def visualize(args):
testimg_dir = args.testimg_dir
testmsk_dir = args.testmsk_dir
test_list = range(args.test_size)
file1 = open(args.save_dir + "recent.txt", "r")
args.weight = file1.read()
file1.close()
try:
model = segnet(args.input_shape, args.n_labels, args.kernel,
args.pool_size, args.output_mode, int(args.weight[-1]))
model.load_weights(args.save_dir + args.weight + ".hdf5")
print(args.save_dir + args.weight + ".hdf5")
except:
print('The model name entered does not exist')
else:
for i in test_list:
img = []
# If grayscale image is used as input
# original_img = cv2.imread(testimg_dir+str(i)+".jpg", 0)
# resized_img = cv2.resize(original_img, (args.input_shape[0], args.input_shape[1]))
# imag = np.stack([resized_img, resized_img, resized_img], axis=2)
# array_img = img_to_array(resized_img) / 255
# img.append(imag)
original_img = cv2.imread(testimg_dir + str(i) + ".jpg")
resized_img = cv2.resize(
original_img, (args.input_shape[0], args.input_shape[1]))
array_img = img_to_array(resized_img) / 255
img.append(resized_img)
array_img = np.expand_dims(array_img, axis=0)
output = model.predict(array_img)
mask = cv2.imread(testmsk_dir + str(i) + ".jpg")
resized_mask = cv2.resize(
mask * 255, (args.input_shape[0], args.input_shape[1]),
interpolation=cv2.INTER_NEAREST)
img.append(resized_mask)
resized_image = np.reshape(
output[0] * 255,
(args.input_shape[0], args.input_shape[1], args.n_labels))
img.append(resized_image[:, :, 1])
fig = plt.figure()
ax = []
for j in range(3):
ax.append(fig.add_subplot(1, 3, j + 1))
if j == 0:
plt.imshow(img[j])
else:
plt.imshow(img[j], cmap='gray', vmin=0, vmax=255)
ax[0].title.set_text("Original Image")
ax[1].title.set_text("Ground Truth")
ax[2].title.set_text("Predicted Mask")
fig.savefig(args.save_dir + "images/" + str(i) + ".png",
dpi=fig.dpi)
plt.close('all')
def main(args):
# set the necessary list
test_file = open(args.test_list, "r")
test_list = test_file.readlines()
model = segnet(args.input_shape, args.n_labels, args.kernel,
args.pool_size, args.output_mode)
print(model.summary())
model.compile(loss=args.loss,
optimizer=Adam(lr=0.0001),
metrics=["accuracy"])
print("Load Model: " + args.resume)
model.load_weights(args.resume)
list_images = image_list(test_list)
size = (args.input_shape[0], args.input_shape[1])
for image in list_images:
print(image)
original_img = cv2.imread(image)[:, :, ::-1]
converted_img = convert_image(original_img, size)
result = model.predict(converted_img)[0]
colored_image = color_image(result, size, args.n_labels, color_table)
last_slash = image.rfind('/')
name = image[last_slash + 1:]
cv2.imwrite(args.save_dir + "/" + name, colored_image)
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
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)
model = segnet(args.input_shape, args.n_labels, args.kernel,
args.pool_size, args.output_mode)
print(model.summary())
model.compile(loss=args.loss,
optimizer=args.optimizer,
metrics=["accuracy"])
model.fit_generator(train_gen,
steps_per_epoch=args.epoch_steps,
epochs=args.n_epochs,
validation_data=val_gen,
validation_steps=args.val_steps)
model.save_weights(args.save_dir + str(args.n_epochs) + ".hdf5")
print("sava weight done..")
def train(args):
train_list = range(args.train_size)
val_list = range(args.val_size)
# set the necessary directories
trainimg_dir = args.trainimg_dir
trainmsk_dir = args.trainmsk_dir
valimg_dir = args.valimg_dir
valmsk_dir = args.valmsk_dir
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)
model = segnet(args.input_shape, args.n_labels, args.kernel,
args.pool_size, args.output_mode, args.model_number)
print(model.summary())
with open(args.save_dir + 'model' + str(args.model_number) + '.txt',
'w') as fh:
model.summary(print_fn=lambda x: fh.write(x + '\n'))
model.compile(loss=args.loss,
optimizer=args.optimizer,
metrics=["accuracy"])
print("Model Compiled")
tensorboard = TensorBoard(log_dir='logs/{}'.format(time()))
model.fit_generator(train_gen,
steps_per_epoch=args.epoch_steps,
epochs=args.n_epochs,
validation_data=val_gen,
validation_steps=args.val_steps,
callbacks=[tensorboard])
model.save(args.save_dir + args.new_model_name + ".hdf5")
print("Model Saved")
K.set_learning_phase(0)
session = K.get_session()
init = tf.global_variables_initializer()
session.run(init)
frozen_graph = freeze_session(
session, output_names=[out.op.name for out in model.outputs])
print([out.op.name for out in model.outputs])
tf.train.write_graph(frozen_graph,
args.save_dir,
args.new_model_name + ".pb",
as_text=False)
session.close()
file1 = open(args.save_dir + "recent.txt", "w+")
file1.write(args.new_model_name)
file1.close()
def main(args):
# set the necessary list
train_file = open(args.train_list, "r")
train_list = train_file.readlines()
random.shuffle(train_list)
val_file = open(args.val_list, "r")
val_list = val_file.readlines()
test_file = open(args.test_list, "r")
test_list = test_file.readlines()
train_gen = data_gen_small(
train_list,
args.batch_size,
[args.input_shape[0], args.input_shape[1]],
args.n_labels,
)
val_gen = data_gen_small(
val_list,
args.batch_size,
[args.input_shape[0], args.input_shape[1]],
args.n_labels,
)
model = segnet(args.input_shape, args.n_labels, args.kernel,
args.pool_size, args.output_mode)
print(model.summary())
model.compile(loss=args.loss,
optimizer=Adam(lr=0.001),
metrics=["accuracy"])
if args.resume:
print("Load Model: " + args.resume)
model.load_weights(args.resume)
model.fit_generator(
train_gen,
steps_per_epoch=args.epoch_steps,
epochs=args.n_epochs,
validation_data=val_gen,
validation_steps=args.val_steps,
)
model.save_weights(args.save_dir + str(args.n_epochs) + ".hdf5")
print("save weight done..")
test_gen = test_data_generator(
test_list,
#args.batch_size,
[args.input_shape[0], args.input_shape[1]],
args.n_labels,
)
model.evaluate(test_gen)
def main(args):
# set the necessary list i.e. list of names of images to be read
test_list = pd.read_csv(args.test_list, header=None)
# set the necessary directories
testimg_dir = args.testimg_dir
testmsk_dir = args.testmsk_dir
# Generate batch data for SGD
# NOTE: This helps control the batch size for each test set
test_gen = data_gen_small(testimg_dir, testmsk_dir, test_list,
args.batch_size,
[args.input_shape[0], args.input_shape[1]],
args.n_labels)
# Create a model
if args.model == 'unet':
model = unet(\
args.input_shape, args.n_labels, args.kernel, \
args.pool_size, args.output_mode, args.gpus)
elif args.model == 'segnet':
model = segnet(\
args.input_shape, args.n_labels, args.kernel, \
args.pool_size, args.output_mode, args.gpus)
elif args.model == 'segunet':
model = segunet(\
args.input_shape, args.n_labels, args.kernel, \
args.pool_size, args.output_mode, args.gpus)
# TODO: Configure the model for training
optimizer = Adadelta(
lr=1.0, rho=0.95, epsilon=None, decay=0.01
) # NOTE: The details doesn't matter here because we are not training the model
model.compile(loss=args.loss, optimizer=optimizer, metrics=[args.metrics])
# Set model trained weights
model.load_weights(args.model_weights)
# Keras moodel summary, for confirmation
print(model.summary())
# Test the model on data generated batch-by-batch by a Python generator
# NOTE: We use evaluate_generator because we do provide our generated dataset with specific batch size
tensorboard = TensorBoard(log_dir=args.save_dir)
fit_start_time = time()
model.evaluate_generator(test_gen,
steps=args.epoch_steps,
verbose=args.verbose,
callbacks=[tensorboard])
print('### Model fit time (s): ', time() - fit_start_time)
def main():
#test_list = pd.read_csv('gdrive/My Drive/Research/TTLAB/AI_For_Earth/Problem_2/test_2/test_2.csv')
test_list = pd.read_csv('gdrive/My Drive/Research/TTLAB/AI_For_Earth/Problem_2/test_2/test.csv')
testing_img_dir = 'gdrive/My Drive/Research/TTLAB/AI_For_Earth/Problem_2/multiclassification/data/SequoiaMulti_30/test_partition_images/'
weights_path = 'gdrive/My Drive/Research/TTLAB/AI_For_Earth/Problem_2/test_2/650.hdf5'
PRED_FOLDER = 'gdrive/My Drive/Research/TTLAB/AI_For_Earth/Problem_2/test_2/'
img_path = 'gdrive/My Drive/Research/TTLAB/AI_For_Earth/Problem_2/multiclassification/data/SequoiaMulti_30/test_partition_images/full_color_training_1.png'
batch_size = 10
input_shape = (256,256,3)
print('test_list')
print(test_list)
test_gen = data_gen_small(
testing_img_dir,
test_list,
batch_size,
#[args.input_shape[0], args.input_shape[1]],
(input_shape[0], input_shape[1]),
)
n_labels = 3
kernel = 3
pool_size = (2,2)
output_mode = 'softmax'
model = segnet(
input_shape, n_labels, kernel, pool_size, output_mode
)
model.load_weights(weights_path)
'''
original_img = cv2.imread(img_path)[:, :, ::-1]
#print(original_img.shape)
#resized_img = cv2.resize(original_img, dims + [3])
resized_img = cv2.resize(original_img, (256,256) )
array_img = img_to_array(resized_img) / 255
output = model.predict(np.array([array_img]))
image = np.reshape(output, (256,256,3))
p = np.argmax(image, axis=-1)
p = np.expand_dims(p, axis=-1)
p = p * (255/3)
p = p.astype(np.int32)
print(p.shape)
p = np.concatenate([p, p, p], axis=2)
'''
'''
image_2 = visualize(np.argmax(image,axis=-1).reshape((256,256)), False)
print(image_2.shape)
print(image_id)
print(np.unique(image_2))
'''
'''
pred_dir = '4_epochs_predictions'
name = 'img_1' + '.png'
cv2.imwrite(os.path.join(PRED_FOLDER, pred_dir, name), p)
'''
imgs_mask_test = model.predict_generator(
test_gen,
steps=np.ceil(len(test_list) / batch_size)-1)
'''
#original_img = cv2.imread(img_path)[:, :, ::-1]
original_img = np.array([cv2.imread(img_path)])
#nop = np.array([None])
#original_img = np.append(nop, original_img)
print(original_img.shape)
output = model.predict(original_img)
pred = visualize(np.argmax(output[0],axis=1).reshape((256,256)), False)
print(pred.shape)
#plt.imshow(pred)
#plt.show()
cv2.imwrite(os.path.join('gdrive/My Drive/Research/TTLAB/AI_For_Earth/Problem_2/test_2/img_2_300_epoch.png'), output)
#cv2.imwrite(os.path.join('gdrive/My Drive/Research/TTLAB/AI_For_Earth/Problem_2/test_2/img_50_epoch.png'), pred)
'''
pred_dir = '650_epochs_predictions'
if not os.path.exists(os.path.join(PRED_FOLDER, pred_dir)):
os.mkdir(os.path.join(PRED_FOLDER, pred_dir))
for image, image_id in zip(imgs_mask_test, test_list.iloc()):
#image = (image[:, :, 0]).astype(np.float32)
#print(image_id)
#print(image.shape)
#print(np.unique(image[:,0] == image[:,1]))
image = np.reshape(image, (256,256,3))
#print('Image shape')
#print(image.shape)
#print(np.unique(image))
#print(np.argmax(image,axis=-1))
#print(np.unique((np.argmax(image,axis=-1))))
#print('Here')
p = np.argmax(image, axis=-1)
p = np.expand_dims(p, axis=-1)
p = p * (255/3)
p = p.astype(np.int32)
#print(p.shape)
p = np.concatenate([p, p, p], axis=2)
#image_2 = visualize(np.argmax(image,axis=-1).reshape((256,256)), False)
#print(image_2.shape)
#print(image_id)
#print(np.unique(image_2))
#print(image_id[0])
if(image_id[0]%50 == 0):
print(image_id)
name = str(image_id[0]) + '.png'
cv2.imwrite(os.path.join(PRED_FOLDER, pred_dir, name), p)
print("Saving predicted cloud masks on disk... \n")
def reset(self, use_residual=False, use_argmax=True):
self.n_classes = 2
self.model = segnet(input_shape=(768, 768, 3), n_labels=self.n_classes,
kernel=3, pool_size=(2, 2), output_mode="softmax", use_residual=False, use_argmax=True)
def main(args):
# set the necessary list i.e. list of names of images to be read
train_list = pd.read_csv(args.train_list, header=None)
train_list_len = len(train_list)
epoch_steps = int(train_list_len / args.batch_size)
val_list = pd.read_csv(args.val_list, header=None)
val_list_len = len(val_list)
val_steps = int(val_list_len / args.batch_size)
# set the necessary directories
trainimg_dir = args.trainimg_dir
trainmsk_dir = args.trainmsk_dir
valimg_dir = args.valimg_dir
valmsk_dir = args.valmsk_dir
# Generate batch data for SGD
# NOTE: This helps control the batch size for each training and validation set
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)
# Create a model
if args.model == 'unet':
model = unet(\
args.input_shape, args.n_labels, args.kernel, \
args.pool, args.output_mode, args.gpus)
elif args.model == 'segnet':
model = segnet(\
args.input_shape, args.n_labels, args.kernel, \
args.pool, args.output_mode, args.gpus)
elif args.model == 'segunet':
model = segunet(\
args.input_shape, args.n_labels, args.kernel, \
args.pool, args.output_mode, args.gpus)
elif args.model == 'fastnet':
model = fastnet(\
args.input_shape, args.n_labels, args.kernel, \
args.pool, args.output_mode, args.gpus)
# Keras moodel summary
print(model.summary())
# TODO: Configure the model for training
optimizer = Adadelta(lr=args.lr,
rho=args.rho,
epsilon=args.epsilon,
decay=args.decay)
model.compile(loss=args.loss, optimizer=optimizer, metrics=[args.metrics])
# model.compile(loss=args.loss, optimizer=args.optimizer, metrics=[args.metrics])
# If pre-trained weights available, use those
if (args.initial_weights):
model.load_weights(args.initial_weights)
print('Initial weights loaded')
# Generate log for tensorboard (currently only useful parameters are added to the dir name)
log_dir = args.save_dir + \
'model_' + args.model + \
'_batch_size_' + str(args.batch_size) + \
'_epoch_steps_' + str(epoch_steps) + \
'_n_epochs_' + str(args.n_epochs) + \
'_lr_' + str(args.lr) + \
'_decay_' + str(args.decay) + \
'_labels_' + str(args.n_labels) + \
'/'
tensorboard = TensorBoard(log_dir=log_dir)
# Generate checkpoints
checkpoint = ModelCheckpoint(
log_dir + 'weights_at_epoch_{epoch:03d}.hdf5',
verbose=1,
save_best_only=False,
save_weights_only=True,
period=args.period) # Create 10 checkpoints only
# Train the model on data generated batch-by-batch by a Python generator
# NOTE: We use fit_generator because we do provide our generated dataset with specific batch size
fit_start_time = time()
model.fit_generator(train_gen,
steps_per_epoch=epoch_steps,
epochs=args.n_epochs,
validation_data=val_gen,
validation_steps=val_steps,
verbose=args.verbose,
callbacks=[checkpoint, tensorboard])
print('### Model fit time (s): ', time() - fit_start_time)
# # NOTE: Cannot save the whole model OR the model structure if the model is not Keras Sequential, it throws "AttributeError: 'NoneType' object has no attribute 'update'"
# # This is a bug in Keras (not sure about TensorFlow)
# # Therefore, the model strutuce must be generated every time it is required to be implemented outside of the current script
# model.save(log_dir + 'model.hdf5') # Model architecture and final weights
# Save final weights
model.save_weights(log_dir + 'weights_at_epoch_%03d.hdf5' %
(args.n_epochs))
print('Final model weights saved.')
import tensorflow as tf
import numpy as np
import matplotlib.pyplot as plt
from model import segnet, cal_loss, cal_accuracy
from data_loader import get_train_iterator, get_test_iterator
from parameter import *
iterator = get_train_iterator()
with tf.Session() as sess:
sess.run(iterator.initializer)
X, Y = iterator.get_next()
X.set_shape([None, INPUT_HEIGHT, INPUT_WIDTH, INPUT_CHANNEL])
Y.set_shape([None, INPUT_HEIGHT * INPUT_WIDTH, NUM_CLASSES])
logits, prediction = segnet(X, True)
cross_entropy_loss = cal_loss(logits, Y)
accuracy = cal_accuracy(prediction, Y)
init = tf.global_variables_initializer()
saver = tf.train.Saver()
sess.run(init)
saver.restore(sess, MODEL_PATH)
image, pred, loss, acc = sess.run(
[X, prediction, cross_entropy_loss, accuracy])
image = image.astype(int)
for y in range(INPUT_HEIGHT):
for x in range(INPUT_WIDTH):
if pred[0, y, x, 0] < pred[0, y, x, 1]:
image[0, y, x, 0] = 255
image[0, y, x, 1] = 0
image[0, y, x, 2] = 0
print("loss:{:.9f}".format(loss), "accuracy:{:.9f}".format(acc))
def main():
model = segnet('./build.json')
model.train('train_eval.txt', '../tgtdir/', get_batch)
def main(args):
# set the necessary list
#train_list = pd.read_csv(args.train_list, header=None)
train_list = pd.read_csv(
'gdrive/My Drive/Research/TTLAB/AI_For_Earth/Problem_2/test_2/train.csv'
)
#val_list = pd.read_csv(args.val_list, header=None)
# set the necessary directories
#trainimg_dir = args.trainimg_dir
trainimg_dir = 'gdrive/My Drive/Research/TTLAB/AI_For_Earth/Problem_2/multiclassification/data/SequoiaMulti_30/train_partition_images/'
trainmsk_dir = 'gdrive/My Drive/Research/TTLAB/AI_For_Earth/Problem_2/multiclassification/data/SequoiaMulti_30/train_partition_ground_truth/'
#trainmsk_dir = args.trainmsk_dir
#valimg_dir = args.valimg_dir
#valmsk_dir = args.valmsk_dir
train_gen = data_gen_small(
trainimg_dir,
trainmsk_dir,
train_list,
args.batch_size,
#[args.input_shape[0], args.input_shape[1]],
(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,
)
'''
model = segnet(args.input_shape, args.n_labels, args.kernel,
args.pool_size, args.output_mode)
print(model.summary())
model.compile(loss=args.loss,
optimizer=args.optimizer,
metrics=["accuracy"])
addtional = 50
train_resume = True
weights_path = 'gdrive/My Drive/Research/TTLAB/AI_For_Earth/Problem_2/test_2/' + str(
600) + ".hdf5"
print(weights_path)
if train_resume:
model.load_weights(weights_path)
print("\nTraining resumed...")
else:
print("\nTraining started from scratch... ")
model.fit_generator(
train_gen,
steps_per_epoch=args.epoch_steps,
epochs=args.n_epochs,
#validation_data=val_gen,
#validation_steps=args.val_steps,
)
model.save_weights(
'gdrive/My Drive/Research/TTLAB/AI_For_Earth/Problem_2/test_2/' +
str(650) + ".hdf5")
print("sava weight done..")
#-- Validation Dataset --#
dataset['val'] = dataset['val'].map(load_image_test)
dataset['val'] = dataset['val'].repeat()
dataset['val'] = dataset['val'].batch(BATCH_SIZE)
dataset['val'] = dataset['val'].prefetch(buffer_size=AUTOTUNE)
print(dataset['train'])
print(dataset['val'])
for image, mask in dataset['train'].take(1):
sample_image, sample_mask = image, mask
# display_sample([sample_image[0], sample_mask[0]])
m = segnet(input_shape=(128, 128, 3), n_class=45)
m.compile(optimizer=tf.keras.optimizers.Adam(LR),
loss="sparse_categorical_crossentropy",
metrics=metrics)
callbacks = [
ModelCheckpoint("files/model.h5"),
ReduceLROnPlateau(monitor="val_loss", factor=0.1, patience=3),
CSVLogger("files/data.csv"),
TensorBoard(),
EarlyStopping(monitor="val_loss", patience=10, restore_best_weights=False)
]
STEPS_PER_EPOCH = TRAINSET_SIZE // BATCH_SIZE
VALIDATION_STEPS = VALSET_SIZE // BATCH_SIZE
print("VALIDATION DATA SIZE", val_images.shape[0])
print("TEST DATA SIZE", test_images.shape[0])
print(
"****************************************************************************************************************"
)
# *********************** Setup model ********************
input_size = train_images.shape[1:4]
if whichmodel == "param_unet":
model = param_unet(input_size, filters, levels,
dropout_rate)
if whichmodel == "segnet":
model = segnet(input_size, 3, dropout_rate)
# *********************************** Train Model **********************************************
gray_image = False
gray_counter = 0
loss_counter = 0
best_val_loss = 1
acc = []
val_acc = []
loss = []
val_loss = []
for i in range(maxepochs):
print("-------------------- Epoch #", i + 1,
"--------------------")
start_time = time.time()
print("Building model.")
input_shape = (256, 256, 3)
if name_model[i] == "UNet":
model_select = model.Unet(size=input_shape) # 搭建新的模型
model_select.load_weights(
".\\result\\model_record\\20210118_256_49461_UNet_CE.h5"
) # 載入現有完成訓練的權重
batch = 10
train_flag = 0
elif name_model[i] == "SegNet":
continue
model_select = model.segnet(input_shape=input_shape,
n_labels=1,
kernel=3,
pool_size=(2, 2),
output_mode="sigmoid")
model_select.load_weights(
".\\result\\model_record\\20210118_256_49461_SegNet_CE.h5")
batch = 10
train_flag = 0
elif name_model[i] == "DV3":
continue
model_select = DV3.Deeplabv3(weights=None,
input_shape=input_shape,
classes=1)
model_select.load_weights(
".\\result\\model_record\\20210118_256_49461_" +
name_model[i] + "_CE.h5")
batch = 3
