def main(argv):
""" The main method for this module.
"""
logfile = os.path.join(os.getcwd(), 'log-mrlite-scheduler.txt')
config_logging(logfile)
start_time = time.time()
logging.info('Job started at %s' % time.asctime())
parser = MRLiteOptionParser()
options, args = parser.parse_args(argv)
scheduler = MRLiteJobScheduler(options)
try:
scheduler.start_jobs()
scheduler.monitor_jobs()
scheduler.quit_jobs()
logging.info('Job finished at %s' % time.asctime())
logging.info('Job run for %.3f seconds' % (time.time() - start_time))
except KeyboardInterrupt:
logging.info('Interrupted by CTRL-C')
scheduler.kill_jobs()
sys.exit(-1)
except Exception:
exc_type, exc_value, exc_traceback = sys.exc_info()
mesg = ''.join(
traceback.format_exception(exc_type, exc_value, exc_traceback))
logging.info('cache exception in mrlite.py')
logging.info(mesg)
scheduler.kill_jobs()
sys.exit(-1)
def main(argv):
""" The main method for this moduler.
"""
if len(argv) != 2:
print 'the config file must be specified.'
sys.exit(1)
logfile = os.path.join(os.getcwd(), 'brook-log-scheduler.txt')
config_logging(logfile)
start_time = time.time()
logging.info('Job started at %s' %time.asctime())
configparser = ConfigFileParser(argv[1])
options = configparser.parser()
logging.info('Parse config file options at %s' %time.asctime())
scheduler = BrookJobScheduler(options)
logging.info('Initialize of scheduler success at %s' %time.asctime())
try:
scheduler.start_jobs()
scheduler.monitor_jobs()
scheduler.quit_jobs()
logging.info('Job finished at %s' %time.asctime())
logging.info('Job run for %.3f seconds' %(time.time() - start_time))
except KeyboardInterrupt:
logging.info('Interrupted by CTRL-C')
scheduler.kill_jobs()
sys.exit(-1)
except Exception:
exc_type, exc_value, exc_traceback = sys.exc_info()
mesg = ''.join(traceback.format_exception(exc_type, exc_value,
exc_traceback))
logging.info('cache exception in brook_scheduler.py')
logging.info(mesg)
scheduler.kill_jobs()
sys.exit(-1)
def __init__(self, server, port, rank, dir):
""" Establish communication with scheduler
"""
self.options = None
logfile = os.path.join(dir, 'log-mrlite-rank-%s.txt' %rank)
config_logging(logfile)
logging.debug('communicator %s started in %s:%s at %s' %(
rank, server, port, time.asctime()))
sockobj = socket.socket(socket.AF_INET, socket.SOCK_STREAM)
sockobj.connect((server, port))
sock = SocketWrapper(sockobj)
sock.send('rank %s %s' %(rank, os.getpid()))
pickled_options = sock.recv()
options = pickle.loads(pickled_options)
self.task_callback = {
'start_mapper' : None,
'start_reducer' : None,
'status': self.report_status,
'quit' : self.quit,
'exit' : self.quit,
}
self.server = server
self.rank = rank
self.sock = sock
self.options = options
self.process = None
if options.maponly_mode:
self.worker = MapOnlyWorker(options, rank, sock)
elif rank < options.num_map_worker:
self.worker = MapWorker(options, rank, sock)
else:
self.worker = ReduceWorker(options, rank, sock)
def __init__(self, server, port, rank, dir):
""" Establish communication with scheduler
"""
self.options = None
logfile = os.path.join(dir, 'log-mrlite-rank-%s.txt' % rank)
config_logging(logfile)
logging.debug('communicator %s started in %s:%s at %s' %
(rank, server, port, time.asctime()))
sockobj = socket.socket(socket.AF_INET, socket.SOCK_STREAM)
sockobj.connect((server, port))
sock = SocketWrapper(sockobj)
sock.send('rank %s %s' % (rank, os.getpid()))
pickled_options = sock.recv()
options = pickle.loads(pickled_options)
self.task_callback = {
'start_mapper': None,
'start_reducer': None,
'status': self.report_status,
'quit': self.quit,
'exit': self.quit,
}
self.server = server
self.rank = rank
self.sock = sock
self.options = options
self.process = None
if options.maponly_mode:
self.worker = MapOnlyWorker(options, rank, sock)
elif rank < options.num_map_worker:
self.worker = MapWorker(options, rank, sock)
else:
self.worker = ReduceWorker(options, rank, sock)
def main(argv):
""" The main method for this module.
"""
logfile = os.path.join(os.getcwd(), 'log-mrlite-scheduler.txt')
config_logging(logfile)
start_time = time.time()
logging.info('Job started at %s' %time.asctime())
parser = MRLiteOptionParser()
options, args = parser.parse_args(argv)
scheduler = MRLiteJobScheduler(options)
try:
scheduler.start_jobs()
scheduler.monitor_jobs()
scheduler.quit_jobs()
logging.info('Job finished at %s' %time.asctime())
logging.info('Job run for %.3f seconds' %(time.time() - start_time))
except KeyboardInterrupt:
logging.info('Interrupted by CTRL-C')
scheduler.kill_jobs()
sys.exit(-1)
except Exception:
exc_type, exc_value, exc_traceback = sys.exc_info()
mesg = ''.join(traceback.format_exception(exc_type, exc_value,
exc_traceback))
logging.info('cache exception in mrlite.py')
logging.info(mesg)
scheduler.kill_jobs()
sys.exit(-1)
def main(argv):
""" The main method for this moduler.
"""
if len(argv) != 2:
print 'the config file must be specified.'
sys.exit(1)
logfile = os.path.join(os.getcwd(), 'brook-log-scheduler.txt')
config_logging(logfile)
start_time = time.time()
logging.info('Job started at %s' % time.asctime())
configparser = ConfigFileParser(argv[1])
options = configparser.parser()
logging.info('Parse config file options at %s' % time.asctime())
scheduler = BrookJobScheduler(options)
logging.info('Initialize of scheduler success at %s' % time.asctime())
try:
scheduler.start_jobs()
scheduler.monitor_jobs()
scheduler.quit_jobs()
logging.info('Job finished at %s' % time.asctime())
logging.info('Job run for %.3f seconds' % (time.time() - start_time))
except KeyboardInterrupt:
logging.info('Interrupted by CTRL-C')
scheduler.kill_jobs()
sys.exit(-1)
except Exception:
exc_type, exc_value, exc_traceback = sys.exc_info()
mesg = ''.join(
traceback.format_exception(exc_type, exc_value, exc_traceback))
logging.info('cache exception in brook_scheduler.py')
logging.info(mesg)
scheduler.kill_jobs()
sys.exit(-1)
def main(command_line_arguments):
config_logging()
parser = argparse.ArgumentParser(
description=
"""Wrapper around docker run that protects against Zombie containers""",
epilog="")
parser.add_argument(
"-u",
"--user",
help="Username or UID (format: <name|uid>[:<group|gid>])",
default=None)
parser.add_argument("-v",
"--volume",
action='append',
type=_volume_mount,
help="Bind mount a volume",
default=[])
parser.add_argument("--cap-add",
help="Add Linux capabilities",
action="append",
type=str,
default=[])
parser.add_argument("--runtime",
help="Runtime to use for this container",
default=None)
parser.add_argument("--name",
help="Assign a name to the container",
default=None)
parser.add_argument("image", metavar="IMAGE")
parser.add_argument("command", metavar="COMMAND")
parser.add_argument("args", nargs='*', metavar="ARG")
args = parser.parse_args(args=command_line_arguments)
docker_client = SafeDockerClient()
return docker_client.run(
args.image, **{
"command": " ".join(list(chain([args.command] + args.args))),
"user": args.user,
"runtime": args.runtime,
"name": args.name,
"volumes": reduce(lambda dct, v: {
**dct,
**v
}, args.volume, {}),
"cap_add": args.cap_add
})
def main(command_line_arguments):
config_logging()
parser = argparse.ArgumentParser(
description="Safe docker login utility to avoid leaking passwords",
epilog="")
parser.add_argument("--secret-name",
help="Secret name",
type=str,
required=True)
parser.add_argument("--secret-endpoint-url",
help="Endpoint Url",
type=str,
default=os.environ.get("DOCKERHUB_SECRET_ENDPOINT_URL",
None))
parser.add_argument("--secret-endpoint-region",
help="AWS Region",
type=str,
default=os.environ.get(
"DOCKERHUB_SECRET_ENDPOINT_REGION", None))
args = parser.parse_args(args=command_line_arguments)
if args.secret_endpoint_url is None:
raise RuntimeError(
"Could not determine secret-endpoint-url, please specify with --secret-endpoint-url"
)
if args.secret_endpoint_region is None:
raise RuntimeError(
"Could not determine secret-endpoint-region, please specify with --secret-endpoint-region"
)
try:
login_dockerhub(args.secret_name, args.secret_endpoint_url,
args.secret_endpoint_region)
except Exception as err:
logging.exception(err)
exit(1)
msgtype = parse_unsigned_byte(stream)
if not msg_spec[msgtype]:
raise UnsupportedPacketException(msgtype)
logger.debug("%s trying to parse message type %x" % (side, msgtype))
msg_parser = msg_spec[msgtype]
msg = msg_parser.parse(stream)
msg['raw_bytes'] = stream.packet_finished()
return Message(msg)
if __name__ == "__main__":
logging.basicConfig(level=logging.ERROR)
(host, port, opts, pcfg) = parse_args()
if opts.logfile:
util.config_logging(opts.logfile)
if opts.loglvl:
logging.root.setLevel(getattr(logging, opts.loglvl.upper()))
# Install signal handler.
signal.signal(signal.SIGINT, sigint_handler)
while True:
cli_sock = wait_for_client(opts.locport)
# Set up client/server main-in-the-middle.
sleep(0.05)
MinecraftSession(pcfg, cli_sock, host, port)
# I/O event loop.
#!/usr/bin/env python
import gzip
from pipeline import *
import util
config = util.read_config('import')
log = util.config_logging(config)
table = 'area.area'
columns=('area', 'zip', 'po_name', 'geom')
filename = 'bayareadata.gz'
area = 'sfbayarea'
db = util.DB.from_config(config)
log.info('importing file %r to table %r' % (filename, table))
# compose import pipeline
cat(gzip.open(filename)) | skip(head=2, tail=2) | split(sep='|') |\
transform([lambda r: area, 0, 1, 2]) |\
join('\t') | load_data(db, table, columns=columns, clean=True)
def __init__(self):
self.config = read_config(self.source_name)
self.log = config_logging(self.config).getLogger(self.source_name)
self.db = DB.from_config(self.config)
venue = os.path.join(api_v1, 'venues/<id:int>'),
venue_nearby = os.path.join(api_v1, 'venues/<id:int>/nearby'),
categories = os.path.join(api_v1, 'categories'),
category = os.path.join(api_v1, 'categories/<id:int>'),
category_venues = os.path.join(api_v1, 'categories/<id:int>/venues'),
zips = os.path.join(api_v1, 'zips'),
zip = os.path.join(api_v1, 'zips/<zip>'),
zip_venues = os.path.join(api_v1, 'zips/<zip>/venues'),
)
### setup
config = util.read_config("api")
log = util.config_logging(config).getLogger("server")
# connection pool
pool = psycopg2.pool.ThreadedConnectionPool(
minconn=1,
maxconn=int(config.max_connections or 10),
dsn=config.db,
cursor_factory=psycopg2.extras.RealDictCursor
)
# install plugin
install(bottle_pgpool.PgSQLPoolPlugin(pool))
### helpers
def fq_url(*path, **subst):
msgtype = parse_unsigned_byte(stream)
if not msg_spec[msgtype]:
raise UnsupportedPacketException(msgtype)
logger.debug("%s trying to parse message type %x" % (side, msgtype))
msg_parser = msg_spec[msgtype]
msg = msg_parser.parse(stream)
msg['raw_bytes'] = stream.packet_finished()
return Message(msg)
if __name__ == "__main__":
logging.basicConfig(level=logging.ERROR)
(host, port, opts, pcfg) = parse_args()
if opts.logfile:
util.config_logging(opts.logfile)
if opts.loglvl:
logging.root.setLevel(getattr(logging, opts.loglvl.upper()))
# Install signal handler.
signal.signal(signal.SIGINT, sigint_handler)
while True:
cli_sock = wait_for_client(opts.locport)
# Set up client/server main-in-the-middle.
sleep(0.05)
MinecraftSession(pcfg, cli_sock, host, port)
# I/O event loop.
#!/usr/bin/env python
import gzip
from pipeline import *
import util
config = util.read_config('import')
log = util.config_logging(config)
table = 'area.area'
columns = ('area', 'zip', 'po_name', 'geom')
filename = 'bayareadata.gz'
area = 'sfbayarea'
db = util.DB.from_config(config)
log.info('importing file %r to table %r' % (filename, table))
# compose import pipeline
cat(gzip.open(filename)) | skip(head=2, tail=2) | split(sep='|') |\
transform([lambda r: area, 0, 1, 2]) |\
join('\t') | load_data(db, table, columns=columns, clean=True)
def main(args):
util.config_logging()
# Settings
lr = args.lr
beta1 = args.beta1
batch_size = args.batch_size
iterations = args.iterations
snapshot = args.snapshot
stepsize = args.stepsize
display = args.display
path_source_cnn = './output/source_cnn'
output_dir = os.path.join('output', 'target_cnn')
save_path = os.path.join(output_dir, 'target_cnn.ckpt')
if not os.path.exists(output_dir):
os.makedirs(output_dir)
# Load source data
train_mat = loadmat('./data/svhn/train_32x32.mat')
train_images = train_mat['X'].transpose((3, 0, 1, 2))
train_images = train_images.astype(np.float32) / 255.
RGB2GRAY = np.array([0.2989, 0.5870, 0.1140], dtype=np.float32)
train_images = np.sum(np.multiply(train_images, RGB2GRAY),
3,
keepdims=True)
# Load target data
target_images = util._read_images(
'./data/mnist/train-images-idx3-ubyte.gz')
# Data generator
idg = ImageDataGenerator()
source_data_gen = idg.flow(train_images,
batch_size=batch_size,
shuffle=True)
target_data_gen = idg.flow(target_images,
batch_size=batch_size,
shuffle=True)
# Define graph
nb_classes = 10
tf.reset_default_graph()
x_source = tf.placeholder(tf.float32, (None, 32, 32, 1))
x_source_resized = tf.image.resize_images(x_source, [28, 28])
x_target = tf.placeholder(tf.float32, (None, 28, 28, 1))
feature_src = source_cnn(x_source_resized,
nb_classes=nb_classes,
trainable=False,
adapt=True)
feature_target = target_cnn(x_target, nb_classes, trainable=True)
d_logits_src = discriminator(feature_src)
d_logits_target = discriminator(feature_target, reuse=True)
# Loss: Discriminator
d_loss_src = tf.reduce_mean(
tf.nn.sigmoid_cross_entropy_with_logits(
logits=d_logits_src, labels=tf.ones_like(d_logits_src)))
d_loss_target = tf.reduce_mean(
tf.nn.sigmoid_cross_entropy_with_logits(
logits=d_logits_target, labels=tf.zeros_like(d_logits_target)))
d_loss = d_loss_src + d_loss_target
# Loss: target CNN
target_loss = tf.reduce_mean(
tf.nn.sigmoid_cross_entropy_with_logits(
logits=d_logits_target, labels=tf.ones_like(d_logits_target)))
t_vars = tf.trainable_variables()
target_vars = [
var for var in tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES,
scope='target_cnn')
]
d_vars = [var for var in t_vars if var.name.startswith('discriminator')]
src_vars = [
var for var in tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES,
scope='source_cnn')
]
lr_var = tf.Variable(lr, name='learning_rate', trainable=False)
optimizer = tf.train.AdamOptimizer(lr_var, beta1)
target_train_op = optimizer.minimize(target_loss, var_list=target_vars)
d_train_op = optimizer.minimize(d_loss, var_list=d_vars)
# Train
source_saver = tf.train.Saver(var_list=src_vars)
target_saver = tf.train.Saver(var_list=target_vars)
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
target_losses = deque(maxlen=10)
d_losses = deque(maxlen=10)
bar = tqdm(range(iterations))
bar.set_description('(lr: {:.0e})'.format(lr))
bar.refresh()
losses = []
with tf.Session(config=config) as sess:
sess.run(tf.global_variables_initializer())
source_saver.restore(sess, tf.train.latest_checkpoint(path_source_cnn))
for i in bar:
batch_source = next(source_data_gen)
batch_target = next(target_data_gen)
target_loss_val, d_loss_val, _, _ = sess.run(
[target_loss, d_loss, target_train_op, d_train_op],
feed_dict={
x_source: batch_source,
x_target: batch_target
})
target_losses.append(target_loss_val)
d_losses.append(d_loss_val)
losses.append([target_loss_val, d_loss_val])
if i % display == 0:
logging.info(
'{:20} Target: {:5.4f} (avg: {:5.4f})'
' Discriminator: {:5.4f} (avg: {:5.4f})'.format(
'Iteration {}:'.format(i), target_loss_val,
np.mean(target_losses), d_loss_val, np.mean(d_losses)))
if stepsize is not None and (i + 1) % stepsize == 0:
lr = sess.run(lr_var.assign(lr * 0.1))
logging.info('Changed learning rate to {:.0e}'.format(lr))
bar.set_description('(lr: {:.0e})'.format(lr))
if (i + 1) % snapshot == 0:
snapshot_path = target_saver.save(sess, save_path)
logging.info('Saved snapshot to {}'.format(snapshot_path))
# Save visualization of training losses
losses = np.array(losses)
plt.plot(losses.T[0], label='Target CNN Loss', alpha=0.5)
plt.plot(losses.T[1], label='Discriminator Loss', alpha=0.5)
plt.title('Training Losses')
plt.legend()
plt.savefig('./losses.png')
def parse_args(self):
"""
Configure the job from command line arguments.
"""
parser = argparse.ArgumentParser(description=self.description)
parser.add_argument("-c",
"--config-file",
help="Config file locations",
action='append')
parser.add_argument("-d", "--run-dir", nargs='?', help="Job run dir")
parser.add_argument("-o",
"--out",
nargs='?',
help="File for component stdout")
parser.add_argument("-e",
"--err",
nargs='?',
help="File for component stderr")
parser.add_argument(
"-l",
"--log",
nargs='?',
help="File for logging output (default is print to terminal)")
parser.add_argument("-L",
"--level",
nargs='?',
help="Global log level")
parser.add_argument("-s",
"--job-steps",
type=int,
default=-1,
help="Job steps to run (single number)")
parser.add_argument("-i",
"--job-id",
type=int,
help="Job ID from JSON job store",
default=None)
parser.add_argument("script",
nargs='?',
help="Path or name of job script")
parser.add_argument("params",
nargs='?',
help="Job param file in JSON format")
# TODO: CL option to disable automatic copying of ouput files.
# The files should be symlinked if copying is disabled.
# parser.add_argument("--no-copy-output-files", help="Disable copying of output files")
#parser.add_argument('--feature', dest='feature', action='store_true')
#parser.add_argument('--no-feature', dest='feature', action='store_false')
#parser.set_defaults(feature=True)
cl = parser.parse_args(self.args)
if cl.run_dir:
self.rundir = cl.run_dir
if cl.level:
num_level = getattr(logging, cl.level.upper(), None)
if not isinstance(num_level, int):
raise ValueError('Invalid log level: %s' % num_level)
logging.getLogger('hpsmc').setLevel(num_level)
#print("Set log level of hpsmc: %s" % logging.getLevelName(logging.getLogger('hpsmc').getEffectiveLevel()))
if cl.log:
self.log_file = cl.log
if not os.path.isabs(self.log_file):
self.log_file = os.path.abspath(self.log_file)
config_logging(stream=open(self.log_file, 'w'))
if cl.out:
self.out_file = cl.out
if not os.path.isabs(self.out_file):
self.out_file = os.path.abspath(self.out_file)
logger.info('Changed stdout file to abs path: %s' %
self.out_file)
if cl.err:
self.err_file = cl.err
if not os.path.isabs(self.err_file):
self.err_file = os.path.abspath(self.err_file)
logger.info('Changed stderr file to abs path: %s' %
self.err_file)
if cl.config_file:
self.config_files = list(map(os.path.abspath, cl.config_file))
else:
self.config_files = []
self.job_steps = cl.job_steps
if cl.script:
self.script = cl.script
else:
raise Exception('Missing required script name or location.')
# Params are actually optional as some job scripts might not need them.
if cl.params:
self.param_file = os.path.abspath(cl.params)
params = {}
if cl.job_id:
# Load data from a job store containing multiple jobs.
self.job_id = cl.job_id
logger.info("Loading job with ID %d from job store '%s'" %
(self.job_id, self.param_file))
jobstore = JobStore(self.param_file)
if jobstore.has_job_id(self.job_id):
params = jobstore.get_job(self.job_id)
else:
raise Exception(
"No job id %d was found in the job store '%s'" %
(self.job_id, self.param_file))
else:
# Load data from a JSON file with a single job definition.
logger.info('Loading job parameters from file: %s' %
self.param_file)
params = load_json_data(self.param_file)
if not isinstance(params, dict):
raise Exception(
'Job ID must be provided when running from a job store.'
)
self._load_params(params)
msgtype = parse_unsigned_byte(stream)
if not msg_spec[msgtype]:
raise UnsupportedPacketException(msgtype)
logger.debug("%s trying to parse message type %x" % (side, msgtype))
msg_parser = msg_spec[msgtype]
msg = msg_parser.parse(stream)
msg['raw_bytes'] = stream.packet_finished()
return Message(msg)
if __name__ == "__main__":
logging.basicConfig(level=logging.ERROR)
(host, port, opts, pcfg) = parse_args()
if opts.logfile:
config_logging(opts.logfile)
if opts.loglvl:
logging.root.setLevel(getattr(logging, opts.loglvl.upper()))
if opts.user:
while True:
password = getpass("Minecraft account password: "******"Authenticating with %s" % opts.user)
if auth.check():
break
logger.error("Authentication failed")
logger.debug("Credentials are valid")
if opts.authenticate or opts.password_file:
from util import config_logging # Enable default logging config config_logging()
venue=os.path.join(api_v1, "venues/<id:int>"),
venue_nearby=os.path.join(api_v1, "venues/<id:int>/nearby"),
categories=os.path.join(api_v1, "categories"),
category=os.path.join(api_v1, "categories/<id:int>"),
category_venues=os.path.join(api_v1, "categories/<id:int>/venues"),
zips=os.path.join(api_v1, "zips"),
zip=os.path.join(api_v1, "zips/<zip>"),
zip_venues=os.path.join(api_v1, "zips/<zip>/venues"),
)
### setup
config = util.read_config("api")
log = util.config_logging(config).getLogger("server")
# connection pool
pool = psycopg2.pool.ThreadedConnectionPool(
minconn=1, maxconn=int(config.max_connections or 10), dsn=config.db, cursor_factory=psycopg2.extras.RealDictCursor
)
# install plugin
install(bottle_pgpool.PgSQLPoolPlugin(pool))
### helpers
def fq_url(*path, **subst):
"""
Builds fully qualified url. Uses bottle request object for scheme and loc
info and works only within request
while not self.closed:
try:
self.handle_read()
except EOFException:
break
gevent.sleep()
self.close()
self.handle_close()
if __name__ == "__main__":
logging.basicConfig(level=logging.ERROR)
(host, port, opts, pcfg) = parse_args()
if opts.logfile:
config_logging(opts.logfile)
if opts.loglvl:
logging.root.setLevel(getattr(logging, opts.loglvl.upper()))
if opts.user:
while True:
password = getpass("Minecraft account password: "******"Authenticating with %s" % opts.user)
if auth.valid():
break
logger.error("Authentication failed")
logger.debug("Credentials are valid")
if opts.authenticate or opts.password_file:
def __init__(self):
self.config = read_config(self.source_name)
self.log = config_logging(self.config).getLogger(self.source_name)
self.db = DB.from_config(self.config)
def main(args):
util.config_logging()
# Parameters
lr = args.lr
batch_size = args.batch_size
iterations = args.iterations
snapshot = args.snapshot
stepsize = args.stepsize
display = args.display
output_dir = 'output/source_cnn/'
output_dir_clf = 'output/classifier/'
save_path = os.path.join(output_dir, 'source_cnn.ckpt')
save_path_clf = os.path.join(output_dir_clf, 'classifier.ckpt')
if not os.path.exists(output_dir):
os.makedirs(output_dir)
if not os.path.exists(output_dir_clf):
os.makedirs(output_dir_clf)
# Load train data
train_mat = loadmat('./data/svhn/train_32x32.mat')
train_images = train_mat['X'].transpose((3, 0, 1, 2))
train_images = train_images.astype(np.float32) / 255.
train_labels = train_mat['y'].squeeze()
train_labels[train_labels == 10] = 0
RGB2GRAY = np.array([0.2989, 0.5870, 0.1140], dtype=np.float32)
train_images = np.sum(np.multiply(train_images, RGB2GRAY),
3,
keepdims=True)
# Data generator
idg = ImageDataGenerator()
train_data_gen = idg.flow(train_images,
train_labels,
batch_size=batch_size,
shuffle=True)
# Define graph
nb_classes = 10
tf.reset_default_graph()
x = tf.placeholder(tf.float32, (None, 32, 32, 1))
x_resized = tf.image.resize_images(x, [28, 28])
t = tf.placeholder(tf.int32, (None, ))
logits = source_cnn(x_resized, nb_classes=nb_classes, trainable=True)
loss = tf.losses.sparse_softmax_cross_entropy(t, logits)
lr_var = tf.Variable(lr, name='learning_rate', trainable=False)
optimizer = tf.train.AdamOptimizer(learning_rate=lr)
train_op = optimizer.minimize(loss)
t_vars = tf.trainable_variables()
source_cnn_vars = [
var for var in t_vars if var.name.startswith('source_cnn')
]
classifier_vars = [
var for var in t_vars if var.name.startswith('classifier')
]
# Train
cnn_saver = tf.train.Saver(var_list=source_cnn_vars)
if len(classifier_vars) > 0:
clf_saver = tf.train.Saver(var_list=classifier_vars)
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
losses = deque(maxlen=10)
training_losses = []
bar = tqdm(range(iterations))
bar.set_description('(lr: {:.0e})'.format(lr))
bar.refresh()
with tf.Session(config=config) as sess:
sess.run(tf.global_variables_initializer())
for i in bar:
batch_images, batch_labels = next(train_data_gen)
loss_val, _ = sess.run([loss, train_op],
feed_dict={
x: batch_images,
t: batch_labels
})
losses.append(loss_val)
training_losses.append(loss_val)
if i % display == 0:
logging.info('{:20} {:10.4f} (avg: {:10.4f})'.format(
'Iteration {}:'.format(i), loss_val, np.mean(losses)))
if stepsize is not None and (i + 1) % stepsize == 0:
lr = sess.run(lr_var.assign(lr * 0.1))
logging.info('Changed learning rate to {:.0e}'.format(lr))
bar.set_description('(lr: {:.0e})'.format(lr))
if (i + 1) % snapshot == 0:
snapshot_path = cnn_saver.save(sess, save_path)
if len(classifier_vars) > 0:
clf_saver.save(sess, save_path_clf)
logging.info('Saved snapshot to {}'.format(snapshot_path))
plt.plot(training_losses, label='Source CNN Loss')
plt.title('Pre-Training Loss')
plt.legend()
plt.savefig('./pre_training_losses.png')
