def background_process_augmentation_example_2():
augseq = iaa.Sequential(
[iaa.Fliplr(0.5),
iaa.CoarseDropout(p=0.1, size_percent=0.1)])
num_epoches = 2
for epoch in range(num_epoches):
print('Started epoch {}.'.format(epoch))
batch_loader = ia.BatchLoader(load_batches_from_numpy)
bg_augmenter = ia.BackgroundAugmenter(batch_loader, augseq)
while True:
print('Next batch...')
batch = bg_augmenter.get_batch()
if batch is None:
print('Finished epoch.')
break
#images = batch.images
images_aug = batch.images_aug
#keypoints = batch.keypoints
#keypoints_aug = batch.keypoints_aug
print('Image IDs:', batch.data)
#ia.imshow(np.hstack(list(images_aug)))
batch_loader.terminate()
bg_augmenter.terminate()
def _augment_small_2():
batch_loader = ia.BatchLoader(load_images(n_batches=2),
queue_size=100)
bg_augmenter = ia.BackgroundAugmenter(batch_loader, augseq_i)
i = 0
while True:
batch = bg_augmenter.get_batch()
if batch is None:
break
i += 1
def __next__(self):
# for train_index, valid_index in kf.split(range(self.n_folds)):
if self.n <= self.n_folds:
train_index, valid_index = next(self.kf_split_iter)
# print(train_index, valid_index)
train_datasets = self._accumulate_datasets_(train_index)
valid_datasets = self._accumulate_datasets_(valid_index)
# train_batchgenerator = WorkflowBatchGenerator(data=train_datasets, batch_size=self.batch_size,
# n_batches=self.n_batches)
# valid_batchgenerator = WorkflowBatchGenerator(data=valid_datasets, batch_size=self.batch_size,
# n_batches=self.n_batches)
train_batchloader = ia.BatchLoader(
load_batches(train_datasets, self.batch_size,
self.n_batches)) #, nb_workers=self.num_workers)
valid_batchloader = ia.BatchLoader(
load_batches(valid_datasets, self.batch_size,
self.n_batches)) #, nb_workers=self.num_workers)
train_bg_augmenter = ia.BackgroundAugmenter(
train_batchloader,
self.image_transform,
nb_workers=self.num_workers)
valid_bg_augmenter = ia.BackgroundAugmenter(
valid_batchloader,
self.image_transform,
nb_workers=self.num_workers)
# train_mt_generator = MultiThreadedAugmenter(train_dataloader, self.image_transform, self.num_workers)
# valid_mt_generator = MultiThreadedAugmenter(valid_dataloader, self.image_transform, self.num_workers)
self.n += 1
return train_bg_augmenter, valid_bg_augmenter
else:
raise StopIteration
def getBatches(self, *args, **kwargs):
if self._is_label_augmented:
# FIXME [fix] >> Do not check.
batch_loader = ia.BatchLoader(self._loadBatchPairs)
augmenter_det = self._augmenter.to_deterministic()
bg_augmenter = ia.BackgroundAugmenter(batch_loader, augmenter_det)
while True:
batch = bg_augmenter.get_batch()
if batch is None:
break
yield batch.images_aug, batch.keypoints_aug
batch_loader.terminate()
bg_augmenter.terminate()
else:
batch_loader = ia.BatchLoader(self._loadBatches)
bg_augmenter = ia.BackgroundAugmenter(batch_loader,
self._augmenter)
while True:
batch = bg_augmenter.get_batch()
if batch is None:
break
#images = batch.images
#images_aug = batch.images_aug
#keypoints = batch.keypoints
#keypoints_aug = batch.keypoints_aug
#data = batch.data
yield batch.images_aug, self._labels[batch.data]
batch_loader.terminate()
bg_augmenter.terminate()
def main():
augseq = iaa.Sequential(
[iaa.Fliplr(0.5),
iaa.CoarseDropout(p=0.1, size_percent=0.1)])
print("------------------")
print("augseq.augment_batches(batches, background=True)")
print("------------------")
batches = list(load_images())
batches_aug = augseq.augment_batches(batches, background=True)
images_aug = []
keypoints_aug = []
for batch_aug in batches_aug:
images_aug.append(batch_aug.images_aug)
keypoints_aug.append(batch_aug.keypoints_aug)
ia.imshow(draw_grid(images_aug, keypoints_aug))
print("------------------")
print("augseq.augment_batches(batches, background=True) -> only images")
print("------------------")
batches = list(load_images())
batches = [batch.images for batch in batches]
batches_aug = augseq.augment_batches(batches, background=True)
images_aug = []
keypoints_aug = None
for batch_aug in batches_aug:
images_aug.append(batch_aug)
ia.imshow(draw_grid(images_aug, keypoints_aug))
print("------------------")
print("BackgroundAugmenter")
print("------------------")
batch_loader = ia.BatchLoader(load_images)
bg_augmenter = ia.BackgroundAugmenter(batch_loader, augseq)
images_aug = []
keypoints_aug = []
while True:
print("Next batch...")
batch = bg_augmenter.get_batch()
if batch is None:
print("Finished.")
break
images_aug.append(batch.images_aug)
keypoints_aug.append(batch.keypoints_aug)
ia.imshow(draw_grid(images_aug, keypoints_aug))
def background_process_augmentation_example_1():
# Example augmentation sequence to run in the background.
augseq = iaa.Sequential(
[iaa.Fliplr(0.5),
iaa.CoarseDropout(p=0.1, size_percent=0.1)])
print('Started epoch.')
# Background augmentation consists of two components:
# (1) BatchLoader, which runs in a Thread and calls repeatedly a user-defined
# function (here: load_batches) to load batches (optionally with keypoints
# and additional information) and sends them to a queue of batches.
# (2) BackgroundAugmenter, which runs several background processes (on other
# CPU cores). Each process takes batches from the queue defined by (1),
# augments images/keypoints and sends them to another queue.
# The main process can then read augmented batches from the queue defined by (2).
batch_loader = ia.BatchLoader(load_batches)
bg_augmenter = ia.BackgroundAugmenter(batch_loader, augseq)
# Run until load_batches() returns nothing anymore. This also allows infinite training.
while True:
print('Next batch...')
batch = bg_augmenter.get_batch()
if batch is None:
print('Finished epoch.')
break
#images = batch.images
images_aug = batch.images_aug
#keypoints = batch.keypoints
#keypoints_aug = batch.keypoints_aug
print('Image IDs:', batch.data)
#ia.imshow(np.hstack(list(images_aug)))
batch_loader.terminate()
bg_augmenter.terminate()
def main():
batch_size = 100
output_txt = '/media/frelam/683cd494-d120-4dbd-81a4-eb3a90330106/ai_challange_scene_dataset_lmdb/aichallenger_arug20171116/dataarymentation/train_arug2.txt'
dataset_root = '/home/frelam/ai_challenger_scene_train_20170904/train/scene_train_images_20170904/'
dataarumation_root = '/media/frelam/683cd494-d120-4dbd-81a4-eb3a90330106/ai_challange_scene_dataset_lmdb/aichallenger_arug20171116/dataarymentation/scene_train_images_arug20171117/'
duplicate_num_per_image = 10
seq = seq_setting()
#input:
# image_dir_array
# image_label_array
image_dir_array, image_label_array = load_image_dir(f1)
dir_chunk_array = [
image_dir_array[x:x + batch_size]
for x in xrange(0, len(image_dir_array), batch_size)
]
label_chunk_array = [
image_label_array[x:x + batch_size]
for x in xrange(0, len(image_label_array), batch_size)
]
dir_for_save = []
label_for_save = []
#batch process
for i in range(len(dir_chunk_array)):
batch_loader = ia.BatchLoader(
process_batch_image(dataset_root, dataarumation_root,
duplicate_num_per_image, dir_chunk_array[i],
label_chunk_array[i], seq))
dir_for_save.append(dirs)
label_for_save.append(labels)
print i
with open(output_txt, 'w') as output:
for i in range(len(dir_for_save)):
for j in range(len(dir_for_save[i])):
output.write(dir_for_save[i][j] + ' ' + label_for_save[i][j] +
'\n')
def finetune(FLAG):
# already exist (uploaded to github)
with open('save/label_dict.pkl', 'rb') as f:
y_dict = pickle.load(f)
dtrain = pd.read_csv(FLAG.train_file,
header=None,
sep=" ",
names=["img", "id"])
dvalid = pd.read_csv(FLAG.valid_file,
header=None,
sep=" ",
names=["img", "id"])
train_list = [
os.path.join(FLAG.train_path, img) for img in list(dtrain.img)
]
valid_list = [
os.path.join(FLAG.valid_path, img) for img in list(dvalid.img)
]
print("Reading train and valid images")
Xtrain = readImgList(train_list)
print("train: {0}".format(Xtrain.shape))
Xvalid = readImgList(valid_list)
print("valid: {0}".format(Xvalid.shape))
ytrain = transformLabel(list(dtrain.id), y_dict)
yvalid = transformLabel(list(dvalid.id), y_dict)
Ytrain = one_hot_encoding(ytrain, len(y_dict))
Yvalid = one_hot_encoding(yvalid, len(y_dict))
print("Building model")
scope_name = "Model"
model = VGG16_GAP(scope_name=scope_name)
model.build(vgg16_npy_path=FLAG.init_from,
shape=Xtrain.shape[1:],
classes=len(y_dict),
prof_type=FLAG.prof_type,
conv_pre_training=True,
fc_pre_training=True,
new_bn=False)
if FLAG.centers is not None:
centers = np.load(FLAG.centers)
model.add_centers(centers.astype(np.float32))
else:
print(
"please specify your center.npy or initialize centers with all zeros"
)
model.add_centers()
print("Setting operations at various levels")
dp = [1.0, 0.75, 0.5]
tasks = [str(int(p * 100)) for p in dp]
model.set_idp_operation(dp=dp,
decay=FLAG.decay,
keep_prob=FLAG.keep_prob,
lambda_c=FLAG.lambda_c)
obj = 0.0
for cur_task in tasks:
print(cur_task)
obj += model.loss_dict[cur_task]
tracking = list()
for cur_task in tasks:
tracking.append(model.accu_dict[cur_task])
# data augmenter
sometimes = lambda aug: iaa.Sometimes(0.5, aug)
transform = iaa.Sequential([
sometimes(
iaa.Affine(translate_percent={
"x": (-0.15, 0.15),
"y": (-0.15, 0.15)
})),
sometimes(iaa.Affine(scale={
"x": (0.85, 1.15),
"y": (0.85, 1.15)
})),
sometimes(iaa.Affine(rotate=(-45, 45))),
sometimes(iaa.Fliplr(0.5))
])
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
augment = True
# hyper parameters
batch_size = 32
epoch = 100
early_stop_patience = 10
min_delta = 0.0001
# recorder
epoch_counter = 0
history = list()
# Passing global_step to minimize() will increment it at each step.
learning_rate = FLAG.learning_rate
opt = tf.train.AdamOptimizer(learning_rate=learning_rate, beta1=0.5)
checkpoint_path = os.path.join(FLAG.save_dir, 'model.ckpt')
# trainable variables
train_vars = list()
for var in tf.trainable_variables():
if model.scope_name in var.name:
train_vars.append(var)
for rm in model.gamma_var:
train_vars.remove(rm)
print('%s is not trainable.' % rm)
for var in train_vars:
if '_mean' in var.name:
train_vars.remove(var)
print('%s is not trainable.' % var)
for var in train_vars:
if '_beta' in var.name:
train_vars.remove(var)
print('%s is not trainable.' % var)
for var in train_vars:
if '_variance' in var.name:
train_vars.remove(var)
print('%s is not trainable.' % var)
print(train_vars)
train_op = opt.minimize(obj, var_list=train_vars)
saver = tf.train.Saver(tf.global_variables(), max_to_keep=len(tasks))
# max step in a epoch
ptrain_max = int(Xtrain.shape[0] / batch_size)
pval_max = int(Xvalid.shape[0] / batch_size)
# re-initialize
def initialize_uninitialized(sess):
global_vars = tf.global_variables()
is_not_initialized = sess.run(
[tf.is_variable_initialized(var) for var in global_vars])
not_initialized_vars = [
v for (v, f) in zip(global_vars, is_not_initialized) if not f
]
if len(not_initialized_vars):
sess.run(tf.variables_initializer(not_initialized_vars))
initialize_uninitialized(sess)
# reset due to adding a new task
patience_counter = 0
current_best_val_accu = 0
# optimize when the aggregated obj
while (patience_counter < early_stop_patience
and epoch_counter < epoch):
# start training
stime = time.time()
train_loss, train_accu = 0.0, 0.0
if augment:
def load_batches():
for i in range(int(Xtrain.shape[0] / batch_size)):
print("Training: {0}/{1}".format(i, ptrain_max),
end='\r')
st = i * batch_size
ed = (i + 1) * batch_size
batch = ia.Batch(images=Xtrain[st:ed, :, :, :],
data=Ytrain[st:ed, :])
yield batch
batch_loader = ia.BatchLoader(load_batches)
bg_augmenter = ia.BackgroundAugmenter(
batch_loader=batch_loader, augseq=transform, nb_workers=1)
while True:
batch = bg_augmenter.get_batch()
if batch is None:
print("Finished epoch.")
break
x_images_aug = batch.images_aug
y_images = batch.data
loss, accu, _, _ = sess.run(
[
obj, model.accu_dict[cur_task], train_op,
model.centers_update_op
],
feed_dict={
model.x: x_images_aug,
model.y: y_images,
model.is_train: True,
model.bn_train: False
})
train_loss += loss
train_accu += accu
batch_loader.terminate()
bg_augmenter.terminate()
else:
for i in range(int(Xtrain.shape[0] / batch_size)):
print("Training: {0}/{1}".format(i, ptrain_max), end='\r')
st = i * batch_size
ed = (i + 1) * batch_size
loss, accu, _, _ = sess.run(
[
obj, model.accu_dict[tasks[0]], train_op,
model.centers_update_op
],
feed_dict={
model.x: Xtrain[st:ed, :],
model.y: Ytrain[st:ed, :],
model.is_train: True,
model.bn_train: False
})
train_loss += loss
train_accu += accu
train_loss = train_loss / ptrain_max
train_accu = train_accu / ptrain_max
# validation
val_loss, val_accu1, val_accu2 = 0.0, 0.0, 0.0
val_accu_dp = list()
for i in range(int(Xvalid.shape[0] / batch_size)):
print("Validating: {0}/{1}".format(i, pval_max), end='\r')
st = i * batch_size
ed = (i + 1) * batch_size
loss, accu1, accu2, accu_dp = sess.run(
[
obj, model.accu_dict[tasks[0]],
model.accu_dict[tasks[-1]], tracking
],
feed_dict={
model.x: Xvalid[st:ed, :],
model.y: Yvalid[st:ed, :],
model.is_train: False,
model.bn_train: False
})
val_loss += loss
val_accu1 += accu1
val_accu2 += accu2
val_accu_dp.append(accu_dp)
val_accu_dp = np.mean(val_accu_dp, axis=0).tolist()
dp_str = ""
for i in range(len(tasks)):
dp_str += "{0}%:{1}, ".format(tasks[i],
np.round(val_accu_dp[i], 4))
print(dp_str)
val_loss = val_loss / pval_max
val_accu1 = val_accu1 / pval_max
val_accu2 = val_accu2 / pval_max
val_accu = val_accu1 # used for early stopping
# early stopping check
if (val_accu - current_best_val_accu) > min_delta:
current_best_val_accu = val_accu
patience_counter = 0
para_dict = sess.run(model.para_dict)
np.save(os.path.join(FLAG.save_dir, "para_dict.npy"),
para_dict)
print("save in %s" %
os.path.join(FLAG.save_dir, "para_dict.npy"))
else:
patience_counter += 1
# shuffle Xtrain and Ytrain in the next epoch
idx = np.random.permutation(Xtrain.shape[0])
Xtrain, Ytrain = Xtrain[idx, :, :, :], Ytrain[idx, :]
# epoch end
epoch_counter += 1
print(
"Epoch %s (%s), %s sec >> train loss: %.4f, train accu: %.4f, val loss: %.4f, val accu at %s: %.4f, val accu at %s: %.4f"
% (epoch_counter, patience_counter,
round(time.time() - stime, 2), train_loss, train_accu,
val_loss, tasks[0], val_accu1, tasks[-1], val_accu2))
history.append([train_loss, train_accu, val_loss, val_accu])
if epoch_counter % 10 == 0:
import matplotlib.pyplot as plt
df = pd.DataFrame(history)
df.columns = [
'train_loss', 'train_accu', 'val_loss', 'val_accu'
]
df[['train_loss', 'val_loss']].plot()
plt.savefig(os.path.join(FLAG.save_dir, 'loss.png'))
plt.close()
df[['train_accu', 'val_accu']].plot()
plt.savefig(os.path.join(FLAG.save_dir, 'accu.png'))
plt.close()
saver.save(sess, checkpoint_path, global_step=epoch_counter)
# extract features and calculate center
output = []
for i in range(int(Xtrain.shape[0] / 200 + 1)):
print(i, end="\r")
st = i * 200
ed = min((i + 1) * 200, Xtrain.shape[0])
prob = sess.run(model.features,
feed_dict={
model.x: Xtrain[st:ed, :],
model.is_train: False,
model.bn_train: False
})
output.append(prob)
for i in range(int(Xvalid.shape[0] / 200 + 1)):
print(i, end="\r")
st = i * 200
ed = min((i + 1) * 200, Xvalid.shape[0])
prob = sess.run(model.features,
feed_dict={
model.x: Xvalid[st:ed, :],
model.is_train: False,
model.bn_train: False
})
output.append(prob)
EX = np.concatenate(output)
print(EX.shape)
EY = np.concatenate([ytrain, yvalid])
print(EY.shape)
centers = np.zeros((len(y_dict), EX.shape[1]))
for i in range(len(y_dict)):
centers[i, :] = np.mean(EX[EY == i, :], axis=0)
np.save(arr=centers,
file=os.path.join(FLAG.save_dir, "centers.npy"))
def main():
augseq = iaa.Sequential(
[iaa.Fliplr(0.5),
iaa.CoarseDropout(p=0.1, size_percent=0.1)])
def func_images(images, random_state, parents, hooks):
time.sleep(0.2)
return images
def func_heatmaps(heatmaps, random_state, parents, hooks):
return heatmaps
def func_keypoints(keypoints_on_images, random_state, parents, hooks):
return keypoints_on_images
augseq_slow = iaa.Sequential([
iaa.Fliplr(0.5),
iaa.Lambda(func_images=func_images,
func_heatmaps=func_heatmaps,
func_keypoints=func_keypoints)
])
print("------------------")
print("augseq.augment_batches(batches, background=True)")
print("------------------")
batches = list(load_images())
batches_aug = augseq.augment_batches(batches, background=True)
images_aug = []
keypoints_aug = []
for batch_aug in batches_aug:
images_aug.append(batch_aug.images_aug)
keypoints_aug.append(batch_aug.keypoints_aug)
ia.imshow(draw_grid(images_aug, keypoints_aug))
print("------------------")
print("augseq.augment_batches(batches, background=True) -> only images")
print("------------------")
batches = list(load_images())
batches = [batch.images for batch in batches]
batches_aug = augseq.augment_batches(batches, background=True)
images_aug = []
keypoints_aug = None
for batch_aug in batches_aug:
images_aug.append(batch_aug)
ia.imshow(draw_grid(images_aug, keypoints_aug))
print("------------------")
print("BackgroundAugmenter")
print("------------------")
batch_loader = ia.BatchLoader(load_images)
bg_augmenter = ia.BackgroundAugmenter(batch_loader, augseq)
images_aug = []
keypoints_aug = []
while True:
print("Next batch...")
batch = bg_augmenter.get_batch()
if batch is None:
print("Finished.")
break
images_aug.append(batch.images_aug)
keypoints_aug.append(batch.keypoints_aug)
ia.imshow(draw_grid(images_aug, keypoints_aug))
print("------------------")
print("BackgroundAugmenter with generator in BL")
print("------------------")
batch_loader = ia.BatchLoader(load_images())
bg_augmenter = ia.BackgroundAugmenter(batch_loader, augseq)
images_aug = []
keypoints_aug = []
while True:
print("Next batch...")
batch = bg_augmenter.get_batch()
if batch is None:
print("Finished.")
break
images_aug.append(batch.images_aug)
keypoints_aug.append(batch.keypoints_aug)
ia.imshow(draw_grid(images_aug, keypoints_aug))
print("------------------")
print("Long running BackgroundAugmenter at BL-queue_size=12")
print("------------------")
batch_loader = ia.BatchLoader(load_images(n_batches=1000), queue_size=12)
bg_augmenter = ia.BackgroundAugmenter(batch_loader, augseq)
i = 0
while True:
if i % 100 == 0:
print("batch=%d..." % (i, ))
batch = bg_augmenter.get_batch()
if batch is None:
print("Finished.")
break
i += 1
print("------------------")
print("Long running BackgroundAugmenter at BL-queue_size=2")
print("------------------")
batch_loader = ia.BatchLoader(load_images(n_batches=1000), queue_size=2)
bg_augmenter = ia.BackgroundAugmenter(batch_loader, augseq)
i = 0
while True:
if i % 100 == 0:
print("batch=%d..." % (i, ))
batch = bg_augmenter.get_batch()
if batch is None:
print("Finished.")
break
i += 1
print("------------------")
print("Long running BackgroundAugmenter (slow loading)")
print("------------------")
batch_loader = ia.BatchLoader(load_images(n_batches=100, sleep=0.2))
bg_augmenter = ia.BackgroundAugmenter(batch_loader, augseq)
i = 0
while True:
if i % 10 == 0:
print("batch=%d..." % (i, ))
batch = bg_augmenter.get_batch()
if batch is None:
print("Finished.")
break
i += 1
print("------------------")
print("Long running BackgroundAugmenter (slow aug) at BL-queue_size=12")
print("------------------")
batch_loader = ia.BatchLoader(load_images(n_batches=100), queue_size=12)
bg_augmenter = ia.BackgroundAugmenter(batch_loader, augseq_slow)
i = 0
while True:
if i % 10 == 0:
print("batch=%d..." % (i, ))
batch = bg_augmenter.get_batch()
if batch is None:
print("Finished.")
break
i += 1
print("------------------")
print("Long running BackgroundAugmenter (slow aug) at BL-queue_size=2")
print("------------------")
batch_loader = ia.BatchLoader(load_images(n_batches=100), queue_size=2)
bg_augmenter = ia.BackgroundAugmenter(batch_loader, augseq_slow)
i = 0
while True:
if i % 10 == 0:
print("batch=%d..." % (i, ))
batch = bg_augmenter.get_batch()
if batch is None:
print("Finished.")
break
i += 1
def train(FLAG):
print("Reading dataset...")
if FLAG.dataset == 'CIFAR-10':
train_data = CIFAR10(train=True)
test_data = CIFAR10(train=False)
vgg16 = VGG16(classes=10)
elif FLAG.dataset == 'CIFAR-100':
train_data = CIFAR100(train=True)
test_data = CIFAR100(train=False)
vgg16 = VGG16(classes=100)
else:
raise ValueError("dataset should be either CIFAR-10 or CIFAR-100.")
print("Build VGG16 models for %s..." % FLAG.dataset)
Xtrain, Ytrain = train_data.train_data, train_data.train_labels
Xtest, Ytest = test_data.test_data, test_data.test_labels
vgg16.build(vgg16_npy_path=FLAG.init_from,
prof_type=FLAG.prof_type,
conv_pre_training=True,
fc_pre_training=False)
vgg16.sparsity_train(l1_gamma=FLAG.lambda_s,
l1_gamma_diff=FLAG.lambda_m,
decay=FLAG.decay,
keep_prob=FLAG.keep_prob)
# define tasks
tasks = ['var_dp']
print(tasks)
# initial task
cur_task = tasks[0]
obj = vgg16.loss_dict[tasks[0]]
saver = tf.train.Saver(tf.global_variables(), max_to_keep=len(tasks))
checkpoint_path = os.path.join(FLAG.save_dir, 'model.ckpt')
tvars_trainable = tf.trainable_variables()
#for rm in vgg16.gamma_var:
# tvars_trainable.remove(rm)
# print('%s is not trainable.'% rm)
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
# hyper parameters
batch_size = 64
epoch = 500
early_stop_patience = 50
min_delta = 0.0001
opt_type = 'adam'
# recorder
epoch_counter = 0
# optimizer
global_step = tf.Variable(0, trainable=False)
# Passing global_step to minimize() will increment it at each step.
if opt_type is 'sgd':
start_learning_rate = 1e-4 # adam # 4e-3 #sgd
half_cycle = 20000
learning_rate = tf.train.exponential_decay(start_learning_rate,
global_step,
half_cycle,
0.5,
staircase=True)
opt = tf.train.MomentumOptimizer(learning_rate=learning_rate,
momentum=0.9,
use_nesterov=True)
else:
start_learning_rate = 1e-4 # adam # 4e-3 #sgd
half_cycle = 10000
learning_rate = tf.train.exponential_decay(start_learning_rate,
global_step,
half_cycle,
0.5,
staircase=True)
opt = tf.train.AdamOptimizer(learning_rate=learning_rate)
train_op = opt.minimize(obj,
global_step=global_step,
var_list=tvars_trainable)
# progress bar
ptrain = IntProgress()
pval = IntProgress()
display(ptrain)
display(pval)
ptrain.max = int(Xtrain.shape[0] / batch_size)
pval.max = int(Xtest.shape[0] / batch_size)
spareness = vgg16.spareness(thresh=0.05)
print("initial spareness: %s" % sess.run(spareness))
# re-initialize
initialize_uninitialized(sess)
# reset due to adding a new task
patience_counter = 0
current_best_val_accu = 0
# optimize when the aggregated obj
while (patience_counter < early_stop_patience
and epoch_counter < epoch):
def load_batches():
for i in range(int(Xtrain.shape[0] / batch_size)):
st = i * batch_size
ed = (i + 1) * batch_size
batch = ia.Batch(images=Xtrain[st:ed, :, :, :],
data=Ytrain[st:ed, :])
yield batch
batch_loader = ia.BatchLoader(load_batches)
bg_augmenter = ia.BackgroundAugmenter(batch_loader=batch_loader,
augseq=transform,
nb_workers=4)
# start training
stime = time.time()
bar_train = Bar(
'Training',
max=int(Xtrain.shape[0] / batch_size),
suffix='%(index)d/%(max)d - %(percent).1f%% - %(eta)ds')
bar_val = Bar(
'Validation',
max=int(Xtest.shape[0] / batch_size),
suffix='%(index)d/%(max)d - %(percent).1f%% - %(eta)ds')
train_loss, train_accu = 0.0, 0.0
while True:
batch = bg_augmenter.get_batch()
if batch is None:
print("Finished epoch.")
break
x_images_aug = batch.images_aug
y_images = batch.data
loss, accu, _ = sess.run(
[obj, vgg16.accu_dict[cur_task], train_op],
feed_dict={
vgg16.x: x_images_aug,
vgg16.y: y_images,
vgg16.is_train: True
})
bar_train.next()
train_loss += loss
train_accu += accu
ptrain.value += 1
ptrain.description = "Training %s/%s" % (ptrain.value,
ptrain.max)
train_loss = train_loss / ptrain.value
train_accu = train_accu / ptrain.value
batch_loader.terminate()
bg_augmenter.terminate()
# # training an epoch
# for i in range(int(Xtrain.shape[0]/batch_size)):
# st = i*batch_size
# ed = (i+1)*batch_size
# augX = transform.augment_images(Xtrain[st:ed,:,:,:])
# sess.run([train_op], feed_dict={vgg16.x: augX,
# vgg16.y: Ytrain[st:ed,:],
# vgg16.is_train: False})
# ptrain.value +=1
# ptrain.description = "Training %s/%s" % (i, ptrain.max)
# bar_train.next()
# validation
val_loss = 0
val_accu = 0
for i in range(int(Xtest.shape[0] / 200)):
st = i * 200
ed = (i + 1) * 200
loss, accu = sess.run(
[obj, vgg16.accu_dict[cur_task]],
feed_dict={
vgg16.x: Xtest[st:ed, :],
vgg16.y: Ytest[st:ed, :],
vgg16.is_train: False
})
val_loss += loss
val_accu += accu
pval.value += 1
pval.description = "Testing %s/%s" % (pval.value, pval.value)
val_loss = val_loss / pval.value
val_accu = val_accu / pval.value
print("\nspareness: %s" % sess.run(spareness))
# early stopping check
if (val_accu - current_best_val_accu) > min_delta:
current_best_val_accu = val_accu
patience_counter = 0
para_dict = sess.run(vgg16.para_dict)
np.save(os.path.join(FLAG.save_dir, "para_dict.npy"),
para_dict)
print("save in %s" %
os.path.join(FLAG.save_dir, "para_dict.npy"))
else:
patience_counter += 1
# shuffle Xtrain and Ytrain in the next epoch
idx = np.random.permutation(Xtrain.shape[0])
Xtrain, Ytrain = Xtrain[idx, :, :, :], Ytrain[idx, :]
# epoch end
# writer.add_summary(epoch_summary, epoch_counter)
epoch_counter += 1
ptrain.value = 0
pval.value = 0
bar_train.finish()
bar_val.finish()
print(
"Epoch %s (%s), %s sec >> train loss: %.4f, train accu: %.4f, val loss: %.4f, val accu at %s: %.4f"
% (epoch_counter, patience_counter,
round(time.time() - stime, 2), train_loss, train_accu,
val_loss, cur_task, val_accu))
saver.save(sess, checkpoint_path, global_step=epoch_counter)
sp, rcut = gammaSparsifyVGG16(para_dict, thresh=0.02)
np.save(os.path.join(FLAG.save_dir, "sparse_dict.npy"), sp)
print("sparsify %s in %s" % (np.round(
1 - rcut, 3), os.path.join(FLAG.save_dir, "sparse_dict.npy")))
#writer.close()
arr_spareness.append(1 - rcut)
np.save(os.path.join(FLAG.save_dir, "sprocess.npy"), arr_spareness)
FLAG.optimizer = opt_type
FLAG.lr = start_learning_rate
FLAG.batch_size = batch_size
FLAG.epoch_end = epoch_counter
FLAG.val_accu = current_best_val_accu
header = ''
row = ''
for key in sorted(vars(FLAG)):
if header is '':
header = key
row = str(getattr(FLAG, key))
else:
header += "," + key
row += "," + str(getattr(FLAG, key))
row += "\n"
header += "\n"
if os.path.exists("/home/cmchang/new_CP_CNN/model.csv"):
with open("/home/cmchang/new_CP_CNN/model.csv", "a") as myfile:
myfile.write(row)
else:
with open("/home/cmchang/new_CP_CNN/model.csv", "w") as myfile:
myfile.write(header)
myfile.write(row)
def example_background_classes():
print("Example: Background Augmentation via Classes")
import imgaug as ia
from imgaug import augmenters as iaa
import numpy as np
from skimage import data
# Example augmentation sequence to run in the background.
augseq = iaa.Sequential(
[iaa.Fliplr(0.5),
iaa.CoarseDropout(p=0.1, size_percent=0.1)])
# A generator that loads batches from the hard drive.
def load_batches():
# Here, load 10 batches of size 4 each.
# You can also load an infinite amount of batches, if you don't train
# in epochs.
batch_size = 4
nb_batches = 10
# Here, for simplicity we just always use the same image.
astronaut = data.astronaut()
astronaut = ia.imresize_single_image(astronaut, (64, 64))
for i in range(nb_batches):
# A list containing all images of the batch.
batch_images = []
# A list containing IDs per image. This is not necessary for the
# background augmentation and here just used to showcase that you
# can transfer additional information.
batch_data = []
# Add some images to the batch.
for b in range(batch_size):
batch_images.append(astronaut)
batch_data.append((i, b))
# Create the batch object to send to the background processes.
batch = ia.Batch(images=np.array(batch_images, dtype=np.uint8),
data=batch_data)
yield batch
# background augmentation consists of two components:
# (1) BatchLoader, which runs in a Thread and calls repeatedly a user-defined
# function (here: load_batches) to load batches (optionally with keypoints
# and additional information) and sends them to a queue of batches.
# (2) BackgroundAugmenter, which runs several background processes (on other
# CPU cores). Each process takes batches from the queue defined by (1),
# augments images/keypoints and sends them to another queue.
# The main process can then read augmented batches from the queue defined
# by (2).
batch_loader = ia.BatchLoader(load_batches)
bg_augmenter = ia.BackgroundAugmenter(batch_loader, augseq)
# Run until load_batches() returns nothing anymore. This also allows infinite
# training.
while True:
print("Next batch...")
batch = bg_augmenter.get_batch()
if batch is None:
print("Finished epoch.")
break
images_aug = batch.images_aug
print("Image IDs: ", batch.data)
ia.imshow(np.hstack(list(images_aug)))
batch_loader.terminate()
bg_augmenter.terminate()
def _augment_small_4():
batch_loader = ia.BatchLoader(load_images(n_batches=10),
queue_size=100)
bg_augmenter = ia.BackgroundAugmenter(batch_loader, augseq_i)
batch = bg_augmenter.get_batch()
def main():
augseq = iaa.Sequential(
[iaa.Fliplr(0.5),
iaa.CoarseDropout(p=0.1, size_percent=0.1)])
def func_images(images, random_state, parents, hooks):
time.sleep(0.2)
return images
def func_heatmaps(heatmaps, random_state, parents, hooks):
return heatmaps
def func_keypoints(keypoints_on_images, random_state, parents, hooks):
return keypoints_on_images
augseq_slow = iaa.Sequential([
iaa.Fliplr(0.5),
iaa.Lambda(func_images=func_images,
func_heatmaps=func_heatmaps,
func_keypoints=func_keypoints)
])
print("------------------")
print("augseq.augment_batches(batches, background=True)")
print("------------------")
batches = list(load_images())
batches_aug = augseq.augment_batches(batches, background=True)
images_aug = []
keypoints_aug = []
for batch_aug in batches_aug:
images_aug.append(batch_aug.images_aug)
keypoints_aug.append(batch_aug.keypoints_aug)
ia.imshow(draw_grid(images_aug, keypoints_aug))
print("------------------")
print("augseq.augment_batches(batches, background=True) -> only images")
print("------------------")
batches = list(load_images())
batches = [batch.images for batch in batches]
batches_aug = augseq.augment_batches(batches, background=True)
images_aug = []
keypoints_aug = None
for batch_aug in batches_aug:
images_aug.append(batch_aug)
ia.imshow(draw_grid(images_aug, keypoints_aug))
print("------------------")
print("BackgroundAugmenter")
print("------------------")
batch_loader = ia.BatchLoader(load_images)
bg_augmenter = ia.BackgroundAugmenter(batch_loader, augseq)
images_aug = []
keypoints_aug = []
while True:
print("Next batch...")
batch = bg_augmenter.get_batch()
if batch is None:
print("Finished.")
break
images_aug.append(batch.images_aug)
keypoints_aug.append(batch.keypoints_aug)
ia.imshow(draw_grid(images_aug, keypoints_aug))
print("------------------")
print("BackgroundAugmenter with generator in BL")
print("------------------")
batch_loader = ia.BatchLoader(load_images())
bg_augmenter = ia.BackgroundAugmenter(batch_loader, augseq)
images_aug = []
keypoints_aug = []
while True:
print("Next batch...")
batch = bg_augmenter.get_batch()
if batch is None:
print("Finished.")
break
images_aug.append(batch.images_aug)
keypoints_aug.append(batch.keypoints_aug)
ia.imshow(draw_grid(images_aug, keypoints_aug))
print("------------------")
print("Long running BackgroundAugmenter at BL-queue_size=12")
print("------------------")
batch_loader = ia.BatchLoader(load_images(n_batches=1000), queue_size=12)
bg_augmenter = ia.BackgroundAugmenter(batch_loader, augseq)
i = 0
while True:
if i % 100 == 0:
print("batch=%d..." % (i, ))
batch = bg_augmenter.get_batch()
if batch is None:
print("Finished.")
break
i += 1
print("------------------")
print("Long running BackgroundAugmenter at BL-queue_size=2")
print("------------------")
batch_loader = ia.BatchLoader(load_images(n_batches=1000), queue_size=2)
bg_augmenter = ia.BackgroundAugmenter(batch_loader, augseq)
i = 0
while True:
if i % 100 == 0:
print("batch=%d..." % (i, ))
batch = bg_augmenter.get_batch()
if batch is None:
print("Finished.")
break
i += 1
print("------------------")
print("Long running BackgroundAugmenter (slow loading)")
print("------------------")
batch_loader = ia.BatchLoader(load_images(n_batches=100, sleep=0.2))
bg_augmenter = ia.BackgroundAugmenter(batch_loader, augseq)
i = 0
while True:
if i % 10 == 0:
print("batch=%d..." % (i, ))
batch = bg_augmenter.get_batch()
if batch is None:
print("Finished.")
break
i += 1
print("------------------")
print("Long running BackgroundAugmenter (slow aug) at BL-queue_size=12")
print("------------------")
batch_loader = ia.BatchLoader(load_images(n_batches=100), queue_size=12)
bg_augmenter = ia.BackgroundAugmenter(batch_loader, augseq_slow)
i = 0
while True:
if i % 10 == 0:
print("batch=%d..." % (i, ))
batch = bg_augmenter.get_batch()
if batch is None:
print("Finished.")
break
i += 1
print("------------------")
print("Long running BackgroundAugmenter (slow aug) at BL-queue_size=2")
print("------------------")
batch_loader = ia.BatchLoader(load_images(n_batches=100), queue_size=2)
bg_augmenter = ia.BackgroundAugmenter(batch_loader, augseq_slow)
i = 0
while True:
if i % 10 == 0:
print("batch=%d..." % (i, ))
batch = bg_augmenter.get_batch()
if batch is None:
print("Finished.")
break
i += 1
for augseq_i in [augseq, augseq_slow]:
print("------------------")
print("Many very small runs (batches=1)")
print("------------------")
for i in range(100):
batch_loader = ia.BatchLoader(load_images(n_batches=1),
queue_size=100)
bg_augmenter = ia.BackgroundAugmenter(batch_loader, augseq_i)
while True:
batch = bg_augmenter.get_batch()
if batch is None:
print("Finished (%d/%d)." % (i + 1, 100))
break
print("------------------")
print("Many very small runs (batches=2)")
print("------------------")
for i in range(100):
batch_loader = ia.BatchLoader(load_images(n_batches=2),
queue_size=100)
bg_augmenter = ia.BackgroundAugmenter(batch_loader, augseq_i)
while True:
batch = bg_augmenter.get_batch()
if batch is None:
print("Finished (%d/%d)." % (i + 1, 100))
break
print("------------------")
print("Many very small runs, separate function (batches=1)")
print("------------------")
def _augment_small_1():
batch_loader = ia.BatchLoader(load_images(n_batches=1),
queue_size=100)
bg_augmenter = ia.BackgroundAugmenter(batch_loader, augseq_i)
i = 0
while True:
batch = bg_augmenter.get_batch()
if batch is None:
break
i += 1
for i in range(100):
_augment_small_1()
print("Finished (%d/%d)." % (i + 1, 100))
print("------------------")
print("Many very small runs, separate function (batches=2)")
print("------------------")
def _augment_small_2():
batch_loader = ia.BatchLoader(load_images(n_batches=2),
queue_size=100)
bg_augmenter = ia.BackgroundAugmenter(batch_loader, augseq_i)
i = 0
while True:
batch = bg_augmenter.get_batch()
if batch is None:
break
i += 1
for i in range(100):
_augment_small_2()
print("Finished (%d/%d)." % (i + 1, 100))
print("------------------")
print(
"Many very small runs, separate function, incomplete fetching (batches=2)"
)
print("------------------")
def _augment_small_3():
batch_loader = ia.BatchLoader(load_images(n_batches=2),
queue_size=100)
bg_augmenter = ia.BackgroundAugmenter(batch_loader, augseq_i)
batch = bg_augmenter.get_batch()
for i in range(100):
_augment_small_3()
print("Finished (%d/%d)." % (i + 1, 100))
#for augseq_i in [augseq, augseq_slow]:
print("------------------")
print(
"Many very small runs, separate function, incomplete fetching (batches=10)"
)
print("------------------")
def _augment_small_4():
batch_loader = ia.BatchLoader(load_images(n_batches=10),
queue_size=100)
bg_augmenter = ia.BackgroundAugmenter(batch_loader, augseq_i)
batch = bg_augmenter.get_batch()
#bg_augmenter.terminate()
for i in range(100):
_augment_small_4()
print("Finished (%d/%d)." % (i + 1, 100))