def train(batch, epochs, size):
"""Train the model.
Arguments:
batch: Integer, The number of train samples per batch.
epochs: Integer, The number of train iterations.
size: Integer, image size.
"""
if not os.path.exists('model'):
os.makedirs('model')
model = MSCNN((size, size, 3))
opt = SGD(lr=1e-5, momentum=0.9, decay=0.0005)
model.compile(optimizer=opt, loss='mse')
lr = ReduceLROnPlateau(monitor='loss', min_lr=1e-7)
indices = list(range(1500))
train_ids, test_ids = train_test_split(indices, test_size=0.25)
hist = model.fit_generator(
generator(train_ids, batch, size),
validation_data=generator(test_ids, batch, size),
steps_per_epoch=len(train_ids) // batch,
validation_steps=len(test_ids) // batch,
epochs=epochs,
callbacks=[lr])
model.save_weights('model\\final_weights.h5')
df = pd.DataFrame.from_dict(hist.history)
df.to_csv('model\\history.csv', index=False, encoding='utf-8')
def test(self, sess, sents, lens, labels, epoch, step, raw_data):
batches = generator(zip(sents, lens, labels), self.batch_size)
res, seq_lens = [], []
for x, seq_len, y in batches:
logits, transition_params = sess.run([self.logits, self.transition_params],
feed_dict={self.input_x:x, self.input_y:y, self.seq_len:seq_len})
seq_lens.extend(seq_len)
for logit, seq_l in zip(logits, seq_len):
viterbi_seq, _ = viterbi_decode(logit[:seq_l], transition_params)
res.append(viterbi_seq)
self.evaluate(res, seq_lens, zip(sents, labels), epoch, step, raw_data)
def train(self, sess, train_sent, train_len, train_label, test_sent, test_len, test_label, raw_data):
epoch = 20
for e in range(epoch):
batches = generator(zip(train_sent, train_len, train_label), self.batch_size)
for step, (x, seq_len, y) in enumerate(batches):
_, loss = sess.run([self.train_op, self.loss],
feed_dict={self.input_x:x,self.input_y:y, self.seq_len:seq_len})
print('epoch '+str(e)+', step '+str(step)+', loss '+str(loss))
if step % 100 == 0:
self.test(sess, test_sent, test_len, test_label, e, step, raw_data)
def val():
datas = load_pickles()
X_test, y_test = datas['X_test'], datas['y_test']
test_gen = generator(X_test,
y_test,
batch_size=30,
aug=False,
shuffle=False)
m = _get_m()
res = m.predict_generator(test_gen, steps=(X_test.shape[0] // 30))
res = np.argmax(res, axis=1)
print(np.mean(res == y_test))
def main(argv):
print('Finding data ...'),
with Timer():
image_filenames = glob(os.path.join(FLAGS.data_dir, 'disc*/OAS1_*_MR1/PROCESSED/MPRAGE/T88_111/OAS1_*_MR1_mpr_n4_anon_111_t88_gfc.hdr'))
label_filenames = glob(os.path.join(FLAGS.data_dir, 'disc*/OAS1_*_MR1/FSL_SEG/OAS1_*_MR1_mpr_n4_anon_111_t88_masked_gfc_fseg.hdr'))
assert(len(image_filenames) == len(label_filenames))
print('Found %i images.' % len(image_filenames))
print('Loading images ...'),
with Timer():
images = [ReadImage(image_filename) for image_filename in image_filenames]
labels = [ReadImage(label_filename) for label_filename in label_filenames]
images_train, images_test, labels_train, labels_test = train_test_split(images, labels, train_size=0.66)
tensor_board = TensorBoard(log_dir='./TensorBoard')
early_stopping = EarlyStopping(monitor='acc', patience=2, verbose=1)
model = create_inception_resnet_v2(nb_classes=nb_classes)
model.compile(optimizer=RMSprop(lr=0.045, rho=0.94, epsilon=1., decay=0.9), loss='categorical_crossentropy', metrics=['acc'])
model.fit_generator(generator(images_train, labels_train, input_shape, nb_classes, FLAGS.patch_size, FLAGS.batch_size),
samples_per_epoch=FLAGS.samples_per_epoch, nb_epoch=FLAGS.nb_epochs, callbacks=[tensor_board, early_stopping],
verbose=1)
import sklearn
from sklearn import preprocessing as skproc
from data import (
load_pickles,
generator,
pregen,
)
BATCH = 128
datas = load_pickles()
X_train, y_train = datas['X_train'], datas['y_train']
X_valid, y_valid = datas['X_valid'], datas['y_valid']
train_gen = generator(X_train, y_train, batch_size=BATCH)
valid_gen = generator(X_valid, y_valid, batch_size=BATCH)
n_classes = datas['n_classes']
image_shape = X_train.shape[1:4]
import os
import keras
from model import get_model
m = get_model(image_shape, n_classes)
epochs = 360
monitor = 'val_acc'
callbacks = [
def main():
"""main function
Main function... (what do you expect me to say...)
Args:
- none
Returns:
- none
"""
# Main function for evaluate
parser = argparse.ArgumentParser(
description="A testing framework for semantic segmentation.")
parser.add_argument(
"--net",
required=True,
default="unet",
type=str,
help=
"(str) The type of net work which is either unet, deeplab or custom.")
parser.add_argument("--epochs", required=False, default=500, type=int)
parser.add_argument("--batch_size", required=False, default=16, type=int)
parser.add_argument("--gpu_id",
required=False,
default="0",
type=str,
help="(str) The id of the gpu used when training.")
parser.add_argument("--img_size",
required=False,
default=192,
type=int,
help="(int) The size of input image")
parser.add_argument(
"--load_weights",
required=False,
default=False,
type=bool,
help="(bool) Use old weights or not (named net_imgSize.h5)")
# Parse argument
args = parser.parse_args()
net_type = args.net
epochs = args.epochs
batch_size = args.batch_size
gpu_number = args.gpu_id
img_size = args.img_size
import os
os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
os.environ["CUDA_VISIBLE_DEVICES"] = gpu_number
# Argument check
if not (net_type in {"unet", "deeplab", "custom"}):
raise ValueError("netType should be either unet, deeplab and custom.")
# Get config
Config = cfg.Config()
# COCO instance
print("Reading COCO ground truth...")
cocoGt = COCO(Config.COCO_training_ann_path)
cocoValGt = COCO(Config.COCO_validation_ann_path)
print("Finished")
# Get all classes
classes = len(cocoGt.getCatIds())
id_to_index = dict()
# There is a wired class of 0 in the feature map of type zero
index_to_id = dict()
# Because the id of COCO dataset starts from 92, we should project those id to index so that keras
# utils can convert the segmentation map into one hot categorical encoding.
for index, id in enumerate(cocoGt.getCatIds()):
id_to_index[id] = index
index_to_id[index] = id
if net_type == "unet":
model = basic_model.unet(input_size=(img_size, img_size, 3),
classes=len(id_to_index))
elif net_type == "deeplab":
deeplab_model = basic_model.Deeplabv3(input_shape=(img_size, img_size,
3),
classes=len(id_to_index),
backbone="xception")
output = KL.Activation("softmax")(deeplab_model.output)
model = KM.Model(deeplab_model.input, output)
elif net_type == "custom":
model = model.custom_model(input_shape=(img_size, img_size, 3),
classes=len(id_to_index))
file_list = glob(Config.COCO_training_path + '*')
val_list = glob(Config.COCO_validation_path + '*')
if args.load_weights:
try:
model.load_weights(net_type + "_" + str(img_size) + ".h5")
print("weights loaded!")
except:
print("weights not found!")
checkpointer = KC.ModelCheckpoint(filepath=net_type + "_" + str(img_size) +
".h5",
verbose=1,
save_best_only=True)
model.compile(optimizer=KO.Adam(),
loss="categorical_crossentropy",
metrics=["accuracy"])
model.fit_generator(data.generator(batch_size, file_list,
(img_size, img_size), cocoGt,
id_to_index, True),
validation_data=data.generator(batch_size, val_list,
(img_size, img_size),
cocoValGt, id_to_index,
False),
validation_steps=10,
steps_per_epoch=100,
epochs=epochs,
use_multiprocessing=True,
workers=8,
callbacks=[checkpointer])
print("Prediction start...")
vfunc = np.vectorize(lambda index: index_to_id[index])
anns = []
# Convert into COCO annotation
for i in trange(len(val_list)):
image = val_list[i]
image_id = int(image.replace(".jpg", '')[-12:])
cropping_image, padding_dims, original_size = utils.padding_and_cropping(
image, (img_size, img_size))
cropping_image = preprocess_input(cropping_image, mode="torch")
result = model.predict(cropping_image)
result = np.argmax(result, axis=3)
seg_result = utils.reverse_padding_and_cropping(
result, padding_dims, original_size)
seg_result = vfunc(seg_result)
COCO_ann = cocostuffhelper.segmentationToCocoResult(seg_result,
imgId=image_id)
for ann in COCO_ann:
ann["segmentation"]["counts"] = ann["segmentation"][
"counts"].decode("ascii") # json can't dump byte string
anns += COCO_ann
with open("result.json", "w") as file:
json.dump(anns, file)
# Read result file
# Test for fake result
#resFile = Config.fake_result
# Evaluate result
resFile = "result.json"
cocoDt = cocoValGt.loadRes(resFile)
cocoEval = COCOStuffeval(cocoValGt, cocoDt)
cocoEval.evaluate()
cocoEval.summarize()
import data
import nvidia
from keras.optimizers import Adam
if __name__ == '__main__':
db = data.database()
train_db, validation_db = data.validation_split(db)
model = nvidia.model_x()
model.summary()
model.compile(loss='mse', optimizer=Adam(), metrics=['mse', 'accuracy'])
# uncomment if you want to update trained network
# model.load_weights('model.h5')
model.fit_generator(data.generator(train_db, augment=True),
validation_data=data.load_data(validation_db),
steps_per_epoch=50,
epochs=10)
model.save('model.h5')
from keras.models import Sequential
from keras.layers.core import Flatten, Dense, Lambda, Activation, Dropout
from keras.layers.convolutional import Convolution2D, Cropping2D
from keras.layers.pooling import MaxPooling2D
import data as DT
batch_size = 32
train_generator = DT.generator(DT.train_samples, batch_size)
validation_generator = DT.generator(DT.validation_samples, batch_size)
nb_train = len(DT.train_samples) * 2
nb_valid = len(DT.validation_samples) * 2
# try to visualize the cropping images/reverse image
model = Sequential()
model.add(Cropping2D(cropping=((50, 20), (0, 0)), input_shape=(160, 320, 3)))
model.add(Lambda(lambda x: x / 127.5 - 1.))
model.add(Convolution2D(6, 5, 5))
# model.add(Convolution2D(6, 5, 5, input_shape=(160, 320, 3)))
model.add(Activation('relu'))
model.add(MaxPooling2D(pool_size=(2, 2))) # default (2, 2)
model.add(Convolution2D(
6,
5,
5,
))
model.add(Activation('relu'))
model.add(MaxPooling2D(pool_size=(2, 2))) # default (2, 2)
model.add(Convolution2D(
6,
def train():
#
# SaveModel is a CallBack class that we can use to save the model for each epoch
# This allows us to easily test each epoch on the simulator. The simulator seems like
# a better validation than just the validation data set
class SaveModel(Callback):
def on_epoch_end(self, epoch, logs={}):
epoch += 1
if (epoch > 0):
#with open ('model-' + str(epoch) + '.json', 'w') as file:
# file.write (model.to_json ())
# file.close ()
#model.save_weights ('model-' + str(epoch) + '.h5')
model.save('model-' + str(epoch) + '.h5')
#
# load the model
#
model = get_model()
# Keras has a nice tool to create an image of our network
from keras.utils.visualize_util import plot
plot(model, to_file='car_model.png', show_shapes=True)
print("Loaded model")
# load the data
xs, ys = data.loadTraining()
# split the dataset into training and validation 80% / 20%
train_xs = xs[:int(len(xs) * 0.8)]
train_ys = ys[:int(len(xs) * 0.8)]
val_xs = xs[-int(len(xs) * 0.2):]
val_ys = ys[-int(len(xs) * 0.2):]
# load the validation dataset, it is better not generate an image each time - process them once
# Use the validation process function, it doesn't augment the image, just resizes it
X, y = data.getValidationDataset(val_xs, val_ys,
data.processImageValidation)
print(model.summary())
print("Loaded validation datasetset")
print("Total of", len(train_ys))
print("Training..")
checkpoint_path = "weights.{epoch:02d}-{val_loss:.2f}.hdf5"
checkpoint = ModelCheckpoint(checkpoint_path,
verbose=1,
save_best_only=False,
save_weights_only=False,
mode='auto')
# I tried using the earlystopping callback, but now I run it for a fixed number of epochs and test to see which is best
earlystopping = EarlyStopping(monitor='val_loss',
min_delta=0,
patience=2,
verbose=1,
mode='auto')
res = model.fit_generator(data.generator(train_xs, train_ys, 256),
validation_data=(X, y),
samples_per_epoch=100 * 256,
nb_epoch=epochs,
verbose=1,
callbacks=[SaveModel()])
# pickle and dump the history so we can graph it in a notebook
history = res.history
with open('history.p', 'wb') as f:
pickle.dump(history, f)
#predictions = tf.squeeze(predictions, axis=[1,2])
correct_prediction = tf.equal(
tf.argmax(predictions, axis=1),
tf.argmax(targets, axis=1)) #shape of correct_prediction is [N]
accuracy = tf.reduce_mean(tf.cast(correct_prediction, 'float'))
loss = tf.reduce_mean(
tf.nn.softmax_cross_entropy_with_logits(labels=targets,
logits=predictions))
tf.summary.scalar('loss', loss)
optimizer = tf.train.GradientDescentOptimizer(learning_rate=learning_rate)
global_steps = tf.Variable(0, trainable=False)
train_step = optimizer.minimize(loss, global_step=global_steps)
filepath = 'train.txt'
gen = generator(filepath, height, width, num_classes, batch_size)
saver = tf.train.Saver()
init = tf.global_variables_initializer()
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
with tf.Session(config=config) as sess:
sess.run(init)
for step in range(total_steps):
images_, labels_ = next(gen)
sess.run(train_step, feed_dict={inputs: images_, targets: labels_})
loss_, accuracy_ = sess.run([loss, accuracy],
feed_dict={
inputs: images_,
targets: labels_
})
from keras.layers import Dense, Flatten, \
Lambda, Conv2D, MaxPooling2D
from keras.layers.noise import AlphaDropout
from sklearn.model_selection import train_test_split
from data import SHAPE, generator
samples = pd.read_csv('driving_log.csv').as_matrix()
train_samples, valid_samples = train_test_split(samples, test_size=0.2)
# Hyperparameters
EPOCHS = 6
BATCH_SIZE = 32
LEARNING_RATE = 1.0e-3
# Data
train_generator = generator(train_samples, batch_size=BATCH_SIZE)
valid_generator = generator(valid_samples, batch_size=BATCH_SIZE)
# Define the model
#model = load_model("./model.h5")
model = Sequential()
model.add(Lambda(lambda x: x / 127.5 - 1., input_shape=SHAPE))
model.add(
Conv2D(2, (3, 3), activation='selu', kernel_initializer='lecun_normal'))
model.add(MaxPooling2D((4, 4)))
model.add(AlphaDropout(0.25))
model.add(Flatten())
model.add(Dense(1))
model.compile(loss='mse', optimizer=optimizers.Adam(lr=LEARNING_RATE))
# Train the model
import data as d
import model as m
from sklearn.model_selection import train_test_split
path_list_train=d.getImagesPath('train')
path_list_test=d.getImagesPath('test')
X_train,Y_train=d.PreprocessData(path_list_train)
X_test,sizes_test=d.PreprocessData(path_list_test,False)
d.savePreparedData(X_test,X_train,Y_train,sizes_test)
xtr, xval, ytr, yval = train_test_split(X_train, Y_train, test_size=0.1, random_state=7)
#X_gen,Y_gen=d.genImagesAndMasks(X_train,Y_train)
#X_train=np.concatenate(X_train,X_gen)
#Y_train=np.concatenate(Y_train,Y_gen)
train_generator, val_generator = d.generator(xtr, xval, ytr, yval, 16)
model = get_unet(256,256,3)
model.fit_generator(train_generator, steps_per_epoch=len(xtr)/6, epochs=250,
validation_data=val_generator, validation_steps=len(xval)/16)
preds_test = model.predict(X_test, verbose=1)
preds_test_t = (preds_test > 0.5).astype(np.uint8)
preds_test_upsampled =d.resizeTest(path_list_test,preds_test,sizes_test)
new_test_ids = []
rles = []
for n, path in enumerate(path_list_test):
rle = list(d.prob_to_rles(preds_test_upsampled[n]))
rles.extend(rle)
new_test_ids.extend([os.path.splitext(os.path.basename(os.path.normpath(str(path))))[0]] * len(rle))
from keras.optimizers import Adam
from data import generator, train_samples, validation_samples, test_samples
import math
def samples_count(total_count, batch_size):
"""
Compute number of ssample per epoch
"""
return math.ceil(total_count/batch_size) * batch_size
learning_rate = 0.0001
batch_size = 32
epochs = 3
# Create generators
train_generator = generator(train_samples, batch_size=batch_size)
validation_generator = generator(validation_samples, batch_size=batch_size, training=False)
test_generetaor = generator(test_samples, batch_size=test_samples.shape[0], training=False)
def build_model(learning_rate=0.0001):
model = Sequential()
model.add(Lambda(lambda x: (x / 255.0) - 0.5, input_shape=(64,64,3)))
model.add(Convolution2D(16, 3,3, subsample=(1, 1), activation="relu"))
model.add(MaxPooling2D(pool_size=(2,2)))
model.add(Convolution2D(32, 3,3, subsample=(1, 1), activation="relu"))
model.add(MaxPooling2D(pool_size=(2,2)))
model.add(Convolution2D(64, 3,3, subsample=(1, 1), activation="relu"))
model.add(MaxPooling2D(pool_size=(2,2)))
model.add(Flatten())
model.add(Dense(512,activation="relu"))
model.add(Dropout(0.5))
