def main(args):
role = role_maker.PaddleCloudRoleMaker(is_collective=True)
fleet.init(role)
config = get_config(args.config, overrides=args.override, show=True)
gpu_id = int(os.environ.get('FLAGS_selected_gpus', 0))
place = fluid.CUDAPlace(gpu_id)
startup_prog = fluid.Program()
valid_prog = fluid.Program()
valid_dataloader, valid_fetchs = program.build(
config, valid_prog, startup_prog, is_train=False)
valid_prog = valid_prog.clone(for_test=True)
exe = fluid.Executor(place)
exe.run(startup_prog)
init_model(config, valid_prog, exe)
valid_reader = Reader(config, 'valid')()
valid_dataloader.set_sample_list_generator(valid_reader, place)
compiled_valid_prog = program.compile(config, valid_prog)
program.run(valid_dataloader, exe, compiled_valid_prog, valid_fetchs, -1,
'eval', config)
def main(args):
config = get_config(args.config, overrides=args.override, show=True)
use_gpu = config.get("use_gpu", True)
places = fluid.cuda_places() if use_gpu else fluid.cpu_places()
startup_prog = fluid.Program()
valid_prog = fluid.Program()
valid_dataloader, valid_fetchs = program.build(config,
valid_prog,
startup_prog,
is_train=False,
is_distributed=False)
valid_prog = valid_prog.clone(for_test=True)
exe = fluid.Executor(places[0])
exe.run(startup_prog)
init_model(config, valid_prog, exe)
valid_reader = Reader(config, 'valid')()
valid_dataloader.set_sample_list_generator(valid_reader, places)
compiled_valid_prog = program.compile(config, valid_prog)
program.run(valid_dataloader, exe, compiled_valid_prog, valid_fetchs, -1,
'eval')
def main(args):
role = role_maker.PaddleCloudRoleMaker(is_collective=True)
fleet.init(role)
config = get_config(args.config, overrides=args.override, show=True)
# assign the place
gpu_id = int(os.environ.get('FLAGS_selected_gpus', 0))
place = fluid.CUDAPlace(gpu_id)
# startup_prog is used to do some parameter init work,
# and train prog is used to hold the network
startup_prog = fluid.Program()
train_prog = fluid.Program()
train_dataloader, train_fetchs = program.build(config,
train_prog,
startup_prog,
is_train=True)
if config.validate:
valid_prog = fluid.Program()
valid_dataloader, valid_fetchs = program.build(config,
valid_prog,
startup_prog,
is_train=False)
# clone to prune some content which is irrelevant in valid_prog
valid_prog = valid_prog.clone(for_test=True)
# create the "Executor" with the statement of which place
exe = fluid.Executor(place=place)
# only run startup_prog once to init
exe.run(startup_prog)
# load model from checkpoint or pretrained model
init_model(config, train_prog, exe)
train_reader = Reader(config, 'train')()
train_dataloader.set_sample_list_generator(train_reader, place)
if config.validate:
valid_reader = Reader(config, 'valid')()
valid_dataloader.set_sample_list_generator(valid_reader, place)
compiled_valid_prog = program.compile(config, valid_prog)
compiled_train_prog = fleet.main_program
for epoch_id in range(config.epochs):
# 1. train with train dataset
program.run(train_dataloader, exe, compiled_train_prog, train_fetchs,
epoch_id, 'train')
# 2. validate with validate dataset
if config.validate and epoch_id % config.valid_interval == 0:
program.run(valid_dataloader, exe, compiled_valid_prog,
valid_fetchs, epoch_id, 'valid')
# 3. save the persistable model
if epoch_id % config.save_interval == 0:
model_path = os.path.join(config.model_save_dir,
config.ARCHITECTURE["name"])
save_model(train_prog, model_path, epoch_id)
def main(args, return_dict={}):
config = get_config(args.config, overrides=args.override, show=True)
config.mode = "valid"
# assign place
use_gpu = config.get("use_gpu", True)
place = paddle.set_device('gpu' if use_gpu else 'cpu')
trainer_num = paddle.distributed.get_world_size()
use_data_parallel = trainer_num != 1
config["use_data_parallel"] = use_data_parallel
if config["use_data_parallel"]:
paddle.distributed.init_parallel_env()
net = program.create_model(config.ARCHITECTURE, config.classes_num)
if config["use_data_parallel"]:
net = paddle.DataParallel(net)
init_model(config, net, optimizer=None)
valid_dataloader = Reader(config, 'valid', places=place)()
net.eval()
with paddle.no_grad():
top1_acc = program.run(valid_dataloader, config, net, None, None, 0,
'valid')
return_dict["top1_acc"] = top1_acc
return top1_acc
def teleportrecieve():
content = request.json
q0 = content["q0"]
q1 = content["q1"]
p = """QUBITS 4
MEASURE 0
MEASURE 1
H 2
CNOT 2 3
CLASSICAL 1 1 1
X 3
CLASSICAL 0 1 1
Z 3"""
p_list = p.splitlines()
if q0 == 1:
p_list.insert(1, "X 0")
if q1 == 1:
p_list.insert(1, "X 1")
p = "\n".join(p_list)
wvf, msg = program.run(p)
msg = isolate_qubit(wvf, 3)
return jsonify({"program": p, 'wvf': msg})
def teleportsend():
sendit = request.form.get('sendit')
if not sendit is None:
url = request.url_root + 'api/teleportrecieve'
p = """QUBITS 3
H 1
CNOT 1 2
CNOT 0 1
H 0
MEASURE 0
MEASURE 1"""
wvf, msg = program.run(p)
# Very hacky way of doing this... @TODO make this better
m = re.findall('====== MEASURE qubit (\d) : (\d)', msg)
if m[0][0] == '0':
q0 = int(m[0][1])
q1 = int(m[1][1])
else:
q1 = int(m[0][1])
q0 = int(m[1][1])
print(url)
res = requests.post(url, json={"q0": q0, "q1": q1})
if res.ok:
j = res.json()
p = j['program']
p_split = p.splitlines()
p_str = "<ol>"
for i in p_split:
p_str += '<li><samp class="code-block">' + i + '</li>'
p_str += '</ol>'
j['program'] = p_str
j['a'] = f"Qubit 0 : {q0}<br />Qubit 1 : {q1}"
return (jsonify(j))
def main(args):
role = role_maker.PaddleCloudRoleMaker(is_collective=True)
fleet.init(role)
config = get_config(args.config, overrides=args.override, show=True)
place = env.place()
startup_prog = fluid.Program()
train_prog = fluid.Program()
train_dataloader, train_fetchs = program.build(
config, train_prog, startup_prog, is_train=True)
if config.validate:
valid_prog = fluid.Program()
valid_dataloader, valid_fetchs = program.build(
config, valid_prog, startup_prog, is_train=False)
valid_prog = valid_prog.clone(for_test=True)
exe = fluid.Executor(place)
exe.run(startup_prog)
init_model(config, train_prog, exe)
train_reader = Reader(config, 'train')()
train_dataloader.set_sample_list_generator(train_reader, place)
if config.validate:
valid_reader = Reader(config, 'valid')()
valid_dataloader.set_sample_list_generator(valid_reader, place)
compiled_valid_prog = program.compile(config, valid_prog)
compiled_train_prog = fleet.main_program
for epoch_id in range(config.epochs):
program.run(train_dataloader, exe, compiled_train_prog, train_fetchs,
epoch_id, 'train')
if config.validate and epoch_id % config.valid_interval == 0:
program.run(valid_dataloader, exe, compiled_valid_prog,
valid_fetchs, epoch_id, 'valid')
if epoch_id % config.save_interval == 0:
model_path = os.path.join(config.model_save_dir,
config.architecture)
save_model(train_prog, model_path, epoch_id)
def main():
global settings
if CONFIG_ENABLED:
if VERBOSITY_ENABLED:
print ("Configuring options ...")
settings = config.load_xml_config(settings)
if VERBOSITY_ENABLED:
print ("Configuration complete.")
pprint(settings)
else:
if VERBOSITY_ENABLED:
print ("Configuration disabled. Using defaults.")
settings = { "logfile" : "{}.log", "tempdir" : "tmp"}
if ENV_VARS_ENABLED:
if VERBOSITY_ENABLED:
print ("Checking environment ...")
settings = envvars(settings)
if VERBOSITY_ENABLED:
print ("Environment variables loaded.")
if ARGS_ENABLED:
if VERBOSITY_ENABLED:
print ("Parsing command line arguments ...")
settings = getargs(settings)
if VERBOSITY_ENABLED:
print ("Command line arguments parsed.")
if LOGGING_ENABLED:
if VERBOSITY_ENABLED:
print ("Setting up log file ...")
if re.search("{}", settings["logfile"]):
settings["logfile"] = settings["logfile"].format(PROGRAM_NAME)
LOG = setuplog(settings)
LOG.info("Logging set up.")
if VERBOSITY_ENABLED:
print ("Logging set up.")
atexit.register(logging.shutdown)
program.run(settings)
sys.exit(0)
def main():
parser = argparse.ArgumentParser()
parser.add_argument(
"-i",
"--max_iterations",
default=10000,
type=int,
help=
"maximum number of iterations to reach solar mass. If more iterations are needed, the integration is stopped and rho is updated."
)
parser.add_argument("-m",
"--goal_mass",
default=1.5,
type=float,
help="Goal mass in solar masses")
parser.add_argument("-r",
"--rho_0",
default=1.,
type=float,
help="Density at r=0 in nuclear densities")
parser.add_argument("-K",
default=1.98183e-15,
type=float,
help="Adiabatic coefficient in km/m^2")
parser.add_argument("-y",
"--gamma",
default=2.75,
type=float,
help="polytropic index")
parser.add_argument("-t",
"--tolerance",
default=0.0001,
type=float,
help="Allowed deviation from goal mass")
args = parser.parse_args()
args.goal_mass *= c.SOLAR_MASS
args.rho_0 *= c.NUCLEAR_DENSITY
program.run(args)
return 0
def main(args, return_dict={}):
config = get_config(args.config, overrides=args.override, show=True)
config.mode = "valid"
# assign place
use_gpu = config.get("use_gpu", True)
place = paddle.set_device('gpu' if use_gpu else 'cpu')
multilabel = config.get("multilabel", False)
trainer_num = paddle.distributed.get_world_size()
use_data_parallel = trainer_num != 1
config["use_data_parallel"] = use_data_parallel
if config["use_data_parallel"]:
paddle.distributed.init_parallel_env()
net = program.create_model(config.ARCHITECTURE, config.classes_num)
init_model(config, net, optimizer=None)
valid_dataloader = Reader(config, 'valid', places=place)()
net.eval()
with paddle.no_grad():
if not multilabel:
top1_acc = program.run(valid_dataloader, config, net, None, None,
0, 'valid')
return_dict["top1_acc"] = top1_acc
return top1_acc
else:
all_outs = []
targets = []
for _, batch in enumerate(valid_dataloader()):
feeds = program.create_feeds(batch, False, config.classes_num,
multilabel)
out = net(feeds["image"])
out = F.sigmoid(out)
use_distillation = config.get("use_distillation", False)
if use_distillation:
out = out[1]
all_outs.extend(list(out.numpy()))
targets.extend(list(feeds["label"].numpy()))
all_outs = np.array(all_outs)
targets = np.array(targets)
mAP = mean_average_precision(all_outs, targets)
return_dict["mean average precision"] = mAP
return mAP
def add_message(uuid):
content = request.json
print(content['mytext'])
p = content['mytext']
#os.chmod("program.txt", 777)
text_file = open("compiler/program.txt", "w")
#os.chmod("program.txt", "w")
text_file.write(p)
text_file.close()
os.system("bash -c './compiler/compile.lisp program.txt'")
os.system("bash -c 'python preprocessor.py a.ir'")
with open('compiler/a.eg', 'r') as myfile:
p = myfile.read()
wvf, msg = program.run(p)
return jsonify({"results": msg})
def run(path):
with open(path, 'r') as myfile:
data = myfile.read()
wvf, msg = program.run(data)
print(msg)
# Very hacky way of doing this... @TODO make this better
m = re.findall('====== MEASURE qubit (\d) : (\d)', msg)
if m[0][0] == '0':
q0 = int(m[0][1])
q1 = int(m[1][1])
else:
q1 = int(m[0][1])
q0 = int(m[1][1])
res = requests.post(url, json={"q0": q0, "q1": q1})
if res.ok:
j = res.json()
print(j["results"])
def test_bell_states():
state_prep = {
"B00": "",
"B01": "X 1\n",
"B10": "X 0\n",
"B11": "X 0\nX 1\n",
}
#@TODO do this with matrices values instead of strings
bell_states = {
"B00": "0.71|00> + 0.71|11>",
"B01": "0.71|01> + 0.71|10>",
"B10": "0.71|00> + -0.71|11>",
"B11": "0.71|01> + -0.71|10>",
}
epr_program = "H 0\nCNOT 0 1"
for i in state_prep:
bell_program = "QUBITS 2\n" + state_prep[i] + epr_program
wvf = run(bell_program)
assert wavefunction(wvf) == bell_states[i]
def main(args):
paddle.seed(12345)
config = get_config(args.config, overrides=args.override, show=True)
# assign the place
use_gpu = config.get("use_gpu", True)
place = paddle.set_device('gpu' if use_gpu else 'cpu')
trainer_num = paddle.distributed.get_world_size()
use_data_parallel = trainer_num != 1
config["use_data_parallel"] = use_data_parallel
if config["use_data_parallel"]:
paddle.distributed.init_parallel_env()
net = program.create_model(config.ARCHITECTURE, config.classes_num)
optimizer, lr_scheduler = program.create_optimizer(
config, parameter_list=net.parameters())
dp_net = net
if config["use_data_parallel"]:
find_unused_parameters = config.get("find_unused_parameters", False)
dp_net = paddle.DataParallel(
net, find_unused_parameters=find_unused_parameters)
# load model from checkpoint or pretrained model
init_model(config, net, optimizer)
train_dataloader = Reader(config, 'train', places=place)()
if config.validate:
valid_dataloader = Reader(config, 'valid', places=place)()
last_epoch_id = config.get("last_epoch", -1)
best_top1_acc = 0.0 # best top1 acc record
best_top1_epoch = last_epoch_id
vdl_writer_path = config.get("vdl_dir", None)
vdl_writer = None
if vdl_writer_path:
from visualdl import LogWriter
vdl_writer = LogWriter(vdl_writer_path)
# Ensure that the vdl log file can be closed normally
try:
for epoch_id in range(last_epoch_id + 1, config.epochs):
net.train()
# 1. train with train dataset
program.run(train_dataloader, config, dp_net, optimizer,
lr_scheduler, epoch_id, 'train', vdl_writer)
# 2. validate with validate dataset
if config.validate and epoch_id % config.valid_interval == 0:
net.eval()
with paddle.no_grad():
top1_acc = program.run(valid_dataloader, config, net, None,
None, epoch_id, 'valid', vdl_writer)
if top1_acc > best_top1_acc:
best_top1_acc = top1_acc
best_top1_epoch = epoch_id
model_path = os.path.join(config.model_save_dir,
config.ARCHITECTURE["name"])
save_model(net, optimizer, model_path, "best_model")
message = "The best top1 acc {:.5f}, in epoch: {:d}".format(
best_top1_acc, best_top1_epoch)
logger.info(message)
# 3. save the persistable model
if epoch_id % config.save_interval == 0:
model_path = os.path.join(config.model_save_dir,
config.ARCHITECTURE["name"])
save_model(net, optimizer, model_path, epoch_id)
except Exception as e:
logger.error(e)
finally:
vdl_writer.close() if vdl_writer else None
def main(args):
config = get_config(args.config, overrides=args.override, show=True)
# assign the place
use_gpu = config.get("use_gpu", True)
places = fluid.cuda_places() if use_gpu else fluid.cpu_places()
# startup_prog is used to do some parameter init work,
# and train prog is used to hold the network
startup_prog = fluid.Program()
train_prog = fluid.Program()
best_top1_acc = 0.0 # best top1 acc record
if not config.get('use_ema'):
train_dataloader, train_fetchs = program.build(config,
train_prog,
startup_prog,
is_train=True,
is_distributed=False)
else:
train_dataloader, train_fetchs, ema = program.build(
config,
train_prog,
startup_prog,
is_train=True,
is_distributed=False)
if config.validate:
valid_prog = fluid.Program()
valid_dataloader, valid_fetchs = program.build(config,
valid_prog,
startup_prog,
is_train=False,
is_distributed=False)
# clone to prune some content which is irrelevant in valid_prog
valid_prog = valid_prog.clone(for_test=True)
# create the "Executor" with the statement of which place
exe = fluid.Executor(places[0])
# Parameter initialization
exe.run(startup_prog)
# load model from 1. checkpoint to resume training, 2. pretrained model to finetune
init_model(config, train_prog, exe)
train_reader = Reader(config, 'train')()
train_dataloader.set_sample_list_generator(train_reader, places)
compiled_train_prog = program.compile(config, train_prog,
train_fetchs['loss'][0].name)
if config.validate:
valid_reader = Reader(config, 'valid')()
valid_dataloader.set_sample_list_generator(valid_reader, places)
compiled_valid_prog = program.compile(config,
valid_prog,
share_prog=compiled_train_prog)
if args.vdl_dir:
from visualdl import LogWriter
vdl_writer = LogWriter(args.vdl_dir)
else:
vdl_writer = None
for epoch_id in range(config.epochs):
# 1. train with train dataset
program.run(train_dataloader, exe, compiled_train_prog, train_fetchs,
epoch_id, 'train', vdl_writer)
# 2. validate with validate dataset
if config.validate and epoch_id % config.valid_interval == 0:
if config.get('use_ema'):
logger.info(logger.coloring("EMA validate start..."))
with ema.apply(exe):
top1_acc = program.run(valid_dataloader, exe,
compiled_valid_prog, valid_fetchs,
epoch_id, 'valid')
logger.info(logger.coloring("EMA validate over!"))
top1_acc = program.run(valid_dataloader, exe, compiled_valid_prog,
valid_fetchs, epoch_id, 'valid')
if top1_acc > best_top1_acc:
best_top1_acc = top1_acc
message = "The best top1 acc {:.5f}, in epoch: {:d}".format(
best_top1_acc, epoch_id)
logger.info("{:s}".format(logger.coloring(message, "RED")))
if epoch_id % config.save_interval == 0:
model_path = os.path.join(config.model_save_dir,
config.ARCHITECTURE["name"])
save_model(train_prog, model_path,
"best_model_in_epoch_" + str(epoch_id))
# 3. save the persistable model
if epoch_id % config.save_interval == 0:
model_path = os.path.join(config.model_save_dir,
config.ARCHITECTURE["name"])
save_model(train_prog, model_path, epoch_id)
def main(args):
paddle.seed(12345)
config = get_config(args.config, overrides=args.override, show=True)
# assign the place
use_gpu = config.get("use_gpu", True)
place = paddle.set_device('gpu' if use_gpu else 'cpu')
trainer_num = paddle.distributed.get_world_size()
use_data_parallel = trainer_num != 1
config["use_data_parallel"] = use_data_parallel
if config["use_data_parallel"]:
paddle.distributed.init_parallel_env()
net = program.create_model(config.ARCHITECTURE, config.classes_num)
optimizer, lr_scheduler = program.create_optimizer(
config, parameter_list=net.parameters())
if config["use_data_parallel"]:
net = paddle.DataParallel(net)
# load model from checkpoint or pretrained model
init_model(config, net, optimizer)
train_dataloader = Reader(config, 'train', places=place)()
if config.validate:
valid_dataloader = Reader(config, 'valid', places=place)()
last_epoch_id = config.get("last_epoch", -1)
best_top1_acc = 0.0 # best top1 acc record
best_top1_epoch = last_epoch_id
for epoch_id in range(last_epoch_id + 1, config.epochs):
net.train()
# 1. train with train dataset
program.run(train_dataloader, config, net, optimizer, lr_scheduler,
epoch_id, 'train')
# 2. validate with validate dataset
if config.validate and epoch_id % config.valid_interval == 0:
net.eval()
with paddle.no_grad():
top1_acc = program.run(valid_dataloader, config, net, None,
None, epoch_id, 'valid')
if top1_acc > best_top1_acc:
best_top1_acc = top1_acc
best_top1_epoch = epoch_id
if epoch_id % config.save_interval == 0:
model_path = os.path.join(config.model_save_dir,
config.ARCHITECTURE["name"])
save_model(net, optimizer, model_path, "best_model")
message = "The best top1 acc {:.5f}, in epoch: {:d}".format(
best_top1_acc, best_top1_epoch)
logger.info("{:s}".format(logger.coloring(message, "RED")))
# 3. save the persistable model
if epoch_id % config.save_interval == 0:
model_path = os.path.join(config.model_save_dir,
config.ARCHITECTURE["name"])
save_model(net, optimizer, model_path, epoch_id)
def test_returns_hello(self):
self.assertEqual(run(), "hello")
def main(args):
role = role_maker.PaddleCloudRoleMaker(is_collective=True)
fleet.init(role)
config = get_config(args.config, overrides=args.override, show=True)
# assign the place
gpu_id = int(os.environ.get('FLAGS_selected_gpus', 0))
place = fluid.CUDAPlace(gpu_id)
# startup_prog is used to do some parameter init work,
# and train prog is used to hold the network
startup_prog = fluid.Program()
train_prog = fluid.Program()
best_top1_acc = 0.0 # best top1 acc record
if not config.get('use_ema'):
train_dataloader, train_fetchs = program.build(config,
train_prog,
startup_prog,
is_train=True)
else:
train_dataloader, train_fetchs, ema = program.build(config,
train_prog,
startup_prog,
is_train=True)
if config.validate:
valid_prog = fluid.Program()
valid_dataloader, valid_fetchs = program.build(config,
valid_prog,
startup_prog,
is_train=False)
# clone to prune some content which is irrelevant in valid_prog
valid_prog = valid_prog.clone(for_test=True)
# create the "Executor" with the statement of which place
exe = fluid.Executor(place)
# Parameter initialization
exe.run(startup_prog)
# load model from 1. checkpoint to resume training, 2. pretrained model to finetune
init_model(config, train_prog, exe)
train_reader = Reader(config, 'train')()
train_dataloader.set_sample_list_generator(train_reader, place)
if config.validate:
valid_reader = Reader(config, 'valid')()
valid_dataloader.set_sample_list_generator(valid_reader, place)
compiled_valid_prog = program.compile(config, valid_prog)
compiled_train_prog = fleet.main_program
vdl_writer = LogWriter(args.vdl_dir) if args.vdl_dir else None
for epoch_id in range(config.epochs):
# 1. train with train dataset
program.run(train_dataloader, exe, compiled_train_prog, train_fetchs,
epoch_id, 'train', vdl_writer)
if int(os.getenv("PADDLE_TRAINER_ID", 0)) == 0:
# 2. validate with validate dataset
if config.validate and epoch_id % config.valid_interval == 0:
if config.get('use_ema'):
logger.info(logger.coloring("EMA validate start..."))
with train_fetchs('ema').apply(exe):
top1_acc = program.run(valid_dataloader, exe,
compiled_valid_prog,
valid_fetchs, epoch_id, 'valid')
logger.info(logger.coloring("EMA validate over!"))
top1_acc = program.run(valid_dataloader, exe,
compiled_valid_prog, valid_fetchs,
epoch_id, 'valid')
if top1_acc > best_top1_acc:
best_top1_acc = top1_acc
message = "The best top1 acc {:.5f}, in epoch: {:d}".format(
best_top1_acc, epoch_id)
logger.info("{:s}".format(logger.coloring(message, "RED")))
if epoch_id % config.save_interval == 0:
model_path = os.path.join(config.model_save_dir,
config.ARCHITECTURE["name"])
save_model(train_prog, model_path,
"best_model_in_epoch_" + str(epoch_id))
# 3. save the persistable model
if epoch_id % config.save_interval == 0:
model_path = os.path.join(config.model_save_dir,
config.ARCHITECTURE["name"])
save_model(train_prog, model_path, epoch_id)
def logout():
program.run()
import program
if __name__ == "__main__":
program.run()
from program import run
if __name__ == "__main__":
run()
def main(args):
role = role_maker.PaddleCloudRoleMaker(is_collective=True)
fleet.init(role)
config = get_config(args.config, overrides=args.override, show=True)
use_fp16 = config.get('use_fp16', False)
if use_fp16:
AMP_RELATED_FLAGS_SETTING = {
'FLAGS_cudnn_exhaustive_search': 1,
'FLAGS_conv_workspace_size_limit': 4000,
'FLAGS_cudnn_batchnorm_spatial_persistent': 1,
'FLAGS_max_inplace_grad_add': 8,
}
os.environ['FLAGS_cudnn_batchnorm_spatial_persistent'] = '1'
paddle.fluid.set_flags(AMP_RELATED_FLAGS_SETTING)
# assign the place
gpu_id = int(os.environ.get('FLAGS_selected_gpus', 0))
place = fluid.CUDAPlace(gpu_id)
# startup_prog is used to do some parameter init work,
# and train prog is used to hold the network
startup_prog = fluid.Program()
train_prog = fluid.Program()
best_top1_acc = 0.0 # best top1 acc record
if not config.get('use_ema'):
train_dataloader, train_fetchs = program.build(
config, train_prog, startup_prog, is_train=True)
else:
train_dataloader, train_fetchs, ema = program.build(
config, train_prog, startup_prog, is_train=True)
if config.validate:
valid_prog = fluid.Program()
valid_dataloader, valid_fetchs = program.build(
config, valid_prog, startup_prog, is_train=False)
# clone to prune some content which is irrelevant in valid_prog
valid_prog = valid_prog.clone(for_test=True)
# create the "Executor" with the statement of which place
exe = fluid.Executor(place)
# Parameter initialization
exe.run(startup_prog)
# load model from 1. checkpoint to resume training, 2. pretrained model to finetune
init_model(config, train_prog, exe)
if not config.get('use_dali', False):
train_reader = Reader(config, 'train')()
train_dataloader.set_sample_list_generator(train_reader, place)
if config.validate:
valid_reader = Reader(config, 'valid')()
valid_dataloader.set_sample_list_generator(valid_reader, place)
compiled_valid_prog = program.compile(config, valid_prog)
else:
import dali
train_dataloader = dali.train(config)
if config.validate and int(os.getenv("PADDLE_TRAINER_ID", 0)):
if int(os.getenv("PADDLE_TRAINER_ID", 0)) == 0:
valid_dataloader = dali.val(config)
compiled_valid_prog = program.compile(config, valid_prog)
compiled_train_prog = fleet.main_program
vdl_writer = None
if args.vdl_dir:
if version_info.major == 2:
logger.info(
"visualdl is just supported for python3, so it is disabled in python2..."
)
else:
from visualdl import LogWriter
vdl_writer = LogWriter(args.vdl_dir)
for epoch_id in range(config.epochs):
# 1. train with train dataset
program.run(train_dataloader, exe, compiled_train_prog, train_fetchs,
epoch_id, 'train', config, vdl_writer)
if int(os.getenv("PADDLE_TRAINER_ID", 0)) == 0:
# 2. validate with validate dataset
if config.validate and epoch_id % config.valid_interval == 0:
if config.get('use_ema'):
logger.info(logger.coloring("EMA validate start..."))
with ema.apply(exe):
top1_acc = program.run(
valid_dataloader, exe, compiled_valid_prog,
valid_fetchs, epoch_id, 'valid', config)
logger.info(logger.coloring("EMA validate over!"))
top1_acc = program.run(valid_dataloader, exe,
compiled_valid_prog, valid_fetchs,
epoch_id, 'valid', config)
if top1_acc > best_top1_acc:
best_top1_acc = top1_acc
message = "The best top1 acc {:.5f}, in epoch: {:d}".format(
best_top1_acc, epoch_id)
logger.info("{:s}".format(logger.coloring(message, "RED")))
if epoch_id % config.save_interval == 0:
model_path = os.path.join(config.model_save_dir,
config.ARCHITECTURE["name"])
save_model(train_prog, model_path, "best_model")
# 3. save the persistable model
if epoch_id % config.save_interval == 0:
model_path = os.path.join(config.model_save_dir,
config.ARCHITECTURE["name"])
save_model(train_prog, model_path, epoch_id)
