagen = ImageDataGenerator(rescale=1.0 / 255.)
validation_datagen = ImageDataGenerator(rescale=1.0 / 255.)
train_generator = train_datagen.flow(X, y, batch_size=300)
validation_generator = validation_datagen.flow(X, y)
nb_epoch = 2000
nb_train_samples = N
nb_validation_samples = 600
old_session = KTF.get_session()
with tf.Graph().as_default():
session = tf.Session('')
KTF.set_session(session)
KTF.set_learning_phase(1)
model = Sequential()
with tf.name_scope("inference") as scope:
model = set_model(model)
model.summary()
fpath = './model/weights.hdf5'
tb_cb = TensorBoard(log_dir="./tensorlog", histogram_freq=1)
cp_cb = ModelCheckpoint(filepath=fpath,
monitor='val_loss',
verbose=1,
save_best_only=False,
mode='auto',
save_weights_only=True)
history = model.fit_generator(train_generator,samples_per_epoch=nb_train_samples, \
nb_epoch=nb_epoch, validation_data=validation_generator, \
nb_val_samples=nb_validation_samples,\
callbacks=[cp_cb, tb_cb])
def _wrap_savedmodel_around_keras(self, preprocess_function,
postprocess_function, model_name,
model_version, h5_filepath, serve_dir,
channels, optional_preprocess_args,
optional_postprocess_args):
""" Injects input and output layers with Keras Lambdas, then
exports to SavedModel.
Args:
preprocess_function: A function from the LayerInjector class
which preprocesses input.
postprocess_function: A function from the LayerInjector class
which postprocesses output.
model_name: The name of the model.
model_version: The version number of the model.
h5_filepath: The filepath to a .h5 file containing an
exported Keras model.
serve_dir: The path to the model's serve directory.
channels: The number of channels of the input image.
optional_preprocess_args: Optional dict of arguments for use
with custom preprocess functions.
optional_postprocess_args: Optional dict of arguments for use
with custom postprocess functions.
"""
# Parses paths
# Note that the serve directory MUST have a model version subdirectory
model_version = str(model_version)
save_path = serve_dir + "/" + model_name + "/" + model_version
# Instantiates a Keras model
K.set_learning_phase(0)
keras_model = load_model(h5_filepath)
# Instantiates placeholder for image bitstring
input_bytes = Input(shape=[], dtype=tf.string)
# Preprocesses image bitstring
pre_map = {"channels": channels, **optional_preprocess_args}
input_tensor = Lambda(preprocess_function,
arguments=pre_map)(input_bytes)
# Gets output tensor(s)
output_tensor = keras_model(input_tensor)
# Postprocesses output tensor(s)
post_map = optional_postprocess_args
output_bytes = Lambda(postprocess_function,
arguments=post_map)(output_tensor)
# Builds new Model
model = Model(input_bytes, output_bytes)
# Builds input/output tensor protos
input_tensor_info = {"input_bytes": build_tensor_info(model.input)}
output_tensor_info = {"output_bytes": build_tensor_info(model.output)}
# Creates and saves SavedModel
self._create_savedmodel(save_path, input_tensor_info,
output_tensor_info)
def encode_and_decode(mode_auto, exp_condition):
config = tf.compat.v1.ConfigProto()
config.gpu_options.allow_growth = True
sess = tf.compat.v1.Session(config=config)
keras.backend.set_session(sess)
old_session = KTF.get_session()
session = tf.compat.v1.Session('')
KTF.set_session(session)
KTF.set_learning_phase(1)
loader_ins = Loader(exp_condition["load_dir"])
loader_ins.load(gray=True, size=(196, 136)) # 横×縦
data = loader_ins.get_data(norm=True) # (None, Height, Width)
if mode_auto == "CAE":
input_shape = (data.shape[1], data.shape[2], 1)
data = np.reshape(data, (data.shape[0], data.shape[1], data.shape[2], 1)) # (None, Height, Width, 1)
elif mode_auto == "AE":
input_shape = (data.shape[1]*data.shape[2],)
data = np.reshape(data, (data.shape[0], data.shape[1]*data.shape[2],)) # (None, Height*Width, 1)
else:
raise Exception
x_train = data[:int(len(data) * exp_condition["train_rate"])]
x_val = data[int(len(data) * exp_condition["train_rate"]):]
train_auto(mode_auto=mode_auto,
x_train=x_train,
x_val=x_val,
input_shape=input_shape,
weights_dir=exp_condition["weights_dir"],
batch_size=exp_condition["batch_size"],
verbose=1,
epochs=exp_condition["epochs"],
num_compare=2
)
data = loader_ins.get_data(norm=True)
model_name = get_latest_modified_file_path(exp_condition["weights_dir"])
print(model_name, "をモデルとして分散表現化します．")
img2vec(data, model_name, mode_auto=mode_auto, mode_out="hwf")
KTF.set_session(old_session)
def main(_):
(x_train, y_train), (x_test, y_test) = load_data()
print(x_train.shape)
print(x_test.shape)
x_train = x_train.astype('float32').reshape(-1, 784) / 255.
x_test = x_test.astype('float32').reshape(-1, 784) / 255.
y_train = to_categorical(y_train)
y_test = to_categorical(y_test)
print(x_train.shape)
print(x_test.shape)
with tf.Graph().as_default() as graph:
with tf.Session() as sess:
K.set_session(sess)
K.set_learning_phase(1)
model = Mnist()
pprint(get_train_ops(graph=model.graph))
def model_fn(input_dim, labels_dim, hidden_units, learning_rate=0.1):
"""Create a Keras Sequential model with layers.
Args:
input_dim: (int) Input dimensions for input layer.
labels_dim: (int) Label dimensions for input layer.
hidden_units: [int] the layer sizes of the DNN (input layer first)
learning_rate: (float) the learning rate for the optimizer.
Returns:
A Keras model.
"""
# "set_learning_phase" to False to avoid:
# AbortionError(code=StatusCode.INVALID_ARGUMENT during online prediction.
K.set_learning_phase(False)
model = models.Sequential()
hidden_units = [int(units) for units in hidden_units.split(',')]
for units in hidden_units:
model.add(
layers.Dense(
units,
activation=relu,
input_shape=[input_dim],
kernel_initializer='glorot_uniform',
))
model.add(layers.Dropout(0.5))
model.add(
layers.BatchNormalization(epsilon=1e-03,
momentum=0.9,
weights=None))
input_dim = units
# activity_regularizer=tf.keras.regularizers.l1(0.01)
# Add a dense final layer with sigmoid function.
model.add(layers.Dense(labels_dim, activation='softmax'))
compile_model(model, learning_rate)
return model
def save_model(model):
# load tensorflow and keras backend
import tensorflow as tf
from tensorflow.python.framework import graph_util
from tensorflow.python.framework import graph_io
import keras.backend.tensorflow_backend as K
ksess = K.get_session()
print(ksess)
# transform keras model to tensorflow graph
# the output will be json-like format
K.set_learning_phase(0)
graph = ksess.graph
kgraph = graph.as_graph_def()
print(kgraph)
# define output node
num_output = 1
prefix = "output"
pred = [None]*num_output
outputName = [None]*num_output
for i in range(num_output):
outputName[i] = prefix + str(i)
pred[i] = tf.identity(model.get_output_at(i), name=outputName[i])
print('output name: ', outputName)
# convert variables in the model graph to constants
constant_graph = graph_util.convert_variables_to_constants(ksess, ksess.graph.as_graph_def(), outputName)
# save the model in .pb and .txt
output_dir = "/home/"
output_graph_name = "keras2tf.pb"
output_text_name = "keras2tf.txt"
graph_io.write_graph(constant_graph, output_dir, output_graph_name, as_text=False)
graph_io.write_graph(constant_graph, output_dir, output_text_name, as_text=True)
print('saved graph .pb at: {0}\nsaved graph .txt at: {1}'.format(
os.path.join(output_dir, output_graph_name),
os.path.join(output_dir, output_text_name)))
def model_fn(input_dim,
labels_dim,
hidden_units,
learning_rate=0.1):
K.set_learning_phase(False)
model = models.Sequential()
hidden_units = [int(units) for units in hidden_units.split(',')]
for units in hidden_units:
model.add(layers.Dense(units, activation=relu, input_shape=[input_dim],
kernel_initializer='glorot_uniform',
))
# model.add(layers.Dropout(0.5))
# model.add(layers.BatchNormalization(epsilon=1e-03, momentum=0.9, weights=None))
input_dim = units
# activity_regularizer=tf.keras.regularizers.l1(0.01)
# Add a dense final layer with sigmoid function.
model.add(layers.Dense(labels_dim, activation='softmax'))
# compile_model(model, learning_rate)
return model
def main(args):
# set the necessary list
train_list = pd.read_csv(args.train_list, header=None)
val_list = pd.read_csv(args.val_list, header=None)
# set the necessary directories
trainimg_dir = args.trainimg_dir
trainmsk_dir = args.trainmsk_dir
valimg_dir = args.valimg_dir
valmsk_dir = args.valmsk_dir
# get old session
old_session = KTF.get_session()
with tf.Graph().as_default():
session = tf.Session('')
KTF.set_session(session)
KTF.set_learning_phase(1)
# set callbacks
fpath = './pretrained_mask/LIP_PSPNet50_mask{epoch:02d}.hdf5'
cp_cb = ModelCheckpoint(filepath=fpath,
monitor='val_loss',
verbose=1,
save_best_only=True,
mode='auto',
period=5)
es_cb = EarlyStopping(monitor='val_loss',
patience=2,
verbose=1,
mode='auto')
tb_cb = TensorBoard(log_dir="./pretrained_mask", write_images=True)
# set generater
train_gen = data_gen_small(trainimg_dir, trainmsk_dir, train_list,
args.batch_size,
[args.input_shape[0], args.input_shape[1]],
args.n_labels)
val_gen = data_gen_small(valimg_dir, valmsk_dir, val_list,
args.batch_size,
[args.input_shape[0], args.input_shape[1]],
args.n_labels)
# set model
pspnet = PSPNet50(input_shape=args.input_shape,
n_labels=args.n_labels,
output_mode=args.output_mode,
upsample_type=args.upsample_type)
print(pspnet.summary())
# compile model
pspnet.compile(loss=args.loss,
optimizer=args.optimizer,
metrics=["accuracy"])
# fit with genarater
pspnet.fit_generator(generator=train_gen,
steps_per_epoch=args.epoch_steps,
epochs=args.n_epochs,
validation_data=val_gen,
validation_steps=args.val_steps,
callbacks=[cp_cb, es_cb, tb_cb])
# save model
with open("./pretrained_mask/LIP_SegUNet_mask.json", "w") as json_file:
json_file.write(json.dumps(json.loads(segunet.to_json()), indent=2))
print("save json model done...")
def Network_config(class_num=4,
epoch=200,
initial_epoch=0,
batch_size=32,
train_data=None,
train_label=None,
test_data=None,
test_label=None,
fold=0):
adam = Adam(lr=0.005, beta_1=0.9, beta_2=0.999, epsilon=1e-08, decay=0.000)
sgd = SGD(lr=0.001, momentum=0.9, decay=0.0, nesterov=False)
K.set_learning_phase(1)
base_model = VGG16(input_tensor=Input(shape=(224, 224, 3)),
weights='imagenet',
include_top=False)
x = base_model.output
x = Flatten()(x)
x = Dense(512, activation='relu')(x)
x = BatchNormalization()(x)
x = Dense(512, activation='relu')(x)
x = BatchNormalization()(x)
predictions = Dense(class_num, activation='softmax')(x)
# this is the model we will train
model = Model(inputs=base_model.input, outputs=predictions)
for layer in (base_model.layers):
layer.trainable = False
if layer.name.startswith('bn'):
layer.call(layer.input, training=False)
model.compile(optimizer=adam,
loss='categorical_crossentropy',
metrics=[keras.metrics.categorical_accuracy])
tools.create_directory('./tmpvgg/')
weights_file = './tmpvgg/' + str(
fold
) + '-weights.{epoch:02d}-{categorical_accuracy:.4f}-{val_loss:.4f}-{val_categorical_accuracy:.4f}.h5'
csv_file = './tmpvgg/record.csv'
lr_reducer = ReduceLROnPlateau(monitor='categorical_accuracy',
factor=0.5,
cooldown=0,
patience=5,
min_lr=0.5e-6)
early_stopper = EarlyStopping(monitor='val_categorical_accuracy',
min_delta=1e-4,
patience=50)
model_checkpoint = ModelCheckpoint(weights_file,
monitor='val_categorical_accuracy',
save_best_only=True,
verbose=2,
save_weights_only=True,
mode='max')
tensorboard = TensorBoard(log_dir='./logs/',
histogram_freq=0,
batch_size=8,
write_graph=True,
write_grads=True,
write_images=True,
embeddings_freq=0,
embeddings_layer_names=None,
embeddings_metadata=None)
CSV_record = CSVLogger(csv_file, separator=',', append=True)
callbacks = [
lr_reducer, early_stopper, model_checkpoint, tensorboard, CSV_record
]
gc.disable()
model.fit_generator(
generator=tools.batch_generator(np.array(train_data),
np.array(train_label), batch_size,
True, class_num),
steps_per_epoch=int(len(train_label) / batch_size) - 1,
max_q_size=50,
initial_epoch=initial_epoch,
epochs=epoch,
verbose=1,
callbacks=callbacks,
validation_data=tools.batch_generator(np.array(test_data),
np.array(test_label), batch_size,
True, class_num),
validation_steps=int(len(test_label) / batch_size) - 1,
class_weight='auto')
all_y_pred = []
all_y_true = []
for test_data_batch, test_label_batch in tools.batch_generator_confusion_matrix(
np.array(test_data), np.array(test_label), batch_size, True,
class_num):
y_pred = model.predict(test_data_batch, batch_size)
y_true = test_label_batch
for y_p in y_pred:
all_y_pred.append(np.where(y_p == max(y_p))[0][0])
for y_t in y_true:
all_y_true.append(np.where(y_t == max(y_t))[0][0])
confusion = confusion_matrix(y_true=all_y_true, y_pred=all_y_pred)
print(confusion)
f = open('confusion_matrix.txt', 'a+')
f.write(str(all_y_true) + "\n")
f.write(str(all_y_pred) + "\n")
f.write(str(confusion) + '\n')
f.close()
gc.enable()
def learning_star(data, label):
X_train, X_test, y_train, y_test = train_test_split(data, label, test_size=0.2, random_state=29)
"""
Add TensorBoard
"""
old_session = tf.Session('')
session = tf.Session('')
KTF.set_session(session)
KTF.set_learning_phase(1)
"""
Define the model
"""
model = Sequential()
model.add(Conv2D(30, 3, input_shape=(data.shape[1], data.shape[2], data.shape[3])))
model.add(Activation('relu'))
model.add(Conv2D(30, 3))
model.add(Activation('relu'))
model.add(MaxPool2D(pool_size=(2, 2)))
model.add(Conv2D(30, 3))
model.add(Activation('relu'))
model.add(MaxPool2D(pool_size=(2, 2)))
model.add(Flatten())
model.add(Dense(1024))
model.add(Activation('relu'))
model.add(Dropout(1.0))
model.add(Dense(nb_classes, activation='softmax'))
adam = Adam(lr=1e-4)
model.compile(optimizer=adam, loss='categorical_crossentropy', metrics=["accuracy"])
"""
Add callbacks (TensorBoard, EarlyStopping)
"""
tb_cb = keras.callbacks.TensorBoard('./logs/', histogram_freq=1)
early_stopping = EarlyStopping(monitor='val_loss', mode='min', patience=20)
cbks = [tb_cb, early_stopping]
"""
Display Model Summary
"""
plot_model(model, to_file='./model.png')
"""
Learning Flow
"""
history = model.fit(X_train, y_train, batch_size=batch_size, epochs=nb_epoch, verbose=1, validation_split=0.1, callbacks=cbks)
score = model.evaluate(X_test, y_test, verbose=0)
print('Test Score:', score[0])
print('Test accuracy:', score[1])
"""
Save Model
"""
print('save the architechture of the model')
json_string = model.to_json()
open(os.path.join('./', 'cnn_model.json'), 'w').write(json_string)
yaml_string = model.to_yaml()
open(os.path.join('./', 'cnn_model.yaml'), 'w').write(yaml_string)
print('save the weights')
model.save_weights(os.path.join('./', 'cnn_model_weight.hdf5'))
"""
def main():
parser = train_parser()
args = parser.parse_args()
if args.use_cpu:
# disable gpu completely
os.environ["CUDA_VISIBLE_DEVICES"] = "-1"
else:
# set device number
os.environ["CUDA_VISIBLE_DEVICES"] = args.gpu
if args.seed != -1:
# set random seed
np.random.seed(args.seed)
random.seed(args.seed)
tf.set_random_seed(args.seed)
# get old session old_session = KTF.get_session()
with tf.Graph().as_default():
session = tf.Session('')
KTF.set_session(session)
KTF.set_learning_phase(1)
# determine some variables from args
dims, input_shape, n_labels, model_dir, checkpoint_path = find_input(
args)
os.makedirs(model_dir, exist_ok=True)
# set callbacks
# for multiple checkpoints set filepath to "...point{epoch:d}.hdf5"
cp_cb = MyCheckpoint(filepath=model_dir + "/checkpoint.hdf5",
monitor='val_loss',
verbose=1,
save_best_only=False,
mode='auto',
period=1)
tb_cb = TensorBoard(log_dir=model_dir, write_images=True)
log = CSVLogger(model_dir + "/training.log", append=True)
# set generator
if args.use_random:
# create random data
train_gen = data_gen_random(dims, n_labels, args.batch_size,
args.use_flow)
val_gen = data_gen_random(dims, n_labels, args.batch_size,
args.use_flow)
elif args.use_numpy:
# load small numpy dataset
n_labels = args.n_classes
datagen = DataStorageNpy(args.train_imgs_npy, args.train_mask_npy,
args.train_flow_npy, args.val_percent,
dims, n_labels, args.use_flow,
args.batch_size)
train_gen = datagen.generate(training=True)
val_gen = datagen.generate(training=False)
else:
# load dataset from folders
datagen = DataStorage(args.data_dir, args.dataset, dims, n_labels)
train_gen = datagen.generate(
args.batch_size,
args.use_flow,
args.use_mb,
args.use_occ,
args.train_categories,
split="train",
shuffle=True,
normalize_images=not args.no_norm_images)
val_gen = datagen.generate(
args.batch_size,
args.use_flow,
args.use_mb,
args.use_occ,
args.train_categories,
split="val",
shuffle=False,
random_crop=False,
normalize_images=not args.no_norm_images)
if args.test_generator:
imgs, labels = next(train_gen)
print("datagen yielded", imgs.shape, labels.shape)
return
# # test has flow and images, but masks are zero
# # so it cannot be used to compute accuracy
# test_gen = datagen.generate(args.batch_size, args.use_flow,
# args.use_mb, args.use_occ,
# split="test")
# ----- determine class weights
weight_mask, weight_mask_iou = find_masks(args, n_labels)
print("using weight mask:", weight_mask)
print("using weight mask for iou:", weight_mask_iou)
# ----- define loss
old_loss = args.loss
print("old loss", old_loss)
loss = weighted_categorical_crossentropy(weight_mask, args.weight_mult)
# ----- define iou metric
def mean_iou(y_true, y_pred):
return mean_iou_func(y_true,
y_pred,
args.batch_size,
dims,
n_labels,
weight_mask_iou,
iou_test=False)
# ----- restart or load model
restart = args.restart
if not args.restart:
# check if model is available
if not os.path.exists(checkpoint_path):
print("no previous model available, restarting from epoch 0")
restart = True
print("model input shape", input_shape)
if restart:
# set model
pspnet = PSPNet50(input_shape=input_shape,
n_labels=n_labels,
output_mode=args.output_mode,
upsample_type=args.upsample_type)
# compile model
pspnet.compile(loss=loss,
optimizer=args.optimizer,
metrics=["accuracy", mean_iou])
# metrics=["acc_iou", mean_iou])
starting_epoch = 0
else:
# load model from dir
try:
pspnet = load_model(checkpoint_path,
custom_objects={
'mean_iou': mean_iou,
'loss': loss
})
except OSError:
raise OSError("failed loading checkpoint", checkpoint_path)
except ValueError:
raise ValueError(
"possible the checkpoint file is corrupt because the "
"script died while saving. use the backup old_checkpoint")
try:
with open(model_dir + "/epochs.txt", "r") as fh:
starting_epoch = int(fh.read())
except OSError:
raise FileNotFoundError("could not find epochs.txt")
print("reloading model epoch", starting_epoch)
# print model summary (weights, layers etc)
if args.summary:
print(pspnet.summary())
# ----- fit with genarater
pspnet.fit_generator(generator=train_gen,
steps_per_epoch=args.epoch_steps,
epochs=args.n_epochs,
validation_data=val_gen,
validation_steps=args.val_steps,
callbacks=[cp_cb, tb_cb, log],
initial_epoch=starting_epoch)
output = clip_ops.clip_by_value(output, epsilon_, 1 - epsilon_)
# print("target shape=", target.shape)
wei = target[:, :, :, :, -1:]
target = target[:, :, :, :, :-1]
# print("target shape=", target.shape, ", wei shape=", wei.shape, "output shape ", output.shape)
output = output[:, :, :, :, :-1]
output = math_ops.log(output / (1 - output))
retval = nn.weighted_cross_entropy_with_logits(
targets=target, logits=output,
pos_weight=10) #900=works #2900=200x200, 125=30x30
retval = retval * (wei + 0.01)
# print("output xentr ", retval)
return tf.reduce_sum(retval, axis=None) / (tf.reduce_sum(wei, axis=None))
K.set_learning_phase(0)
if args.theano_backend:
K.set_image_data_format('channels_first')
else:
K.set_image_data_format('channels_last')
try:
net_model = load_model(weight_file_path,
custom_objects={
'loss_mse_select_clipped':
loss_mse_select_clipped,
'loss_ROIsoft_crossentropy':
loss_ROI_crossentropy,
'_to_tensor': _to_tensor
})
except ValueError as err:
def run():
env = RoboEnv(client)
nb_actions = env.action_space.n
for i in range(50):
print('simulation start: %d' % i)
old_session = KTF.get_session()
with tf.Graph().as_default():
session = tf.Session('')
KTF.set_session(session)
KTF.set_learning_phase(1)
if not os.path.isfile(model_save_path):
model = Sequential()
model.add(
Flatten(input_shape=(1, ) + env.observation_space.shape))
model.add(Dense(16))
model.add(Activation('relu'))
model.add(Dense(16))
model.add(Activation('relu'))
model.add(Dense(16))
model.add(Activation('relu'))
model.add(Dense(nb_actions))
model.add(Activation('linear'))
else:
with open(model_save_path, 'r') as f:
model = model_from_json(f.read())
if os.path.isfile(weight_save_path):
model.load_weights(weight_save_path)
print(model.summary())
# experience replay用のmemory
memory = SequentialMemory(limit=50000, window_length=1)
# 行動方策はオーソドックスなepsilon-greedy。ほかに、各行動のQ値によって確率を決定するBoltzmannQPolicyが利用可能
policy = EpsGreedyQPolicy(eps=0.1)
dqn = DQNAgent(model=model,
nb_actions=nb_actions,
memory=memory,
nb_steps_warmup=100,
target_model_update=1e-2,
policy=policy)
dqn.compile(Adam(lr=1e-3), metrics=['mae'])
tb = TensorBoard(log_dir='data/tensorboard',
histogram_freq=0,
write_graph=True)
history = dqn.fit(env,
nb_steps=50000,
visualize=False,
callbacks=[tb],
verbose=2,
nb_max_episode_steps=5000)
json_string = model.to_json()
with open(model_save_path, 'w') as f:
f.write(json_string)
model.save_weights(weight_save_path)
with open(
'data/history/history%s.pickle' %
time.strftime('%Y%m%d-%H%M%S'), 'wb') as f:
pickle.dump(history.history, f)
KTF.set_session(old_session)
from keras.utils import np_utils
from keras.models import model_from_json
M = 10
# data
(x_train, y_train), (x_test, y_test) = mnist.load_data()
x_train = x_train.astype('f') / 255.0 # (60000, 28, 28)
x_test = x_test.astype('f') / 255.0
y_train = np_utils.to_categorical(y_train, M) # int -> one-of-vector
y_test = np_utils.to_categorical(y_test, M)
with tf.Graph().as_default():
session = tf.Session('')
K.set_session(session)
K.set_learning_phase(1)
try:
json_file = open('model.json', 'r')
loaded_model_json = json_file.read()
json_file.close()
loaded_model = model_from_json(loaded_model_json)
loaded_model.load_weights("model.h5")
print("Loaded model from disk")
except:
# model construction
model = Sequential()
model.add(Reshape((28, 28, 1),
input_shape=(28, 28))) # tensorflow-order!!
model.add(Conv2D(8, (5, 5), strides=(2, 2), activation='relu'))
model.add(Conv2D(16, (5, 5), strides=(2, 2), activation='relu'))
def main(args):
# set the necessary list
train_list = pd.read_csv(args.train_list, header=None)
val_list = pd.read_csv(args.val_list, header=None)
# set the necessary directories
trainimg_dir = args.trainimg_dir
trainmsk_dir = args.trainmsk_dir
valimg_dir = args.valimg_dir
valmsk_dir = args.valmsk_dir
# get old session old_session = KTF.get_session()
with tf.Graph().as_default():
session = tf.Session('')
KTF.set_session(session)
KTF.set_learning_phase(1)
# class weights
classes = ['background', 'hat', 'hair', 'glove', 'sunglasses', 'upperclothes',
'dress', 'coat', 'socks', 'pants', 'jumpsuits', 'scarf', 'skirt',
'face', 'leftArm', 'rightArm', 'leftLeg', 'rightLeg', 'leftShoe','rightShoe']
if args.class_weights:
"""
class_weights = {0:1, 1:40, 2:1, 3:114, 4:151, 5:3, 6:53, 7:7, 8:165, 9:7, 10:106,
11:249, 12:150, 13:1, 14:1, 15:1, 16:1, 17:1, 18:114, 19:118}
"""
class_weights = [1, 40, 1, 114, 151, 3, 53, 7, 165, 7, 106, 249, 150, 1, 1, 1, 1, 1, 114, 118]
# set callbacks
fpath = "./pretrained_class_weights/LIP_PSPNet50_class_weights{epoch:02d}.hdf5"
cp_cb = ModelCheckpoint(filepath = fpath, monitor='val_loss', verbose=1, save_best_only=True, mode='auto', period=2)
es_cb = EarlyStopping(monitor='val_loss', patience=2, verbose=1, mode='auto')
tb_cb = TensorBoard(log_dir="./pretrained_class_weights", write_images=True)
# set generater
train_gen = data_gen_small(
trainimg_dir,
trainmsk_dir,
train_list,
args.batch_size,
[args.input_shape[0], args.input_shape[1]],
args.n_labels)
val_gen = data_gen_small(
valimg_dir,
valmsk_dir,
val_list,
args.batch_size,
[args.input_shape[0], args.input_shape[1]],
args.n_labels)
# set model
pspnet = PSPNet50(
input_shape=args.input_shape,
n_labels=args.n_labels,
output_mode=args.output_mode,
upsample_type=args.upsample_type)
print(pspnet.summary())
# compile model
pspnet.compile(
loss=args.loss,
optimizer=args.optimizer,
metrics=["accuracy"])
# fit with genarater
pspnet.fit_generator(
generator=train_gen,
steps_per_epoch=args.epoch_steps,
epochs=args.n_epochs,
validation_data=val_gen,
validation_steps=args.val_steps,
class_weight=class_weights,
callbacks=[cp_cb, es_cb, tb_cb])
# save model
with open("./pretrained_class_weights/LIP_PSPNet50.json", "w") as json_file:
json_file.write(json.dumps(json.loads(pspnet.to_json()), indent=2))
print("save json model done...")
def main(fname_lift_train,
fname_shape_train,
fname_lift_test,
fname_shape_test,
case_number,
case_type=3,
env="Lab"):
r_rate = [1, 2, 4, 8]
for rr in r_rate:
if rr == 1:
s_odd = 0 # 全部読みだす
elif fname_shape_train.find("fourier") != -1:
s_odd = 3 # 前方から読み出す(fourier用)
else:
s_odd = 4 # 全体にわたって等間隔に読み出す(equidistant, dense用)
old_session = KTF.get_session()
with tf.Graph().as_default():
source = "Incompressible_Invicid\\training_data\\"
if env == "Lab":
source = "D:\\Toyota\\Data\\" + source
case_num = get_case_number(source, env, case_number)
log_name = "learned\\" + case_num + "_tb_log.hdf5"
json_name = "learned\\" + case_num + "_mlp_model_.json"
weight_name = "learned\\" + case_num + "_mlp_weight.h5"
elif env == "Colab":
source = "/content/drive/Colab Notebooks/" + source.replace(
"\\", "/")
case_num = get_case_number(source, env, case_number)
log_name = "learned/" + case_num + "_log.hdf5"
json_name = "learned/" + case_num + "_mlp_model_.json"
weight_name = "learned/" + case_num + "_mlp_weight.h5"
session = tf.Session('')
KTF.set_session(session)
KTF.set_learning_phase(1)
model = Sequential()
if case_type == 3:
X_train, y_train = read_csv_type3(source,
fname_lift_train,
fname_shape_train,
shape_odd=s_odd,
read_rate=rr)
x_test, y_test = read_csv_type3(source,
fname_lift_test,
fname_shape_test,
shape_odd=s_odd,
read_rate=rr)
input_vector_dim = X_train.shape[1]
with tf.name_scope("inference") as scope:
model.add(Dense(units=2, input_dim=input_vector_dim))
model.add(LeakyReLU())
model.add(Dense(units=16))
model.add(LeakyReLU())
"""
model.add(Dense(units=128))
model.add(LeakyReLU())
model.add(Dense(units=192))
model.add(LeakyReLU())
model.add(Dense(units=2048))
model.add(LeakyReLU())
model.add(Dense(units=2048))
model.add(LeakyReLU())
"""
"""
model.add(Dense(units=512))
model.add(LeakyReLU())
for i in range(5):
model.add(Dense(units = 512))
model.add(LeakyReLU())
# model.add(Dropout(0.5))
# model.add(Dense(units=half, activation='relu'))
# model.add(Dropout(0.5))
"""
model.add(Dense(units=1))
model.summary()
save_my_log(source, case_number, fname_lift_train,
fname_shape_train, model.summary())
# es_cb = EarlyStopping(monitor='val_loss', patience=0, verbose=0, mode='auto')
tb_cb = TensorBoard(log_dir=source + log_name,
histogram_freq=0,
write_grads=True)
model.compile(loss="mean_squared_error", optimizer='Adam')
batch_size = 500
train_steps, train_batches = batch_iter(X_train, y_train,
batch_size)
valid_steps, valid_batches = batch_iter(x_test, y_test, batch_size)
"""
model.fit(x=X_train, y=y_train,
batch_size=600, nb_epoch=1000,
validation_split=0.05, callbacks=[tb_cb])
"""
model.fit_generator(train_batches,
train_steps,
epochs=1000,
validation_data=valid_batches,
validation_steps=valid_steps,
callbacks=[tb_cb])
# X_train: [number, angle, shape001, shape002, ..., shapeMAX]
# y_train: [number, lift]
# 適当に中央付近の翼を抜き出しての-40-38degreeをプロットさせてみる
tekito = 1306 * 40 # NACA2613 or NACA2615
plt.figure()
plt.plot(X_train[tekito:tekito + 40, 0],
y_train[tekito:tekito + 40])
plt.plot(X_train[tekito:tekito + 40, 0],
model.predict(X_train)[tekito:tekito + 40])
plt.savefig(source + case_num + "_train.png")
y_predict = model.predict(x_test)
tekito = (99 + 13) * 40 # 22012
plt.figure()
plt.plot(x_test[tekito:tekito + 40, 0], y_test[tekito:tekito + 40])
plt.plot(x_test[tekito:tekito + 40, 0],
y_predict[tekito:tekito + 40])
plt.savefig(source + case_num + "_test.png")
json_string = model.to_json()
open(source + json_name, 'w').write(json_string)
model.save_weights(source + weight_name)
KTF.set_session(old_session)
def cnn_train(self, X_train, Y_train, X_test, Y_test):
conv1 = self.conv1
conv2 = self.conv2
conv3 = self.conv3
dense1 = self.dense1
dense2 = self.dense2
# ニュートラルネットワークで使用するモデル作成
old_session = KTF.get_session()
# old_session = tf.compat.v1.keras.backend.get_session()
with tf.Graph().as_default():
session = tf.Session('')
KTF.set_session(session)
KTF.set_learning_phase(1)
model = Sequential()
model.add(
Conv2D(conv1,
kernel_size=(3, 3),
activation='relu',
input_shape=(X_train.shape[1:])))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Conv2D(conv2, (3, 3), activation='relu'))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Conv2D(conv3, (3, 3), activation='relu'))
# model.add(Dropout(0.25))
# model.add(Conv2D(128, (3, 3), padding='same',activation='relu'))
# model.add(MaxPooling2D(pool_size=(2, 2)))
# model.add(Dropout(0.25))
model.add(Flatten())
model.add(Dense(dense1, activation='relu'))
# model.add(Dropout(0.5))
model.add(Dense(self.classnum, activation='softmax'))
model.summary()
# optimizer には adam を指定
adam = keras.optimizers.Adam(lr=self.learning_rate)
model.compile(loss='categorical_crossentropy',
optimizer='adam',
metrics=['accuracy'])
# model.compile(loss='sparse_categorical_crossentropy',optimizer='adam',metrics=['accuracy'])
es_cb = keras.callbacks.EarlyStopping(monitor='val_loss',
min_delta=0,
patience=100,
verbose=0,
mode='auto')
tb_cb = keras.callbacks.TensorBoard(log_dir=self.f_log,
histogram_freq=1)
# cp_cb = keras.callbacks.ModelCheckpoint(filepath = os.path.join(f_model,'cnn_model{epoch:02d}-loss{loss:.2f}-acc{acc:.2f}-vloss{val_loss:.2f}-vacc{val_acc:.2f}.hdf5'), monitor='val_loss', verbose=1, save_best_only=True, mode='auto')
# cbks = [es_cb, tb_cb, cp_cb]
cbks = [es_cb, tb_cb]
# cbks = [tb_cb]
history = model.fit(X_train,
Y_train,
batch_size=self.nb_batch,
epochs=self.nb_epoch,
validation_data=(X_test, Y_test),
callbacks=cbks,
verbose=1)
score = model.evaluate(X_test, Y_test, verbose=0)
print('Test score:', score[0])
print('Test accuracy:', score[1])
print('save the architecture of a model')
json_string = model.to_json()
open(os.path.join(self.f_model, 'cnn_model.json'),
'w').write(json_string)
yaml_string = model.to_yaml()
open(os.path.join(self.f_model, 'cnn_model.yaml'),
'w').write(yaml_string)
print('save weights')
model.save_weights(os.path.join(self.f_model, 'cnn_weights.hdf5'))
KTF.set_session(old_session)
return
# https://keras.io/getting-started/faq/#how-can-i-obtain-the-output-of-an-intermediate-layer
if SAVE_HIDDEN:
layers_of_interest = [
'conv1d_1', 'conv1d_2', 'reshape_2', 'lstm_1', 'dense_1', 'dense_2'
]
np.random.seed(0)
rand_indices = np.random.randint(0, x_test.shape[0], 250)
print('Saving hidden layers: ', layers_of_interest)
tfs.get_keras_layers(
model,
layers_of_interest,
x_test[rand_indices],
y_test[rand_indices],
output_dir=tfs.prep_dir(
'I:/_ecg_data_backup/classification/hidden_layers'),
fname='rat_hidden_all_' + file_name + '.mat')
print('Elapsed Time (ms): ', tfs.current_time_ms() - start_time_ms)
print('Elapsed Time (min): ',
(tfs.current_time_ms() - start_time_ms) / 60000)
data_dict = {'x_val': x_test, 'y_val': y_test, 'y_prob': yy_probabilities}
savemat(tfs.prep_dir(output_folder) + file_name + '.mat', mdict=data_dict)
tf_backend.set_learning_phase(0)
tfs.export_model_keras(keras_model_name,
tfs.prep_dir("graph_rat"),
model_name=description)
# Results:
def __enter__(self):
self.learning_phase_placeholder = K.learning_phase()
K.set_learning_phase(self.value)
def main(args):
aggregator = Aggregator(_base_dir = args.base_dir,
_img_dir = args.img_dir,
_label_dir = args.label_dir,
_inf_dir = args.inf_dir,
_dag_dir = args.dag_dir,
_poses_dir = args.poses_dir)
print('Num of ground truth labeled images %d\n\n' % len(aggregator.agg_list))
for i in range(0, 100):
print('\nDAgger Iteration: %d\n' % aggregator.dag_it_num)
# creates new directories each iteration
aggregator.on_new_iter()
# returns the training, validation lists and knowledge of index at which index evaluation in agg_list starts
train, val, idx_eval = aggregator.get_training_data()
# set directory to save predictions of inference
inf_dir = aggregator.dag_dir + '%02d/inf/' % aggregator.dag_it_num
# initiates the evaluator
evaluator = Evaluator(aggregator.base_dir, inf_dir, aggregator.label_dir,
_agg_list = aggregator.agg_list)
# estimating a batch each process should evaluate later to don't exceed a given process number
evaluator.estimate_batch_size(len(aggregator.agg_list[idx_eval:]))
print('Evaluating %d images in %d threads' % (evaluator.batch_size, evaluator.num_max_threads))
if aggregator.dag_done or evaluator.stop_dagger:
aggregator.save_list()
print('DAgger stopped!')
break
aggregator.save_list(train, 'train')
aggregator.save_list(val, 'val')
# Training and Prediction
os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
os.environ["CUDA_VISIBLE_DEVICES"] = '0'
with tf.Graph().as_default():
session = tf.Session('')
KTF.set_session(session)
KTF.set_learning_phase(1)
# initializes a trainer with already separated training and validation list
# an inference is done for all images in the agg_list after the idx_eval
trainer = Trainer(_train_list = train,
_val_list = val,
_inf_list = aggregator.agg_list[idx_eval:],
_base_dir = aggregator.base_dir,
_img_dir = aggregator.img_dir,
_label_dir = aggregator.label_dir,
_inf_dir = aggregator.inf_dir,
_dag_dir = aggregator.dag_dir,
_log_dir = 'log/')
# setting of some hyper parameters
trainer.batch_size = 8
# increasing epoch steps so the net has the chance to see all images of the training set in an epoch
# else it is observable that the validation loss doesn't decrease and the training is stopped
trainer.epoch_steps = len(train) // trainer.batch_size
trainer.val_steps = len(val) // trainer.batch_size
trainer.n_epochs = 25
trainer.dag_it = aggregator.dag_it_num
trainer.update_callback()
# trains model for defined number of epochs with the actual dataset
trainer.train()
print('\nTraining done!\nStarting Prediction\n')
# safes inferences of images that are unseen by the net
trainer.predict()
session.close()
print('\nInference done!\n')
print('Evaluating %d images' % len(aggregator.agg_list[idx_eval:]))
# Training and prediction done
# Evaluation
aggregator.agg_list = evaluator.process_prediction(agg_chunk = aggregator.agg_list,
idx_eval = idx_eval)
print('Evaluation done. Saving evaluated data.')
aggregator.save_list(aggregator.agg_list[idx_eval:], 'eval')
# Evaluation done and saved for next iteration
# save full aggregation list with all information of all iterations until this in iteration's folder
aggregator.save_list()
# delete all images of inference step to save space on the drive
aggregator.delete_inf()
aggregator.prepare_next_it()
def cnn_train_noneval(self, X_train, Y_train):
conv1 = self.conv1
conv2 = self.conv2
conv3 = self.conv3
dense1 = self.dense1
dense2 = self.dense2
# ニュートラルネットワークで使用するモデル作成
old_session = KTF.get_session()
#old_session = tf.compat.v1.keras.backend.get_session()
with tf.Graph().as_default():
session = tf.Session('')
KTF.set_session(session)
KTF.set_learning_phase(1)
model = Sequential()
model.add(
Conv2D(conv1,
kernel_size=(3, 3),
activation='relu',
input_shape=(X_train.shape[1:])))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Conv2D(conv2, (3, 3), activation='relu'))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Conv2D(conv3, (3, 3), activation='relu'))
# model.add(Dropout(0.25))
# model.add(Conv2D(128, (3, 3), padding='same',activation='relu'))
# model.add(MaxPooling2D(pool_size=(2, 2)))
# model.add(Dropout(0.25))
model.add(Flatten())
model.add(Dense(dense1, activation='relu'))
# model.add(Dropout(0.5))
model.add(Dense(self.classnum, activation='softmax'))
model.summary()
# optimizer には adam を指定
adam = keras.optimizers.Adam(lr=self.learning_rate)
model.compile(loss='categorical_crossentropy',
optimizer='adam',
metrics=['accuracy'])
# model.compile(loss='sparse_categorical_crossentropy',optimizer='adam',metrics=['accuracy'])
history = model.fit(X_train,
Y_train,
batch_size=self.nb_batch,
epochs=self.nb_epoch,
validation_data=None,
callbacks=None,
verbose=1)
print('save the architecture of a model')
json_string = model.to_json()
open(os.path.join(self.f_model, 'cnn_model.json'),
'w').write(json_string)
yaml_string = model.to_yaml()
open(os.path.join(self.f_model, 'cnn_model.yaml'),
'w').write(yaml_string)
print('save weights')
model.save_weights(os.path.join(self.f_model, 'cnn_weights.hdf5'))
KTF.set_session(old_session)
return
def main(args):
# device number
if args.gpu_num:
os.environ["CUDA_VISIBLE_DEVICES"] = args.gpu_num
print(os.environ["CUDA_VISIBLE_DEVICES"])
# set the necessary directories
train_dir = 'resources/train'
test_dir = 'resources/test'
train_list = []
train_list_file = "resources/train_list.txt"
val_list_file = "resources/val_list.txt"
with open(train_list_file, "r") as f:
for l in f:
train_list.append(l.replace("\n", ""))
val_list = []
with open(val_list_file, "r") as f:
for l in f:
val_list.append(l.replace("\n", ""))
with tf.Graph().as_default():
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
config.gpu_options.per_process_gpu_memory_fraction = 1.0
session = tf.Session(config=config)
KTF.set_session(session)
KTF.set_learning_phase(1)
# set callbacks
cp_cb = ModelCheckpoint(filepath='resources/checkpoints/checkpoint',
monitor='val_loss',
verbose=1,
save_best_only=True,
mode='auto',
period=2)
es_cb = EarlyStopping(monitor='val_loss',
patience=2,
verbose=1,
mode='auto')
tb_cb = TensorBoard(log_dir='resources/logs/', write_images=True)
csv_logger = CSVLogger('resources/logs/training.log')
# set generater
train_gen = data_gen_small('resources/train/', 'resources/train/',
train_list, args.batch_size,
[args.input_shape[0], args.input_shape[1]],
args.n_labels)
val_gen = data_gen_small('resources/val/', 'resources/val/', val_list,
args.batch_size,
[args.input_shape[0], args.input_shape[1]],
args.n_labels)
# set model
model = segunet(args.input_shape, args.n_labels, args.kernel,
args.pool_size, args.output_mode)
print(model.summary())
# compile model
model.compile(loss=args.loss,
optimizer=args.optimizer,
metrics=["accuracy"])
# fit with genarater
model.fit_generator(generator=train_gen,
steps_per_epoch=args.epoch_steps,
epochs=args.n_epochs,
validation_data=val_gen,
validation_steps=args.val_steps,
callbacks=[cp_cb, es_cb, tb_cb, csv_logger])
model.save_weights("resources/weights/weights_01.hdf5")
def dnn_classification(train_x, train_y, test_x, test_y, class_number, base_dir, from_frequency, to_frequency, frequency_list):
# conv1 = 30
nb_epoch = 10000
nb_batch = 32
learning_rate = 1e-2
try: ##convolutionを使う場合
conv1
train_x.resize(train_x.shape[0], train_x.shape[1], 1)
test_x.resize(test_x.shape[0], test_x.shape[1], 1)
except:
pass
dense1 = 60
dense2 = 30
dense3 = 14
dense4 = class_number
regularizers_l2_1 = 0
regularizers_l2_2 = 0
regularizers_l2_3 = 0
try:
model_structure = 'conv{0}relu_{1}relul2{2}_{3}relul2{4}_{5}relul2{6}_{7}softmax'.format(conv1, dense1,
regularizers_l2_1,
dense2,
regularizers_l2_2,
dense3,
regularizers_l2_3,
dense4)
except:
model_structure = '{0}relul2{1}_{2}relul2{3}_{4}relul2{5}_{6}softmax'.format(dense1, regularizers_l2_1, dense2,
regularizers_l2_2, dense3,
regularizers_l2_3, dense4)
f_log = base_dir + '/logs/fit' + 'freq' + str(
from_frequency) + 'to' + str(to_frequency) + 'num' + str(
len(frequency_list)) + '/' + model_structure + '_lr' + str(learning_rate) + '/Adam_epoch' + str(
nb_epoch) + '_batch' + str(nb_batch)
# print(f_log)
f_model = base_dir + '/model' + 'freq' + str(
from_frequency) + 'to' + str(to_frequency) + 'num' + str(
len(frequency_list)) + '/' + model_structure + '_lr' + str(learning_rate) + '/Adam_epoch' + str(
nb_epoch) + '_batch' + str(nb_batch)
os.makedirs(f_model, exist_ok=True)
# ニュートラルネットワークで使用するモデル作成
old_session = KTF.get_session()
with tf.Graph().as_default():
session = tf.Session('')
KTF.set_session(session)
KTF.set_learning_phase(1)
model = keras.models.Sequential()
try:
model.add(Conv1D(conv1, 4, padding='same', input_shape=(train_x.shape[1:]), activation='relu'))
model.add(Flatten())
model.add(Dense(dense1, activation='relu', kernel_regularizer=regularizers.l2(regularizers_l2_1)))
except:
model.add(Dense(dense1, activation='relu', kernel_regularizer=regularizers.l2(regularizers_l2_1),
input_shape=(train_x.shape[1:])))
# model.add(Dropout(0.25))
model.add(Dense(dense2, activation='relu', kernel_regularizer=regularizers.l2(regularizers_l2_2)))
# model.add(Dropout(0.25))
model.add(Dense(dense3, activation='relu', kernel_regularizer=regularizers.l2(regularizers_l2_3)))
model.add(Dense(dense4, activation='softmax'))
model.summary()
# optimizer には adam を指定
adam = keras.optimizers.Adam(lr=learning_rate)
model.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])
# model.compile(loss='sparse_categorical_crossentropy',optimizer='adam',metrics=['accuracy'])
train_y = np.array(train_y)
test_y = np.array(test_y)
# print(test_y)
# print(test_y.shape)
# print(type(test_y))
es_cb = keras.callbacks.EarlyStopping(monitor='val_loss', min_delta=0, patience=1000, verbose=0, mode='auto')
tb_cb = keras.callbacks.TensorBoard(log_dir=f_log, histogram_freq=1)
# cp_cb = keras.callbacks.ModelCheckpoint(filepath = os.path.join(f_model,'tag_model{epoch:02d}-loss{loss:.2f}-acc{acc:.2f}-vloss{val_loss:.2f}-vacc{val_acc:.2f}.hdf5'), monitor='val_loss', verbose=1, save_best_only=True, mode='auto')
# cbks = [es_cb, tb_cb, cp_cb]
cbks = [es_cb, tb_cb]
history = model.fit(train_x, train_y, batch_size=nb_batch, epochs=nb_epoch,
validation_data=(test_x, test_y), callbacks=cbks, verbose=1)
score = model.evaluate(test_x, test_y, verbose=0)
print('Test score:', score[0])
print('Test accuracy:', score[1])
predict = model.predict(test_x)
# print('predict:{}'.format(predict))
print('save the architecture of a model')
json_string = model.to_json()
open(os.path.join(f_model, 'tag_model.json'), 'w').write(json_string)
yaml_string = model.to_yaml()
open(os.path.join(f_model, 'tag_model.yaml'), 'w').write(yaml_string)
print('save weights')
model.save_weights(os.path.join(f_model, 'tag_weights.hdf5'))
KTF.set_session(old_session)
best_pred = []
probability = []
category = np.arange(1, class_number+1)
for (i, pre) in enumerate(predict):
y = pre.argmax() # preがそれぞれの予測確率で一番高いものを取ってきている。Y_testはone-hotベクトル
best_pred.append(category[y])
probability.append(pre[y])
return best_pred, probability
MultiprocessIterator(
valid,
batch_size=8,
#repeat=False,
shuffle=False,
n_processes=12,
n_prefetch=120,
shared_mem=1000 * 1000 *
5)) # type: Iterator[Tuple[np.ndarray, np.ndarray]]
old_session = tensorflow_backend.get_session()
with K.tf.Graph().as_default():
session = K.tf.Session("")
tensorflow_backend.set_session(session)
tensorflow_backend.set_learning_phase(1)
loss = "" # type: Union[Callable, str]
metrics = [] # type: List[Union[Callable, str]]
input_shape = (resize_shape[0], resize_shape[1], 3)
if args.dice_coef:
output_ch = 1
loss = dice_coef_loss
metrics = [dice_coef]
filename = "_weights.epoch{epoch:04d}-val_loss{val_loss:.2f}-val_dice_coef{val_dice_coef:.2f}.hdf5"
else:
output_ch = 2
loss = "categorical_crossentropy"
metrics = ['accuracy']
filename = "_weights.epoch{epoch:04d}-val_loss{val_loss:.2f}-val_acc{val_acc:.2f}.hdf5"
def main(args):
# device number
if args.gpu_num:
os.environ["CUDA_VISIBLE_DEVICES"] = args.gpu
# set the necessary list
train_list = pd.read_csv(args.train_list, header=None)
val_list = pd.read_csv(args.val_list, header=None)
# set the necessary directories
trainimg_dir = args.trainimg_dir
trainmsk_dir = args.trainmsk_dir
valimg_dir = args.valimg_dir
valmsk_dir = args.valmsk_dir
# get old session old_session = KTF.get_session()
with tf.Graph().as_default():
session = tf.Session('')
KTF.set_session(session)
KTF.set_learning_phase(1)
# set callbacks
cp_cb = ModelCheckpoint(filepath=args.log_dir,
monitor='val_loss',
verbose=1,
save_best_only=True,
mode='auto',
period=2)
es_cb = EarlyStopping(monitor='val_loss',
patience=2,
verbose=1,
mode='auto')
tb_cb = TensorBoard(log_dir=args.log_dir, write_images=True)
# set generater
train_gen = data_gen_small(trainimg_dir, trainmsk_dir, train_list,
args.batch_size,
[args.input_shape[0], args.input_shape[1]],
args.n_labels)
val_gen = data_gen_small(valimg_dir, valmsk_dir, val_list,
args.batch_size,
[args.input_shape[0], args.input_shape[1]],
args.n_labels)
# set model
pspnet = PSPNet50(input_shape=args.input_shape,
n_labels=args.n_labels,
output_mode=args.output_mode,
upsample_type=args.upsample_type)
print(pspnet.summary())
# compile model
pspnet.compile(loss=args.loss,
optimizer=args.optimizer,
metrics=["accuracy"])
# fit with genarater
pspnet.fit_generator(generator=train_gen,
steps_per_epoch=args.epoch_steps,
epochs=args.n_epochs,
validation_data=val_gen,
validation_steps=args.val_steps,
callbacks=[cp_cb, es_cb, tb_cb])
def main(args):
# set the necessary list
# train_list = pd.read_csv(args.train_list,header=None)
# val_list = pd.read_csv(args.val_list,header=None)
# set the necessary directories
trainimg_dir = args.trainimg_dir
trainmsk_dir = args.trainmsk_dir
valimg_dir = args.valimg_dir
valmsk_dir = args.valmsk_dir
# get old session
# old_session = KTF.get_session()
with tf.Graph().as_default():
session = tf.Session('')
KTF.set_session(session)
KTF.set_learning_phase(1)
# set callbacks
fpath = './pretrained_mask/' + args.name + '{epoch:02d}.hdf5'
cp_cb = ModelCheckpoint(filepath=fpath,
monitor='val_loss',
verbose=1,
mode='auto',
period=1)
tb_cb = TensorBoard(log_dir="./pretrained_mask",
write_graph=True,
write_images=True)
seq = iaa.Sequential(
[
iaa.Crop(
px=(0, 16)
), # crop images from each side by 0 to 16px (randomly chosen)
iaa.Fliplr(0.5), # horizontally flip 50% of the images
sometimes(
iaa.Affine(
scale={
"x": (0.8, 1.2),
"y": (0.8, 1.2)
},
# scale images to 80-120% of their size, individually per axis
translate_percent={
"x": (-0.2, 0.2),
"y": (-0.2, 0.2)
}, # translate by -20 to +20 percent (per axis)
rotate=(-10, 10), # rotate by -45 to +45 degrees
)),
],
random_order=True)
if args.dataset == 'coco':
train_gen = Datahandler_COCO(trainimg_dir,
trainmsk_dir).make_batches(
batchsize=args.batch_size,
inputshape=args.input_shape,
augmentation=seq)
val_gen = Datahandler_COCO(valimg_dir, valmsk_dir).make_batches(
batchsize=args.batch_size,
inputshape=args.input_shape,
augmentation=None)
elif args.dataset == 'pascal_khamba':
train_gen = Pascal_Generator(trainimg_dir,
trainmsk_dir).make_batches(
batchsize=args.batch_size,
inputshape=args.input_shape,
augmentation=seq)
val_gen = Pascal_Generator(valimg_dir, valmsk_dir).make_batches(
batchsize=args.batch_size,
inputshape=args.input_shape,
augmentation=None)
else:
train_gen = Default_Generator(trainimg_dir,
trainmsk_dir).make_batches(
batchsize=args.batch_size,
inputshape=args.input_shape,
augmentation=seq)
val_gen = Default_Generator(valimg_dir, valmsk_dir).make_batches(
batchsize=args.batch_size,
inputshape=args.input_shape,
augmentation=None)
# set model
pspnet = PSPNet50(input_shape=args.input_shape,
n_labels=args.n_labels,
output_mode=args.output_mode,
upsample_type=args.upsample_type)
print(pspnet.summary())
if args.load is not None:
print("loadinf weights")
pspnet.load_weights(args.load)
# compile model
pspnet.compile(loss=args.loss,
optimizer=args.optimizer,
metrics=["accuracy"])
# fit with genarater
pspnet.fit_generator(generator=train_gen,
steps_per_epoch=args.steps,
epochs=args.epochs,
validation_data=val_gen,
validation_steps=args.val_steps,
callbacks=[cp_cb, tb_cb],
verbose=True)
# save model
with open("./pretrained_mask/" + args.name + ".json", "w") as json_file:
json_file.write(json.dumps(json.loads(pspnet.to_json()), indent=2))
print("save json model done...")
def run_cnn(X_train,
X_val,
X_test,
network_func,
n_classes=30,
img_shape=(128, 128, 3),
optimizer='adam',
loss='categorical_crossentropy',
metrics=['accuracy'],
callbacks=None,
batch_size=32,
steps_per_epoch=64,
num_epochs=300,
validation_steps=50,
path_save_weights=None,
path_plot_model=None,
dropout=False):
train_iter = batch_generator.Batch_generator(X_train,
image_shape=img_shape,
num_classes=n_classes,
batch_size=batch_size,
segment=False)
val_iter = batch_generator.Batch_generator(X_val,
image_shape=img_shape,
num_classes=n_classes,
batch_size=batch_size,
segment=False)
old_session = KTF.get_session()
# with tf.Graph().as_default():
session = tf.Session()
KTF.set_session(session)
# if using dropout (a situation when networks would be different between training and testing phase)
if dropout:
KTF.set_learning_phase(1)
inputs = layers.Input(img_shape)
output_layer = network_func(inputs, n_classes)
model = models.Model(inputs, output_layer)
model.summary()
model.compile(optimizer=optimizer, loss=loss, metrics=metrics)
history = model.fit_generator(train_iter,
steps_per_epoch=steps_per_epoch,
epochs=num_epochs,
verbose=1,
validation_data=val_iter,
validation_steps=validation_steps,
callbacks=callbacks,
workers=2)
# save model
# model.save('./test_log/models/fcn_model.hdf5')
if path_save_weights != None:
try:
model.save_weights(path_save_weights)
except:
print("Could not save weigths data.")
if path_plot_model != None:
try:
plot_model(model, to_file=path_plot_model)
except:
print("Could not plot model properly.")
# test the prediction
print("\n")
print("Testing model...")
score = model_helper.predict_from_path(model, X_test, img_shape, n_classes)
print("\n")
print("=====" * 13)
print("\n")
print("Test score : {:.6f}".format(score[0]))
print("Test accuracy : {:.6f}\n".format(score[1]))
KTF.set_session(old_session)
def main():
base_dir = '/media/localadmin/Test/11Nils/kitti/dataset/sequences/Data/'
label_dir = 'labels/'
train_img_dir = 'images/'
eval_base_dir = '/media/localadmin/Test/11Nils/kitti/dataset/sequences/08/'
eval_lbl_dir = 'labels/'
eval_img_dir = 'image_2/'
inf_dir = 'pooling_test/'
test_no_indices = True
if test_no_indices:
inf_dir_tested = 'MaxPooling2D/'
log = 'log/'
gpu_num = '0, 1, 2, 3'
else:
inf_dir_tested = 'MaxPooling2DWithIndices/'
log = 'log_indices/'
gpu_num = '2, 3'
segnet = True
if segnet:
inf_dir_tested = 'SegNet/'
log = 'log_seg/'
gpu_num = '2, 3'
label_list = sorted(os.listdir(base_dir + label_dir))
shuffle(label_list)
train = label_list[:int(len(label_list) * 0.8)]
val = label_list[int(len(label_list) * 0.8):]
eval = os.listdir(eval_base_dir + eval_lbl_dir)
print(len(train), len(val), len(eval))
with tf.Graph().as_default():
os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
os.environ["CUDA_VISIBLE_DEVICES"] = gpu_num
session = tf.Session('')
KTF.set_session(session)
KTF.set_learning_phase(1)
trainer = Trainer(_train_list=train,
_val_list=val,
_inf_list=eval,
_gpu_num=gpu_num,
_no_inidices=test_no_indices,
_segnet=segnet)
trainer.base_dir = base_dir
trainer.label_dir = label_dir
trainer.img_dir = train_img_dir
trainer.log_dir = eval_base_dir + inf_dir + log
trainer.inf_dir = inf_dir + inf_dir_tested
trainer.batch_size = 16
trainer.epoch_steps = 750
trainer.val_steps = 200
trainer.n_epochs = 30
trainer.dag_it = 0
trainer.update_callback()
# trains model for defined number of epochs with the actual dataset
print('Loading labels from %s' %
(trainer.base_dir + trainer.label_dir))
print('Loading imgs from %s' % (trainer.base_dir + trainer.img_dir))
trainer.train()
print('\nTraining done!\nStarting Prediction\n')
# safes inferences of images that are unseen by the net
trainer.base_dir = eval_base_dir
trainer.img_dir = eval_img_dir
print('Loading labels from %s' %
(trainer.base_dir + trainer.label_dir))
print('Loading imgs from %s' % (trainer.base_dir + trainer.img_dir))
trainer.predict()
session.close()
ax.text(x - 1,
y - 1,
part_place_name,
color='black',
fontsize=15)
place_name_list.append(part_place_name)
return ax, place_name_list
#実際に確認したいターゲットの画像を設定する。
img = cv2.imread("hrei-sign105.png")
input_shape = (50, 100, 1)
with tf.Graph().as_default():
session = tf.Session('')
KTF.set_session(session)
KTF.set_learning_phase(1)
def schedule(epoch, decay=0.9):
return base_lr * decay**(epoch)
base_lr = 3e-4
optim = Adam(lr=base_lr)
model = spp_model(input_shape, NUM_CLASSES, optim)
model.load_weights('./param/learning_place_name.hdf5', by_name=True)
result_label_list = predict_img(img, model, place_list, input_shape)
print(len(result_label_list))
def run_segmentation(X_train,
y_train,
X_val,
y_val,
X_test,
y_test,
network_func,
n_classes=30,
img_shape=(128, 128, 3),
optimizer='adam',
loss='categorical_crossentropy',
metrics=['accuracy'],
callbacks=None,
batch_size=32,
steps_per_epoch=500,
num_epochs=30,
validation_steps=65,
path_save_weights=None,
path_save_pred_images=None,
num_save_pred_image=None,
dropout=False):
# import tensorflow as tf
# from keras import backend as K
config = tf.ConfigProto()
config.gpu_options.per_process_gpu_memory_fraction = 0.90
train_iter = batch_generator.Batch_generator(X_train,
target=y_train,
image_shape=img_shape,
num_classes=n_classes,
batch_size=batch_size,
segment=True)
val_iter = batch_generator.Batch_generator(X_val,
target=y_val,
image_shape=img_shape,
num_classes=n_classes,
batch_size=batch_size,
segment=True)
old_session = KTF.get_session()
# with tf.Graph().as_default():
session = tf.Session(config=config)
KTF.set_session(session)
# if using dropout (a situation when networks would be different between training and testing phase)
if dropout:
KTF.set_learning_phase(1)
inputs = layers.Input(img_shape)
output_layer = network_func(inputs, n_classes)
model = models.Model(inputs, output_layer)
model.summary()
model.compile(optimizer=optimizer, loss=loss, metrics=metrics)
model.fit_generator(train_iter,
steps_per_epoch=steps_per_epoch,
epochs=num_epochs,
verbose=1,
validation_data=val_iter,
validation_steps=validation_steps,
callbacks=callbacks,
workers=2)
# save model
# model.save('./test_log/models/fcn_model.hdf5')
if path_save_weights != None:
try:
model.save_weights(path_save_weights)
except:
print("Path to save weigths data is not valid.")
# predicting and calculate IoU
print("\n")
print("=====" * 15)
print("\n")
print("Testing model...")
start = time()
# extract data
test = np.vstack([
np.expand_dims(misc.imresize(misc.imread(t), img_shape), axis=0)
for t in X_test
])
pred = model_helper.prediction(model, test) # predicted data
target = pre.pixelwise_class_array(y_test) # ground truths
iou = model_helper.ave_iou_score(pred, target)
end = time()
print("\n")
print("IoU score : {:.6f}".format(iou))
print("Calcuration time : {:.6f} sec.".format(end - start))
# Save predicted image
if path_save_pred_images != None:
print("\n")
print("=====" * 15)
print("\n")
print("Saving predict image...")
path_save_pred_images = os.path.join(path_save_pred_images,
'predictions')
# create directory
pre.makedirs_if_none(path_save_pred_images)
# reduce save data if need
if num_save_pred_image != None:
pred = pred[:num_save_pred_image]
X_test = X_test[:num_save_pred_image]
# convert from class array to image(rgb) array
pred = pre.pixelwise_array_to_img(pred)
# Save data
for img, file_path in zip(pred, X_test):
misc.imsave(
os.path.join(path_save_pred_images,
os.path.basename(file_path)), img)
print("Done.")
# close current session
KTF.set_session(old_session)
