def load(self, session, model_path):
x_input = tf.placeholder(tf.float32, shape=(None, ) + self.x_shape)
x_input = tf.transpose(x_input, [0, 3, 1, 2])
with TowerContext('', is_training=False):
with tf.variable_scope('', reuse=tf.AUTO_REUSE):
_ = self.model.get_logits(x_input)
get_model_loader(model_path).init(session)
def run(self, nr_gpus):
"""
Do the 2 phase training, both phases have the same nr_epochs and lr_sched
Phase 1: finetune decoder portion (together with the first 7x7 in the encoder),
using the pretrained preact-resnet50 on ImageNet
Phase 2: unfreeze all layer, training whole, weigths taken from last epoch of
Phase 1
"""
def get_last_chkpt_path(phase1_dir):
stat_file_path = phase1_dir + '/stats.json'
with open(stat_file_path) as stat_file:
info = json.load(stat_file)
chkpt_list = [epoch_stat['global_step'] for epoch_stat in info]
last_chkpts_path = "%smodel-%d.index" % (phase1_dir,
max(chkpt_list))
return last_chkpts_path
save_dir = self.save_dir + '/base/'
self.train_batch_size = self.train_phase1_batch_size
init_weights = get_model_loader(self.pretrained_preact_resnet50_path)
self.run_once(nr_gpus, True, sess_init=init_weights, save_dir=save_dir)
save_dir = self.save_dir + '/tune/'
self.train_batch_size = self.train_phase2_batch_size
phase1_last_model_path = get_last_chkpt_path(self.save_dir + '/base/')
init_weights = SaverRestore(phase1_last_model_path,
ignore=['learning_rate'])
self.run_once(nr_gpus,
False,
sess_init=init_weights,
save_dir=save_dir)
return
def run(model):
instance = Model(model, model.conf.data_format)
if not model.conf.is_train:
batch = 64
dataset = get_data(model.conf.data_dir, 'val', batch)
eval_on_ILSVRC12(
instance,
get_model_loader(model.conf.logdir + '/' + model.conf.test_step),
dataset)
else:
logger.set_logger_dir(os.path.join(model.conf.logdir))
config = get_config(instance, model.conf)
if model.conf.reload_step:
config.session_init = get_model_loader(model.conf.logdir + '/' +
model.conf.reload_step)
trainer = SyncMultiGPUTrainerParameterServer(max(get_nr_gpu(), 1))
launch_train_with_config(config, trainer)
def do_visualize(model, model_path, nr_visualize=100, output_dir='output'):
"""
Visualize some intermediate results (proposals, raw predictions) inside the pipeline.
"""
df = get_train_dataflow()
df.reset_state()
pred = OfflinePredictor(
PredictConfig(model=model,
session_init=get_model_loader(model_path),
input_names=['image', 'gt_boxes', 'gt_labels'],
output_names=[
'generate_{}_proposals/boxes'.format(
'fpn' if cfg.MODE_FPN else 'rpn'),
'generate_{}_proposals/scores'.format(
'fpn' if cfg.MODE_FPN else 'rpn'),
'fastrcnn_all_scores',
'output/boxes',
'output/scores',
'output/labels',
]))
if os.path.isdir(output_dir):
shutil.rmtree(output_dir)
fs.mkdir_p(output_dir)
with tqdm.tqdm(total=nr_visualize) as pbar:
for idx, dp in itertools.islice(enumerate(df), nr_visualize):
img, gt_boxes, gt_labels = dp['image'], dp['gt_boxes'], dp[
'gt_labels']
rpn_boxes, rpn_scores, all_scores, \
final_boxes, final_scores, final_labels = pred(img, gt_boxes, gt_labels)
# draw groundtruth boxes
gt_viz = draw_annotation(img, gt_boxes, gt_labels)
# draw best proposals for each groundtruth, to show recall
proposal_viz, good_proposals_ind = draw_proposal_recall(
img, rpn_boxes, rpn_scores, gt_boxes)
# draw the scores for the above proposals
score_viz = draw_predictions(img, rpn_boxes[good_proposals_ind],
all_scores[good_proposals_ind])
results = [
DetectionResult(*args)
for args in zip(final_boxes, final_scores, final_labels,
[None] * len(final_labels))
]
final_viz = draw_final_outputs(img, results)
viz = tpviz.stack_patches(
[gt_viz, proposal_viz, score_viz, final_viz], 2, 2)
if os.environ.get('DISPLAY', None):
tpviz.interactive_imshow(viz)
cv2.imwrite("{}/{:03d}.png".format(output_dir, idx), viz)
pbar.update()
def run(self):
def get_last_chkpt_path(prev_phase_dir):
stat_file_path = os.path.join(prev_phase_dir, "stats.json")
with open(stat_file_path) as stat_file:
info = json.load(stat_file)
chkpt_list = [epoch_stat["global_step"] for epoch_stat in info]
last_chkpts_path = os.path.join(
prev_phase_dir, "model-{}.index".format(max(chkpt_list))
)
return last_chkpts_path
phase_opts = self.training_phase
if len(phase_opts) > 1:
for idx, opt in enumerate(phase_opts):
random.seed(self.seed)
np.random.seed(self.seed)
tf.random.set_random_seed(self.seed)
log_dir = os.path.join(self.save_dir, str(idx))
pretrained_path = opt["pretrained_path"]
if pretrained_path == -1:
pretrained_path = get_last_chkpt_path(prev_log_dir)
init_weights = SaverRestore(
pretrained_path, ignore=["learning_rate"]
)
elif pretrained_path is not None:
init_weights = get_model_loader(pretrained_path)
self.run_once(opt, sess_init=init_weights, save_dir=log_dir)
prev_log_dir = log_dir
else:
random.seed(self.seed)
np.random.seed(self.seed)
tf.random.set_random_seed(self.seed)
opt = phase_opts[0]
init_weights = None
if "pretrained_path" in opt:
assert opt["pretrained_path"] != -1
init_weights = get_model_loader(opt["pretrained_path"])
self.run_once(opt, sess_init=init_weights, save_dir=self.save_dir)
return
def run(self):
def get_last_chkpt_path(prev_phase_dir):
stat_file_path = prev_phase_dir + "/stats.json"
with open(stat_file_path) as stat_file:
info = json.load(stat_file)
chkpt_list = [epoch_stat["global_step"] for epoch_stat in info]
last_chkpts_path = "%smodel-%d.index" % (prev_phase_dir,
max(chkpt_list))
return last_chkpts_path
phase_opts = self.training_phase
if len(phase_opts) > 1:
for idx, opt in enumerate(phase_opts):
log_dir = "%s/%02d" % (self.save_dir, idx)
if opt["pretrained_path"] == -1:
pretrained_path = get_last_chkpt_path(prev_log_dir)
init_weights = SaverRestore(pretrained_path,
ignore=["learning_rate"])
elif opt["pretrained_path"] is not None:
init_weights = get_model_loader(pretrained_path)
self.run_once(opt,
sess_init=init_weights,
save_dir=log_dir + "/")
prev_log_dir = log_dir
else:
opt = phase_opts[0]
if "pretrained_path" in opt:
if opt["pretrained_path"] == None:
init_weights = None
elif opt["pretrained_path"] == -1:
log_dir_prev = "%s" % self.save_dir
pretrained_path = get_last_chkpt_path(log_dir_prev)
init_weights = SaverRestore(pretrained_path,
ignore=["learning_rate"])
else:
init_weights = get_model_loader(opt["pretrained_path"])
self.run_once(opt, sess_init=init_weights, save_dir=self.save_dir)
return
def load(self, **kwargs):
session = kwargs["session"]
assert isinstance(session, tf.Session)
x_input = tf.placeholder(tf.float32, shape=(None, ) + self.x_shape)
x_input = tf.transpose(x_input, [0, 3, 1, 2])
with TowerContext('', is_training=False):
with tf.variable_scope('', reuse=tf.AUTO_REUSE):
logits = self.model.get_logits(x_input)
model_path = get_model_path('R152-Denoise.npz')
url = 'https://github.com/facebookresearch/ImageNet-Adversarial-Training/releases/download/v0.1/R152-Denoise.npz'
if not os.path.exists(model_path):
if not os.path.exists(os.path.dirname(model_path)):
os.makedirs(os.path.dirname(model_path))
from six.moves import urllib
urllib.request.urlretrieve(url, model_path, show_progress)
get_model_loader(model_path).init(session)
def run(self):
def get_last_chkpt_path(prev_phase_dir):
stat_file_path = prev_phase_dir + '/stats.json'
with open(stat_file_path) as stat_file:
info = json.load(stat_file)
chkpt_list = [epoch_stat['global_step'] for epoch_stat in info]
last_chkpts_path = "%smodel-%d.index" % (prev_phase_dir,
max(chkpt_list))
return last_chkpts_path
phase_opts = self.training_phase
if len(phase_opts) > 1:
for idx, opt in enumerate(phase_opts):
random.seed(self.seed)
np.random.seed(self.seed)
tf.random.set_random_seed(self.seed)
log_dir = '%s/%02d/' % (self.save_dir, idx)
pretrained_path = opt['pretrained_path']
if pretrained_path == -1:
pretrained_path = get_last_chkpt_path(prev_log_dir)
init_weights = SaverRestore(pretrained_path,
ignore=['learning_rate'])
elif pretrained_path is not None:
init_weights = get_model_loader(pretrained_path)
self.run_once(opt, sess_init=init_weights, save_dir=log_dir)
prev_log_dir = log_dir
else:
random.seed(self.seed)
np.random.seed(self.seed)
tf.random.set_random_seed(self.seed)
opt = phase_opts[0]
init_weights = None
if 'pretrained_path' in opt:
assert opt['pretrained_path'] != -1
init_weights = get_model_loader(opt['pretrained_path'])
self.run_once(opt, sess_init=init_weights, save_dir=self.save_dir)
return
def main(_):
# Images for inception classifier are normalized to be in [-1, 1] interval,
# eps is a difference between pixels so it should be in [0, 2] interval.
# Renormalizing epsilon from [0, 255] to [0, 2].
full_start = timer()
eps = 2.0 * FLAGS.max_epsilon / 255.0
tf.logging.set_verbosity(tf.logging.INFO)
with tf.Graph().as_default():
# Prepare graph
x_raw = tf.placeholder(tf.float32, shape=batch_shape)
x_max = tf.clip_by_value(x_raw + eps, -1.0, 1.0)
x_min = tf.clip_by_value(x_raw - eps, -1.0, 1.0)
x_input = tf.placeholder(tf.float32, shape=batch_shape)
target_class_input = tf.placeholder(tf.int32, shape=[FLAGS.batch_size])
raw_class_input = tf.placeholder(tf.int32, shape=[FLAGS.batch_size])
i = tf.constant(0)
grad = tf.zeros(shape=batch_shape)
adv_images = np.zeros(shape=batch_shape)
x_adv1, _, _, _, _, _, _, = tf.while_loop(stop, graph, [
x_input, target_class_input, raw_class_input, i, x_max, x_min, grad
])
print('Created Graph')
# Run computation
with tf.Session() as sess:
processed = 0.0
s1 = tf.train.Saver(slim.get_model_variables(scope='resnet_v2_50'))
s1.restore(sess, FLAGS.checkpoint_path_resnet_v2)
get_model_loader(args1.load).init(sess)
#如果有allow_pickle的错误，请修改tensorpack的库，使得里面dict(np.load(path,allow_pickle=True))
get_model_loader(args2.load).init(sess)
get_model_loader(args3.load).init(sess)
print('Initialized Models')
for filenames, image_mix, image_raw in load_images(
FLAGS.input_dir, batch_shape):
target_class_for_batch = (
[all_images_taget_class[n] for n in filenames] + [0] *
(FLAGS.batch_size - len(filenames)))
raw_class_for_batch = (
[all_images_raw_class[n] for n in filenames] + [0] *
(FLAGS.batch_size - len(filenames)))
adv_images = sess.run(x_adv1,
feed_dict={
x_input: image_mix,
target_class_input:
target_class_for_batch,
x_raw: image_raw,
raw_class_input: raw_class_for_batch
})
save_images(adv_images, filenames, FLAGS.output_dir)
processed += FLAGS.batch_size
full_end = timer()
print("DONE: Processed {} images in {} sec".format(
processed, full_end - full_start))
def prepare_model(model_name, classes, use_pretrained,
pretrained_model_file_path):
kwargs = {'pretrained': use_pretrained, 'classes': classes}
net_lambda, net_file_path = get_model(model_name, **kwargs)
net = ImageNetModel(model_lambda=net_lambda)
if use_pretrained and not pretrained_model_file_path:
pretrained_model_file_path = net_file_path
inputs_desc = None
if pretrained_model_file_path:
assert (os.path.isfile(pretrained_model_file_path))
logging.info('Loading model: {}'.format(pretrained_model_file_path))
inputs_desc = get_model_loader(pretrained_model_file_path)
return net, inputs_desc
def export(params, model_path, export_type, output_dir):
if not os.path.exists(output_dir):
os.makedirs(output_dir)
model_params = params['model']
model = BaseVQVAE.from_params(model_params)
pred_config = PredictConfig(
session_init=get_model_loader(model_path),
model=model,
input_names=['input'],
output_names=['x_recon', 'embeddings', 'latent_zq'])
if export_type == 'compact':
checkpoint_name = os.path.split(model_path)[1]
ModelExporter(pred_config).export_compact(
os.path.join(output_dir, 'model.pb'))
else:
ModelExporter(pred_config).export_serving(
os.path.join(output_dir, 'exported'))
def prepare_model(model_name, use_pretrained, pretrained_model_file_path):
kwargs = {'pretrained': use_pretrained}
raw_net = get_model(model_name, **kwargs)
input_image_size = raw_net.in_size[0] if hasattr(raw_net,
'in_size') else 224
net = ImageNetModel(model_lambda=raw_net, image_size=input_image_size)
if use_pretrained and not pretrained_model_file_path:
pretrained_model_file_path = raw_net.file_path
inputs_desc = None
if pretrained_model_file_path:
assert (os.path.isfile(pretrained_model_file_path))
logging.info('Loading model: {}'.format(pretrained_model_file_path))
inputs_desc = get_model_loader(pretrained_model_file_path)
return net, inputs_desc
def create():
# fetch weights
weights_path = zoo.fetch_weights(
'https://github.com/facebookresearch/ImageNet-Adversarial-Training/releases/download/v0.1/R152-Denoise.npz',
unzip=False)
# model constructer expects an ArgumentParser as argument
parser = argparse.ArgumentParser()
parser.add_argument('-d',
'--depth',
help='ResNet depth',
type=int,
default=152,
choices=[50, 101, 152])
parser.add_argument('--arch',
help='Name of architectures defined in nets.py',
default='ResNetDenoise')
args = parser.parse_args([])
model = getattr(nets, args.arch + 'Model')(args)
image = tf.placeholder(tf.float32, shape=(None, 3, 224, 224))
with TowerContext(tower_name='', is_training=False):
logits = model.get_logits(image)
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
session = tf.Session(config=config)
with session.as_default():
model = get_model_loader(weights_path).init(session)
fmodel = TensorFlowModel(image,
logits,
channel_axis=1,
bounds=[0, 255.],
preprocessing=(127.5, 127.5))
return fmodel
def prepare_model(model_name,
use_pretrained,
pretrained_model_file_path,
data_format="channels_last"):
kwargs = {"pretrained": use_pretrained}
raw_net = get_model(name=model_name, data_format=data_format, **kwargs)
input_image_size = raw_net.in_size[0] if hasattr(raw_net,
"in_size") else 224
net = ImageNetModel(model_lambda=raw_net,
image_size=input_image_size,
data_format=data_format)
if use_pretrained and not pretrained_model_file_path:
pretrained_model_file_path = raw_net.file_path
inputs_desc = None
if pretrained_model_file_path:
assert (os.path.isfile(pretrained_model_file_path))
logging.info("Loading model: {}".format(pretrained_model_file_path))
inputs_desc = get_model_loader(pretrained_model_file_path)
return net, inputs_desc
def gen_pred_config(self):
if self.inf_auto_find_chkpt:
self.inf_model_path = os.path.join(
self.save_dir,
str(
max([
int(x) for x in [
name for name in os.listdir(self.save_dir)
if os.path.isdir(os.path.join(self.save_dir, name))
]
])),
)
print(f"Inference model path: <{self.inf_model_path}>")
print(
'-----Auto Selecting Checkpoint Basing On "%s" Through "%s" Comparison'
% (self.inf_auto_metric, self.inf_auto_comparator))
model_path, stat = get_best_chkpts(self.inf_model_path,
self.inf_auto_metric,
self.inf_auto_comparator)
print("Selecting: %s" % model_path)
print("Having Following Statistics:")
for key, value in stat.items():
print("\t%s: %s" % (key, value))
sess = get_model_loader(model_path)
else:
model_path = self.inf_model_path
sess = SaverRestoreRelaxed(self.inf_model_path)
model_constructor = self.get_model()
pred_config = PredictConfig(
model=model_constructor(),
session_init=sess,
input_names=self.eval_inf_input_tensor_names,
output_names=self.eval_inf_output_tensor_names,
)
return pred_config
with tf.Graph().as_default() as G:
if args.config:
logger.warn(
"Using a config script is not reliable. Please use metagraph.")
MODEL = imp.load_source('config_script', args.config).Model
M = MODEL()
with TowerContext('', is_training=False):
input = PlaceholderInput()
input.setup(M.get_inputs_desc())
M.build_graph(input)
else:
tf.train.import_meta_graph(args.meta)
# loading...
init = get_model_loader(args.model)
sess = tf.Session(config=tf.ConfigProto(allow_soft_placement=True))
sess.run(tf.global_variables_initializer())
sess.run(tf.local_variables_initializer())
init.init(sess)
# dump ...
with sess.as_default():
if args.output.endswith('npy') or args.output.endswith('npz'):
varmanip.dump_session_params(args.output)
else:
var = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES)
var.extend(tf.get_collection(tf.GraphKeys.MODEL_VARIABLES))
gvars = set([k.name for k in tf.global_variables()])
var = [v for v in var if v.name in gvars]
var_dict = {}
"""
Sec 2.3: We keep the per-worker sample size n constant when we change the number of workers k.
In this work, we use n = 32 which has performed well for a wide range of datasets and networks.
"""
parser.add_argument('--batch', help='per-GPU batch size', default=32, type=int)
args = parser.parse_args()
model = Model(args.depth, args.norm, args.use_ws)
model.accum_grad = args.accum_grad
if args.weight_decay_norm:
model.weight_decay_pattern = ".*/W|.*/gamma|.*/beta"
if args.eval:
batch = 128 # something that can run on one gpu
ds = get_val_dataflow(args.data, batch, fbresnet_augmentor(False))
eval_classification(model, get_model_loader(args.load), ds)
sys.exit()
logger.info("Training on {}".format(socket.gethostname()))
# Print some information for sanity check.
os.system("nvidia-smi")
assert args.load is None
hvd.init()
if args.logdir is None:
args.logdir = os.path.join('train_log', 'Horovod-{}GPUs-{}Batch'.format(hvd.size(), args.batch))
if hvd.rank() == 0:
logger.set_logger_dir(args.logdir, 'd')
logger.info("Rank={}, Local Rank={}, Size={}".format(hvd.rank(), hvd.local_rank(), hvd.size()))
from tensorpack import logger
from tensorpack.tfutils import varmanip, get_model_loader
if __name__ == '__main__':
parser = argparse.ArgumentParser(
description='Keep only TRAINABLE and MODEL variables in a checkpoint.')
parser.add_argument('--meta', help='metagraph file', required=True)
parser.add_argument(dest='input', help='input model file, has to be a TF checkpoint')
parser.add_argument(dest='output', help='output model file, can be npz or TF checkpoint')
args = parser.parse_args()
tf.train.import_meta_graph(args.meta, clear_devices=True)
# loading...
init = get_model_loader(args.input)
sess = tf.Session(config=tf.ConfigProto(allow_soft_placement=True))
sess.run(tf.global_variables_initializer())
sess.run(tf.local_variables_initializer())
init.init(sess)
# dump ...
with sess.as_default():
if args.output.endswith('npy') or args.output.endswith('npz'):
varmanip.dump_session_params(args.output)
else:
var = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES)
var.extend(tf.get_collection(tf.GraphKeys.MODEL_VARIABLES))
gvars = set([k.name for k in tf.global_variables()])
var = [v for v in var if v.name in gvars]
var_dict = {}
parser.add_argument('--eval', action='store_true')
parser.add_argument('--flops', action='store_true', help='print flops and exit')
args = parser.parse_args()
if args.gpu:
os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu
if args.v2 and args.group != parser.get_default('group'):
logger.error("group= is not used in ShuffleNetV2!")
model = Model()
if args.eval:
batch = 128 # something that can run on one gpu
ds = get_data('val', batch)
eval_on_ILSVRC12(model, get_model_loader(args.load), ds)
elif args.flops:
# manually build the graph with batch=1
input_desc = [
InputDesc(tf.float32, [1, 224, 224, 3], 'input'),
InputDesc(tf.int32, [1], 'label')
]
input = PlaceholderInput()
input.setup(input_desc)
with TowerContext('', is_training=False):
model.build_graph(*input.get_input_tensors())
model_utils.describe_trainable_vars()
tf.profiler.profile(
tf.get_default_graph(),
cmd='op',
# CAM_viz or eval or train
if args.CAM_viz != None:
batch = 128
# @ 20171125
if args.CAM_viz == 'train':
assert args.crop_method_TR not in ['CAMCrop', 'CAMCropR'], '*** CAMCrop or CAMCropR IS NOT ALLOWED IN CAM_viz ***'
# @ 20171122 : visualize the CAM for better understanding of the CNN
# introduce strict_order to make it convenient for image saving process
# @ 20171124 : introduce remainder_TR to extract all the CAM_viz of images in the training set
ds = get_data(args.CAM_viz, args.data, batch, args.crop_method_TS if args.CAM_viz == 'val' else args.crop_method_TR, \
repeat_times = 1, strict_order = True, remainder_TR = True if args.CAM_viz == 'train' else False)
viz_CAM(model, get_model_loader(args.load), args.CAM_viz, ds, args.CAM_dir,
save_PKL = args.CAM_viz_PKL, save_REP = args.CAM_viz_REP)
elif args.eval:
batch = 128 # something that can run on one gpu
# @ 20171121 to make it convenient to calculate average estimation
batch = int(batch // args.repeat_times * args.repeat_times)
ds = get_data('val', args.data, batch, args.crop_method_TS, repeat_times = args.repeat_times, strict_order = True)
eval_on_AVA2012(model, get_model_loader(args.load), ds, args.repeat_times)
else:
# add @ 20171128: the strategy of parameter initalization with a ImageNet pre-trained model
# should be recorded within the name string of the directory of training log
def main(_):
# Images for inception classifier are normalized to be in [-1, 1] interval,
# eps is a difference between pixels so it should be in [0, 2] interval.
# Renormalizing epsilon from [0, 255] to [0, 2].
full_start = timer()
eps = 2.0 * FLAGS.max_epsilon / 255.0
batch_shape = [FLAGS.batch_size, FLAGS.image_height, FLAGS.image_width, 3]
tf.logging.set_verbosity(tf.logging.INFO)
all_images_taget_class, all_images_true_label = load_target_class(FLAGS.input_dir)
with tf.Graph().as_default():
# Prepare graph
x_input = tf.placeholder(tf.float32, shape=batch_shape)
x_max = tf.clip_by_value(x_input + eps, -1.0, 1.0)
x_min = tf.clip_by_value(x_input - eps, -1.0, 1.0)
target_class_input = tf.placeholder(tf.int32, shape=[FLAGS.batch_size])
true_label_input = tf.placeholder(tf.int32, shape=[FLAGS.batch_size])
i = tf.constant(0)
grad = tf.zeros(shape=batch_shape)
x_adv, _, _, _, _, _, _ = tf.while_loop(stop, graph, [x_input, target_class_input,true_label_input, i, x_max, x_min, grad])
s1 = tf.train.Saver(slim.get_model_variables(scope='AdvInceptionResnetV2'))
s2 = tf.train.Saver(slim.get_model_variables(scope='Ens3AdvInceptionV3'))
s3 = tf.train.Saver(slim.get_model_variables(scope='Ens4AdvInceptionV3'))
s4 = tf.train.Saver(slim.get_model_variables(scope='EnsAdvInceptionResnetV2'))
s5 = tf.train.Saver(slim.get_model_variables(scope='AdvInceptionV3'))
s6 = tf.train.Saver(slim.get_model_variables(scope='InceptionV3'))
s7 = tf.train.Saver(slim.get_model_variables(scope='resnet_v2_101'))
s8 = tf.train.Saver(slim.get_model_variables(scope='resnet_v2_50'))
#s4 = tf.train.Saver(slim.get_model_variables(scope='EnsAdvInceptionResnetV2'))
model = resnext.ResNeXtDenoiseAllModel()
print('Created Graph')
#byheziwen
'''x_input1 = tf.placeholder(tf.float32, shape=batch_shape)
x_input2 = tf.placeholder(tf.float32, shape=batch_shape)
x_all_16 = all_16(x_input1, x_input2)'''
# Run computation
with tf.Session() as sess:
processed = 0.0
s1.restore(sess, FLAGS.checkpoint_path_adv_inception_resnet_v2)
s2.restore(sess, FLAGS.checkpoint_path_ens3_adv_inception_v3)
s3.restore(sess, FLAGS.checkpoint_path_ens4_adv_inception_v3)
s4.restore(sess, FLAGS.checkpoint_path_ens_adv_inception_resnet_v2)
s5.restore(sess, FLAGS.checkpoint_path_adv_inception_v3)
s6.restore(sess, FLAGS.checkpoint_path_inception_v3)
s7.restore(sess, FLAGS.checkpoint_path_resnet_v2_101)
s8.restore(sess, FLAGS.checkpoint_path_resnet_v2_50)
#s4.restore(sess, FLAGS.checkpoint_path_ens_adv_inception_resnet_v2)
model = get_model_loader(FLAGS.checkpoint_path_resnext_101).init(sess)
print('Initialized Models')
for filenames, images in load_images(FLAGS.input_dir, batch_shape):
target_class_for_batch = (
[all_images_taget_class[n] for n in filenames]
+ [0] * (FLAGS.batch_size - len(filenames)))
true_label_for_batch = (
[all_images_true_label[n] for n in filenames]
+ [0] * (FLAGS.batch_size - len(filenames)))
adv_images = sess.run(x_adv, feed_dict={x_input: images, target_class_input: target_class_for_batch, true_label_input:true_label_for_batch})
#by heziwen
#adv_images = sess.run(x_all_16, feed_dict={x_input1:adv_images, x_input2:images})
save_images(adv_images, filenames, FLAGS.output_dir)
processed += FLAGS.batch_size
full_end = timer()
print("DONE: Processed {} images in {} sec".format(processed, full_end - full_start))
from tensorpack import logger
from tensorpack.tfutils import varmanip, get_model_loader
if __name__ == '__main__':
parser = argparse.ArgumentParser(
description='Keep only TRAINABLE and MODEL variables in a checkpoint.')
parser.add_argument('--meta', help='metagraph file', required=True)
parser.add_argument(dest='input', help='input model file, has to be a TF checkpoint')
parser.add_argument(dest='output', help='output model file, can be npz or TF checkpoint')
args = parser.parse_args()
tf.train.import_meta_graph(args.meta)
# loading...
init = get_model_loader(args.input)
sess = tf.Session(config=tf.ConfigProto(allow_soft_placement=True))
sess.run(tf.global_variables_initializer())
sess.run(tf.local_variables_initializer())
init.init(sess)
# dump ...
with sess.as_default():
if args.output.endswith('npy') or args.output.endswith('npz'):
varmanip.dump_session_params(args.output)
else:
var = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES)
var.extend(tf.get_collection(tf.GraphKeys.MODEL_VARIABLES))
gvars = set([k.name for k in tf.global_variables()])
var = [v for v in var if v.name in gvars]
var_dict = {}
# create GhostNet
from ghost_net import GhostNet
model = GhostNet(width=args.width_ratio, se=args.se,
weight_decay=args.weight_decay,
dw_code=dw_code, ratio_code=ratio_code,
label_smoothing=args.label_smoothing)
model.data_format = args.data_format
print('model created')
# start evaluation
if args.eval:
batch = 256 # something that can run on your gpu
ds = get_data('val', batch)
start = time()
eval_on_ILSVRC12(model, get_model_loader(args.load), ds)
stop = time()
print('Evaluation used time: %.2fs.' % (stop-start))
elif args.flops > 0:
# manually build the graph with batch=1
image_shape = 224
input_desc = [
InputDesc(tf.float32, [1, image_shape, image_shape, 3], 'input'),
InputDesc(tf.int32, [1], 'label')
]
input = PlaceholderInput()
input.setup(input_desc)
with TowerContext('', is_training=False):
model.build_graph(*input.get_input_tensors())
model_utils.describe_trainable_vars()
def eval_retrieval(evalargs):
# Set up evaluation:
save_dir = evalargs.save_dir
mkdir_p(evalargs.save_dir)
model_configs = get_model_config(evalargs.ModelPath)
# Set up graph:
pred_config = PredictConfig(
model=DH3D(model_configs),
session_init=get_model_loader(evalargs.ModelPath),
input_names=['pointclouds'],
output_names=['globaldesc'], # ['globaldesc'], output_weights
)
predictor = OfflinePredictor(pred_config)
# Data:
df, totalbatch = get_eval_global_testdata(model_configs,
evalargs.data_path,
evalargs.ref_gt_file)
# Predict:
pcdnum = 0
for [pcds, names] in df: # pcds is a list, batchsize x numpts x 3
batch = pcds.shape[0]
if totalbatch > batch:
numpts = pcds.shape[1]
pcddim = pcds.shape[2]
padzeros = np.zeros([totalbatch - batch, numpts, pcddim],
dtype=np.float32)
pcds = np.vstack([pcds, padzeros])
results = predictor(pcds)
global_feats = results[0]
for i in range(batch):
pcdnum += 1
globaldesc = global_feats[i]
name = names[i]
savename = os.path.join(evalargs.save_dir, name)
basedir = os.path.dirname(savename)
mkdir_p(basedir)
globaldesc.tofile(savename)
print('predicted {} poitnclouds \n'.format(pcdnum))
# Evaluation recall:
if evalargs.eval_recall:
evaluator = GlobalDesc_eval(
result_savedir='./',
desc_dir=save_dir,
database_file=os.path.join(evalargs.data_path,
evalargs.ref_gt_file),
query_file=os.path.join(evalargs.data_path, evalargs.qry_gt_file),
max_num_nn=25)
evaluator.evaluate()
print("evaluation finished!\n")
if evalargs.delete_tmp:
# delete all the descriptors
descdirs = [os.path.join(save_dir, f) for f in os.listdir(save_dir)]
descdirs = [d for d in descdirs if os.path.isdir(d)]
for d in descdirs:
shutil.rmtree(d)
if not tf.test.is_gpu_available():
from tensorflow.python.framework import test_util
assert get_tf_version_tuple() >= (1, 7) and test_util.IsMklEnabled(), \
"Inference requires either GPU support or MKL support!"
assert args.load
finalize_configs(is_training=False)
if args.predict or args.visualize:
cfg.TEST.RESULT_SCORE_THRESH = cfg.TEST.RESULT_SCORE_THRESH_VIS
if args.visualize:
do_visualize(MODEL, args.load)
else:
predcfg = PredictConfig(
model=MODEL,
session_init=get_model_loader(args.load),
input_names=MODEL.get_inference_tensor_names()[0],
output_names=MODEL.get_inference_tensor_names()[1])
if args.compact:
ModelExporter(predcfg).export_compact(args.compact, optimize=False)
elif args.serving:
ModelExporter(predcfg).export_serving(args.serving, optimize=False)
if args.predict:
predictor = OfflinePredictor(predcfg)
for image_file in args.predict:
do_predict(predictor, image_file)
elif args.evaluate:
assert args.evaluate.endswith('.json'), args.evaluate
do_evaluate(predcfg, args.evaluate)
parser.add_argument('--mode',
choices=['resnet', 'preact', 'se'],
help='variants of resnet to use',
default='resnet')
args = parser.parse_args()
if args.gpu:
os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu
model = Model(args.depth, args.mode)
model.data_format = args.data_format
if args.eval:
batch = 1 # something that can run on one gpu
ds = get_data('val', batch)
eval_on_iNaturalist(model, get_model_loader(args.load), ds)
print('eval Done!')
elif args.test:
batch = 1 # something that can run on one gpu
ds = get_data('test', batch)
test_on_iNaturalist(model, get_model_loader(args.load), ds)
else:
if args.fake:
logger.set_logger_dir(os.path.join('train_log', 'tmp'), 'd')
else:
logger.set_logger_dir(
os.path.join(
'train_log',
'iNaturalist-{}-d{}'.format(args.mode, args.depth)))
parser.add_argument('--batch', default=256, type=int,
help='total batch size. 32 per GPU gives best accuracy, higher values should be similarly good')
parser.add_argument('--mode', choices=['resnet', 'preact', 'se'],
help='variants of resnet to use', default='resnet')
args = parser.parse_args()
if args.gpu:
os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu
model = Model(args.depth, args.mode)
model.data_format = args.data_format
if args.eval:
batch = 1 # something that can run on one gpu
ds = get_data('val', batch)
eval_on_iNaturalist(model, get_model_loader(args.load), ds)
print('eval Done!')
elif args.test:
batch = 1 # something that can run on one gpu
ds = get_data('test', batch)
test_on_iNaturalist(model, get_model_loader(args.load), ds)
else:
if args.fake:
logger.set_logger_dir(os.path.join('train_log', 'tmp'), 'd')
else:
logger.set_logger_dir(
os.path.join('train_log', 'iNaturalist-{}-d{}'.format(args.mode, args.depth)))
config = get_config(model, fake=args.fake)
"Note that it's best to keep per-GPU batch size in [32, 64] to obtain the best accuracy."
"Pretrained models listed in README were trained with batch=32x8.")
parser.add_argument('--mode',
choices=['resnet', 'preact', 'se'],
help='variants of resnet to use',
default='resnet')
args = parser.parse_args()
if args.gpu:
os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu
model = Model(args.depth, args.mode)
model.data_format = args.data_format
if args.eval:
batch = 128 # something that can run on one gpu
ds = get_data('val', batch)
eval_on_ILSVRC12(model, get_model_loader(args.load), ds)
else:
if args.fake:
logger.set_logger_dir(os.path.join('train_log', 'tmp'), 'd')
else:
logger.set_logger_dir(
os.path.join('train_log',
'imagenet-{}-d{}'.format(args.mode, args.depth)))
config = get_config(model, fake=args.fake)
if args.load:
config.session_init = get_model_loader(args.load)
trainer = SyncMultiGPUTrainerReplicated(max(get_nr_gpu(), 1))
launch_train_with_config(config, trainer)
def do_train(model):
batch = args.batch
total_batch = batch * hvd.size()
if args.fake:
data = FakeData([[batch, 224, 224, 3], [batch]],
1000,
random=False,
dtype=['uint8', 'int32'])
data = StagingInput(QueueInput(data))
callbacks = []
steps_per_epoch = 50
else:
logger.info("#Tower: {}; Batch size per tower: {}".format(
hvd.size(), batch))
zmq_addr = 'ipc://@imagenet-train-b{}'.format(batch)
if args.no_zmq_ops:
dataflow = RemoteDataZMQ(zmq_addr, hwm=150, bind=False)
data = QueueInput(dataflow)
else:
data = ZMQInput(zmq_addr, 30, bind=False)
data = StagingInput(data)
steps_per_epoch = int(np.round(1281167 / total_batch))
BASE_LR = 0.1 * (total_batch // 256)
"""
ImageNet in 1 Hour, Sec 2.1:
Linear Scaling Rule: When the minibatch size is
multiplied by k, multiply the learning rate by k.
"""
logger.info("Base LR: {}".format(BASE_LR))
callbacks = [
ModelSaver(max_to_keep=10),
EstimatedTimeLeft(),
ScheduledHyperParamSetter('learning_rate', [(0, BASE_LR),
(35, BASE_LR * 1e-1),
(70, BASE_LR * 1e-2),
(95, BASE_LR * 1e-3)])
]
"""
Feature Denoising, Sec 5:
Our models are trained for a total of
110 epochs; we decrease the learning rate by 10× at the 35-
th, 70-th, and 95-th epoch
"""
max_epoch = 110
if BASE_LR > 0.1:
callbacks.append(
ScheduledHyperParamSetter('learning_rate',
[(0, 0.1),
(5 * steps_per_epoch, BASE_LR)],
interp='linear',
step_based=True))
"""
ImageNet in 1 Hour, Sec 2.2:
we start from a learning rate of η and increment it by a constant amount at
each iteration such that it reaches ηˆ = kη after 5 epochs
"""
if not args.fake:
# add distributed evaluation, for various attackers that we care.
def add_eval_callback(name, attacker, condition):
cb = create_eval_callback(
name,
model.get_inference_func(attacker),
# always eval in the last 2 epochs no matter what
lambda epoch_num: condition(epoch_num) or epoch_num > max_epoch
- 2)
callbacks.append(cb)
add_eval_callback('eval-clean', NoOpAttacker(), lambda e: True)
add_eval_callback(
'eval-10step',
PGDAttacker(10, args.attack_epsilon, args.attack_step_size),
lambda e: True)
add_eval_callback(
'eval-50step',
PGDAttacker(50, args.attack_epsilon, args.attack_step_size),
lambda e: e % 20 == 0)
add_eval_callback(
'eval-100step',
PGDAttacker(100, args.attack_epsilon, args.attack_step_size),
lambda e: e % 10 == 0 or e > max_epoch - 5)
for k in [20, 30, 40, 60, 70, 80, 90]:
add_eval_callback(
'eval-{}step'.format(k),
PGDAttacker(k, args.attack_epsilon, args.attack_step_size),
lambda e: False)
trainer = HorovodTrainer(average=True)
trainer.setup_graph(model.get_inputs_desc(), data, model.build_graph,
model.get_optimizer)
trainer.train_with_defaults(callbacks=callbacks,
steps_per_epoch=steps_per_epoch,
session_init=get_model_loader(args.load)
if args.load is not None else None,
max_epoch=max_epoch,
starting_epoch=args.starting_epoch)
if args.v2 and args.group != parser.get_default('group'):
logger.error("group= is not used in ShuffleNetV2!")
if args.batch != 1024:
logger.warn(
"Total batch size != 1024, you need to change other hyperparameters to get the same results."
)
TOTAL_BATCH_SIZE = args.batch
model = Model()
if args.eval:
batch = 128 # something that can run on one gpu
ds = get_data('val', batch)
eval_classification(model, get_model_loader(args.load), ds)
elif args.flops:
# manually build the graph with batch=1
with TowerContext('', is_training=False):
model.build_graph(
tf.placeholder(tf.float32, [1, 224, 224, 3], 'input'),
tf.placeholder(tf.int32, [1], 'label'))
model_utils.describe_trainable_vars()
tf.profiler.profile(
tf.get_default_graph(),
cmd='op',
options=tf.profiler.ProfileOptionBuilder.float_operation())
logger.info(
"Note that TensorFlow counts flops in a different way from the paper."
)
help='ResNet depth',
type=int,
default=152,
choices=[50, 101, 152])
parser.add_argument('--arch',
help='Name of architectures defined in nets.py',
default='ResNetDenoise')
parser.add_argument('--load', help='path to checkpoint')
parser.add_argument('--input', help='path to input image')
args = parser.parse_args()
model = getattr(nets, args.arch + 'Model')(args)
input = tf.placeholder(tf.float32, shape=(None, 224, 224, 3))
image = input / 127.5 - 1.0
image = tf.transpose(image, [0, 3, 1, 2])
with TowerContext('', is_training=False):
logits = model.get_logits(image)
sess = tf.Session()
get_model_loader(args.load).init(sess)
sample = cv2.imread(args.input) # this is a BGR image, not RGB
# imagenet evaluation uses standard imagenet pre-processing
# (resize shortest edge to 256 + center crop 224).
# However, for images of unknown sources, let's just do a naive resize.
sample = cv2.resize(sample, (224, 224))
l = sess.run(logits, feed_dict={input: np.array([sample])})
print("Prediction: ", l.argmax())
