recent_acc = np.mean(self.agent_memory['accuracy'])
else:
recent_reward = np.mean(
self.agent_memory['reward'][-self.recent_zone:])
recent_acc = np.mean(
self.agent_memory['accuracy'][-self.recent_zone:])
return recent_acc, recent_reward
#
if __name__ == '__main__':
# api = Baidu()
# api = FacePP()
api = AmazonRekognition()
rm = ResultManager('evaluation_results')
running_agent = RunningAgent( # dqn_path='evaluation_results/agent_DQN_train_baidu_DNIM.h5',
# dqn_path='evaluation_results/agent_DQN_train_face_imagenet.h5',
# dqn_path='evaluation_results/agent_DQN_train_face_DNIM.h5',
# dqn_path='evaluation_results/agent_DQN_retrain_face_DNIM.h5',
# dqn_path='evaluation_results/agent_DQN_retrain_baidu_DNIM.h5',
# dqn_path='evaluation_results/agent_DQN_initial_baidu_DNIM.h5',
dqn_path='evaluation_results/agent_DQN_initial_aws_DNIM.h5',
banchmark_q=75,
cloud_backend=api,
)
# imagenet_paths = _gen_sample_set_imagenet('/home/hsli/gnode02/imagenet-data/train/', 2)
imagenet_paths = _gen_sample_set_imagenet('/home/imagenet-data/train/', 2)
if len(self.agent_memory['reward']) < self.recent_zone:
recent_reward = np.mean(self.agent_memory['reward'])
recent_acc = np.mean(self.agent_memory['accuracy'])
else:
recent_reward = np.mean(
self.agent_memory['reward'][-self.recent_zone:])
recent_acc = np.mean(
self.agent_memory['accuracy'][-self.recent_zone:])
return recent_acc, recent_reward
#
if __name__ == '__main__':
api = Baidu()
rm = ResultManager('evaluation_results')
running_agent = RunningAgent(
dqn_path='evaluation_results/agent_DQN_train_baidu_imagenet.h5',
banchmark_q=75,
cloud_backend=api,
)
imagenet_paths = _gen_sample_set_imagenet(
'/home/hsli/gnode02/imagenet-data/train/', 3)[-500:]
test_image_paths = imagenet_paths
for idx, path in enumerate(test_image_paths):
error_code, log_dict = running_agent.agent_upload(path)
if error_code > 0: continue
if len(self.agent_memory['reward']) < self.recent_zone:
recent_reward = np.mean(self.agent_memory['reward'])
recent_acc = np.mean(self.agent_memory['accuracy'])
else:
recent_reward = np.mean(
self.agent_memory['reward'][-self.recent_zone:])
recent_acc = np.mean(
self.agent_memory['accuracy'][-self.recent_zone:])
return recent_acc, recent_reward
#
if __name__ == '__main__':
api = Baidu()
rm = ResultManager('evaluation_results')
running_agent = RunningAgent(
dqn_path='evaluation_results/agent_DQN_train_baidu_DNIM.h5',
banchmark_q=75,
cloud_backend=api,
)
# imagenet_paths = _gen_sample_set_imagenet('/home/hsli/gnode02/imagenet-data/train/', 2)
np.random.shuffle(subset)
test_image_paths = subset + rm.load(7)['image_path'][:1200] + rm.load(
7)['image_path']
for idx, path in enumerate(test_image_paths):
error_code, log_dict = running_agent.agent_upload(path)
np.random.seed(2)
BATCH_SIZE = 128
def list_split(l, size):
return [l[m:m + size] for m in range(0, len(l), size)]
if __name__ == '__main__':
images_dir = '/home/hsli/gnode02/imagenet-data/train/'
feature_extractor = load_model(
'checkpoints/mobilenetv2_predictor_2W_acc_0.6955_epoch50.hdf5')
rm = ResultManager('results')
agent_acc_size_dict = []
origin_acc_size_dict = []
agent = Q_Agent(s_dim=2,
a_dim=10,
epsilon_decay=0.9993,
epsilon_min=0.2,
lr=0.1,
gamma=0.95)
step_count = 0
env = BatchImgEnvironment(imagenet_train_path=images_dir,
samples_per_class=3,
backbone_model=InceptionV3(),
recent_reward = np.mean(self.agent_memory['reward'])
recent_acc = np.mean(self.agent_memory['accuracy'])
else:
recent_reward = np.mean(
self.agent_memory['reward'][-self.recent_zone:])
recent_acc = np.mean(
self.agent_memory['accuracy'][-self.recent_zone:])
return recent_acc, recent_reward
#
if __name__ == '__main__':
api = Baidu()
# api = FacePP()
rm = ResultManager('evaluation_results')
running_agent = RunningAgent( # dqn_path='evaluation_results/agent_DQN_train_baidu_DNIM.h5',
# dqn_path='evaluation_results/agent_DQN_train_face_imagenet.h5',
# dqn_path='evaluation_results/agent_DQN_train_face_DNIM.h5',
# dqn_path='evaluation_results/agent_DQN_retrain_face_DNIM.h5',
# dqn_path='evaluation_results/agent_DQN_retrain_baidu_DNIM.h5.retrain',
# dqn_path='evaluation_results/agent_DQN_initial_baidu_DNIM.h5',
# dqn_path='evaluation_results/agent_DQN_retrain_baidu_DNIM.h5.retrain.retrain_imagenet1',
# dqn_path='evaluation_results/agent_DQN_baidu_all_DNIM.h5',
# dqn_path='evaluation_results/agent_DQN_baidu_FLIR.h5',
dqn_path='evaluation_results/agent_DQN_train_1.6K_baidu.h5',
banchmark_q=75,
cloud_backend=api,
)
from keras.utils import multi_gpu_model
from keras.models import Model
from keras import regularizers
from keras.layers import Dense, GlobalAveragePooling2D, Dropout, BatchNormalization, Flatten
from keras.applications.mobilenetv2 import preprocess_input
from keras.callbacks import ModelCheckpoint, ReduceLROnPlateau, EarlyStopping
import tensorflow as tf
import cv2
import os
tf.set_random_seed(2)
np.random.seed(2)
os.environ["CUDA_VISIBLE_DEVICES"] = "1, 2"
rm = ResultManager('results')
train_size = 19000
test_size = 2000
train_img_paths = [
path.replace('gnode02/', '') for path in rm.load(6)['paths'][:train_size]
]
train_label_data = np.array([(0., 1.) if item > 3 else (1., 0.)
for item in rm.load(6)['robustness'][:train_size]
])
test_img_paths = [
path.replace('gnode02/', '') for path in rm.load(6)['paths'][-test_size:]
]
test_label_data = np.array([(0., 1.) if item > 3 else (1., 0.)
plt.subplot(514)
plt.plot(y[:, 3])
plt.ylabel('epsilon')
plt.subplot(515)
plt.plot(y[:, 4])
plt.ylabel('action')
plt.pause(0.0001)
if __name__ == '__main__':
images_dir = '/home/hsli/imagenet-data/train/'
feature_extractor = load_model(
'checkpoints/mobilenetv2_predictor_2W_acc_0.6955_epoch50.hdf5')
rm = ResultManager('results')
agent_acc_size_dict = []
origin_acc_size_dict = []
agent = Q_Agent(s_dim=2,
a_dim=10,
epsilon_decay=0.991,
epsilon_min=0.02,
lr=0.4,
gamma=0.92)
step_count = 0
env = EnvironmentAPI(imagenet_train_path=images_dir,
samples_per_class=37,
cloud_agent=FacePP())
from keras.applications import MobileNetV2
from Agents import PG_Agent, DDPG_Agent
from ImageCompressionEnvironment import BatchImgEnvironment
np.random.seed(2)
if __name__ == '__main__':
model = MobileNetV2()
model.compile(optimizer='adam',
loss='categorical_crossentropy',
metrics=['accuracy'])
images_dir = '/home/hsli/imagenet-data/train/'
rm = ResultManager('results')
agent_acc_size_dict = []
origin_acc_size_dict = []
agent = DDPG_Agent(a_dim=64, s_dim=64, train_batchsize=128)
epsilon = 0.8 # control exploration
epsilon_decay = 0.99
env = BatchImgEnvironment(imagenet_train_path=images_dir,
samples_per_class=1024,
step_batch_size=128,
deep_model=model)
try:
def estimate(self):
if len(self.agent_memory['reward']) < self.recent_zone:
recent_reward = np.mean(self.agent_memory['reward'])
recent_acc = np.mean(self.agent_memory['accuracy'])
else:
recent_reward = np.mean(self.agent_memory['reward'][-self.recent_zone:])
recent_acc = np.mean(self.agent_memory['accuracy'][-self.recent_zone:])
return recent_acc, recent_reward
#
if __name__ == '__main__':
api = Baidu()
# api = FacePP()
rm = ResultManager('result')
running_agent = RunningAgent(# dqn_path='evaluation_results/agent_DQN_train_baidu_DNIM.h5',
# dqn_path='evaluation_results/agent_DQN_train_face_imagenet.h5',
# dqn_path='evaluation_results/agent_DQN_train_face_DNIM.h5',
# dqn_path='evaluation_results/agent_DQN_retrain_face_DNIM.h5',
# dqn_path='evaluation_results/agent_DQN_retrain_baidu_DNIM.h5',
dqn_path='result/agent_DQN_train_baidu_imagenet.h5',
banchmark_q=75,
cloud_backend=api,
)
# imagenet_paths = _gen_sample_set_imagenet('/home/hsli/gnode02/imagenet-data/train/', 2)
# imagenet_paths = _gen_sample_set_imagenet('/home/imagenet-data/train/', 2)
with open("result/imagenet_baidu_ref2000.pkl", "rb") as file:
plt.subplot(515)
plt.plot(y[:, 4])
plt.ylabel('action')
plt.pause(0.0001)
if __name__ == '__main__':
# images_dir = '/home/hsli/gnode02/imagenet-data/train/'
images_dir = '/home/imagenet-data/train/'
feature_extractor = MobileNetV2(include_top=False)
x = feature_extractor.output
x = AveragePooling2D(pool_size=(4, 4))(x)
feature_extractor = Model(inputs=feature_extractor.input, outputs=x)
rm = ResultManager('evaluation_results')
agent_acc_size_dict = []
origin_acc_size_dict = []
agent = DQN_Agent(s_dim=1280,
a_dim=10,
epsilon_decay=0.99,
epsilon_min=0.02,
gamma=0.95,
replay_batchsize=256)
if EVALUATION:
agent.model = load_model(
'evaluation_results/agent_DQN_train_amazon_imagenet.h5'
) # If in evaluation phase, replace this with the actual pretrained model
agent.curr_exploration_rate = 0
from ImageCompressionEnvironment import BatchImgEnvironment
np.set_printoptions(precision=3)
tf.set_random_seed(2)
np.random.seed(2)
EVALUATION = True
if __name__ == '__main__':
images_dir = '/home/hsli/gnode02/imagenet-data/train/'
feature_extractor = load_model(
'checkpoints/mobilenetv2_predictor_2W_acc_0.6955_epoch50.hdf5')
rm = ResultManager('results')
agent_acc_size_dict = []
origin_acc_size_dict = []
agent = Q_Agent(s_dim=2,
a_dim=10,
epsilon_decay=0.9993,
epsilon_min=0.2,
lr=0.1,
gamma=0.95)
step_count = 0
env = BatchImgEnvironment(imagenet_train_path=images_dir,
samples_per_class=2,
backbone_model=InceptionV3(),
for image_name in np.random.choice(os.listdir(
"%s/%s" % (imagenet_data_path, img_class)),
size=samples_per_class):
sample_image_path = ("%s/%s/%s" %
(imagenet_data_path, img_class, image_name))
image_paths.append(sample_image_path)
return image_paths
def size_Q(img, Q):
f = BytesIO()
img.save(f, format="JPEG", quality=Q)
return len(f.getvalue())
rm = ResultManager('results')
sample_set = gen_sample_set('/home/hsli/imagenet-data/train/', 1)
aver_size_list = []
for i in np.arange(1, 101, 1):
print("Parsing %s" % i)
aver_size = [
size_Q(Image.open(path).convert("RGB"), int(i)) for path in sample_set
]
aver_size_list.append(aver_size)
rm.save(aver_size_list,
name='aver_size',
topic='measurements',
comment="finegrained average size upon different Q on 2W images")
sums = []
ids = []
curr_sum = 0
for idx, item in enumerate(l):
if item == 1:
curr_sum += 1
if curr_sum == len(l):
return len(l) - 1
else:
sums.append(curr_sum)
curr_sum = 0
ids.append(idx - 1)
return np.array(ids)[np.argmax(sums)]
rm = ResultManager('evaluation_results')
sample_paths, sample_labels = gen_sample_set(
'/home/hsli/gnode02/imagenet-data/train/', 5)
model_labels = to_categorical(sample_labels, 1000)
model = ResNet50()
model.compile('adam', 'categorical_crossentropy',
['top_k_categorical_accuracy'])
robust_dict = defaultdict(list)
measurement_dict = defaultdict(list)
banchmark = defaultdict(list)
for idx, path in enumerate(sample_paths):
if idx % 100 == 0:
print(time.strftime("%Y-%m-%d %H:%M:%S", time.localtime()), end='\t')
loc=8, ncol=3)
plt.subplots_adjust(top=1.0,
bottom=0.33,
left=0.105,
right=0.91,
hspace=0.2,
wspace=0.2)
plt.show()
x = np.linspace(0, 2 * np.pi, 50)
y = np.sin(x)
plt.plot(x, y)
plt.show()
rm = ResultManager('evaluation_results')
rm.print_meta_info()
not_reload = rm.load(16)
reload = rm.load(20)
initial_log = reload
# plot_keys = ['status', 'step_count', 'comp_size', 'upload_size']
# plot_keys = ['agent_epsilon', 'agent_accuracy', 'recent_reward', 'explor_rate']
plot_keys = ['agent_epsilon', 'recent_reward', 'explor_rate']
plot_durations(np.array([initial_log[key] for key in plot_keys]),
title_list=plot_keys)
plot_keys = ['status', 'recent_accuracy', 'accuracy', 'upload_size']
plot_durations(np.array([initial_log[key] for key in plot_keys]),
title_list=plot_keys)
# # evaluation_framework(not_reload)
evaluation_framework(rm.load(17))
plt.ylabel('epsilon')
plt.subplot(515)
plt.plot(y[:, 4])
plt.ylabel('action')
plt.pause(0.0001)
if __name__ == '__main__':
images_dir = '/home/hsli/gnode02/imagenet-data/train/'
feature_extractor = MobileNetV2(include_top=False)
x = feature_extractor.output
x = AveragePooling2D(pool_size=(4, 4))(x)
feature_extractor = Model(inputs=feature_extractor.input, outputs=x)
rm = ResultManager('evaluation_results')
agent_acc_size_dict = []
origin_acc_size_dict = []
agent = DQN_Agent(s_dim=1280,
a_dim=10,
epsilon_decay=0.99,
epsilon_min=0.02,
gamma=0.95,
replay_batchsize=256)
if EVALUATION:
agent.model = load_model('evaluation_results/agent_DQN_train_amazon_imagenet.h5') # If in evaluation phase, replace this with the actual pretrained model
agent.curr_exploration_rate = 0
step_count = 0