def init_reader(self):
if fleet.is_server():
return
self.reader, self.file_list = get_reader(self.input_data, config)
self.example_nums = 0
self.count_method = self.config.get(
"static_benchmark.example_count_method", "example")
if self.count_method == "example":
self.example_nums = get_example_num(self.file_list)
elif self.count_method == "word":
self.example_nums = get_word_num(self.file_list)
else:
raise ValueError(
"Set static_benchmark.example_count_method for example / word for example count."
)
def main(args):
paddle.seed(12345)
# load config
config = load_yaml(args.config_yaml)
config["yaml_path"] = args.config_yaml
config["config_abs_dir"] = args.abs_dir
# load static model class
static_model_class = load_static_model_class(config)
input_data = static_model_class.create_feeds()
input_data_names = [data.name for data in input_data]
fetch_vars = static_model_class.net(input_data)
#infer_target_var = model.infer_target_var
logger.info("cpu_num: {}".format(os.getenv("CPU_NUM")))
static_model_class.create_optimizer()
use_gpu = config.get("runner.use_gpu", True)
use_auc = config.get("runner.use_auc", False)
auc_num = config.get("runner.auc_num", 1)
train_data_dir = config.get("runner.train_data_dir", None)
epochs = config.get("runner.epochs", None)
print_interval = config.get("runner.print_interval", None)
model_save_path = config.get("runner.model_save_path", "model_output")
model_init_path = config.get("runner.model_init_path", None)
batch_size = config.get("runner.train_batch_size", None)
reader_type = config.get("runner.reader_type", "DataLoader")
os.environ["CPU_NUM"] = str(config.get("runner.thread_num", 1))
logger.info("**************common.configs**********")
logger.info(
"use_gpu: {}, train_data_dir: {}, epochs: {}, print_interval: {}, model_save_path: {}"
.format(use_gpu, train_data_dir, epochs, print_interval,
model_save_path))
logger.info("**************common.configs**********")
place = paddle.set_device('gpu' if use_gpu else 'cpu')
exe = paddle.static.Executor(place)
# initialize
exe.run(paddle.static.default_startup_program())
last_epoch_id = config.get("last_epoch", -1)
if reader_type == 'QueueDataset':
dataset, file_list = get_reader(input_data, config)
elif reader_type == 'DataLoader':
train_dataloader = create_data_loader(config=config, place=place)
for epoch_id in range(last_epoch_id + 1, epochs):
epoch_begin = time.time()
if use_auc:
reset_auc(auc_num)
if reader_type == 'DataLoader':
fetch_batch_var = dataloader_train(epoch_id, train_dataloader,
input_data_names, fetch_vars,
exe, config)
metric_str = ""
for var_idx, var_name in enumerate(fetch_vars):
metric_str += "{}: {}, ".format(var_name,
fetch_batch_var[var_idx])
logger.info("epoch: {} done, ".format(epoch_id) + metric_str +
"epoch time: {:.2f} s".format(time.time() -
epoch_begin))
elif reader_type == 'QueueDataset':
fetch_batch_var = dataset_train(epoch_id, dataset, fetch_vars, exe,
config)
logger.info("epoch: {} done, ".format(epoch_id) +
"epoch time: {:.2f} s".format(time.time() -
epoch_begin))
else:
logger.info("reader type wrong")
save_static_model(paddle.static.default_main_program(),
model_save_path,
epoch_id,
prefix='rec_static')
def main(args):
paddle.seed(12345)
# load config
config = load_yaml(args.config_yaml)
config["yaml_path"] = args.config_yaml
config["config_abs_dir"] = args.abs_dir
# modify config from command
if args.opt:
for parameter in args.opt:
parameter = parameter.strip()
key, value = parameter.split("=")
config[key] = value
# load static model class
static_model_class = load_static_model_class(config)
input_data = static_model_class.create_feeds()
input_data_names = [data.name for data in input_data]
fetch_vars = static_model_class.net(input_data)
#infer_target_var = model.infer_target_var
logger.info("cpu_num: {}".format(os.getenv("CPU_NUM")))
static_model_class.create_optimizer()
use_gpu = config.get("runner.use_gpu", True)
use_auc = config.get("runner.use_auc", False)
use_visual = config.get("runner.use_visual", False)
use_inference = config.get("runner.use_inference", False)
auc_num = config.get("runner.auc_num", 1)
train_data_dir = config.get("runner.train_data_dir", None)
epochs = config.get("runner.epochs", None)
print_interval = config.get("runner.print_interval", None)
model_save_path = config.get("runner.model_save_path", "model_output")
model_init_path = config.get("runner.model_init_path", None)
batch_size = config.get("runner.train_batch_size", None)
reader_type = config.get("runner.reader_type", "DataLoader")
os.environ["CPU_NUM"] = str(config.get("runner.thread_num", 1))
logger.info("**************common.configs**********")
logger.info(
"use_gpu: {}, use_visual: {}, train_batch_size: {}, train_data_dir: {}, epochs: {}, print_interval: {}, model_save_path: {}"
.format(use_gpu, use_visual, batch_size, train_data_dir, epochs,
print_interval, model_save_path))
logger.info("**************common.configs**********")
place = paddle.set_device('gpu' if use_gpu else 'cpu')
exe = paddle.static.Executor(place)
# initialize
exe.run(paddle.static.default_startup_program())
last_epoch_id = config.get("last_epoch", -1)
# Create a log_visual object and store the data in the path
if use_visual:
from visualdl import LogWriter
log_visual = LogWriter(args.abs_dir + "/visualDL_log/train")
else:
log_visual = None
step_num = 0
if reader_type == 'QueueDataset':
dataset, file_list = get_reader(input_data, config)
elif reader_type == 'DataLoader':
train_dataloader = create_data_loader(config=config, place=place)
for epoch_id in range(last_epoch_id + 1, epochs):
epoch_begin = time.time()
if use_auc:
reset_auc(auc_num)
if reader_type == 'DataLoader':
fetch_batch_var, step_num = dataloader_train(
epoch_id, train_dataloader, input_data_names, fetch_vars, exe,
config, use_visual, log_visual, step_num)
metric_str = ""
for var_idx, var_name in enumerate(fetch_vars):
metric_str += "{}: {}, ".format(var_name,
fetch_batch_var[var_idx])
logger.info("epoch: {} done, ".format(epoch_id) + metric_str +
"epoch time: {:.2f} s".format(time.time() -
epoch_begin))
elif reader_type == 'QueueDataset':
fetch_batch_var = dataset_train(epoch_id, dataset, fetch_vars, exe,
config)
logger.info("epoch: {} done, ".format(epoch_id) +
"epoch time: {:.2f} s".format(time.time() -
epoch_begin))
else:
logger.info("reader type wrong")
save_static_model(paddle.static.default_main_program(),
model_save_path,
epoch_id,
prefix='rec_static')
if use_inference:
feed_var_names = config.get("runner.save_inference_feed_varnames",
[])
feedvars = []
fetch_var_names = config.get(
"runner.save_inference_fetch_varnames", [])
fetchvars = []
for var_name in feed_var_names:
if var_name not in paddle.static.default_main_program(
).global_block().vars:
raise ValueError(
"Feed variable: {} not in default_main_program, global block has follow vars: {}"
.format(
var_name,
paddle.static.default_main_program().global_block(
).vars.keys()))
else:
feedvars.append(paddle.static.default_main_program().
global_block().vars[var_name])
for var_name in fetch_var_names:
if var_name not in paddle.static.default_main_program(
).global_block().vars:
raise ValueError(
"Fetch variable: {} not in default_main_program, global block has follow vars: {}"
.format(
var_name,
paddle.static.default_main_program().global_block(
).vars.keys()))
else:
fetchvars.append(paddle.static.default_main_program().
global_block().vars[var_name])
save_inference_model(model_save_path, epoch_id, feedvars,
fetchvars, exe)
