def build(config,
main_prog,
startup_prog,
class_num=None,
step_each_epoch=100,
is_train=True,
is_distributed=True):
"""
Build a program using a model and an optimizer
1. create feeds
2. create a dataloader
3. create a model
4. create fetchs
5. create an optimizer
Args:
config(dict): config
main_prog(): main program
startup_prog(): startup program
class_num(int): the class number of network, required if use_mix
is_train(bool): train or eval
is_distributed(bool): whether to use distributed training method
Returns:
dataloader(): a bridge between the model and the data
fetchs(dict): dict of model outputs(included loss and measures)
"""
with paddle.static.program_guard(main_prog, startup_prog):
with paddle.utils.unique_name.guard():
mode = "Train" if is_train else "Eval"
use_mix = "batch_transform_ops" in config["DataLoader"][mode][
"dataset"]
feeds = create_feeds(config["Global"]["image_shape"],
use_mix,
class_num=class_num,
dtype="float32")
# build model
# data_format should be assigned in arch-dict
input_image_channel = config["Global"]["image_shape"][
0] # default as [3, 224, 224]
model = build_model(config)
out = model(feeds["data"])
# end of build model
fetchs = create_fetchs(out,
feeds,
config["Arch"],
epsilon=config.get('ls_epsilon'),
class_num=class_num,
use_mix=use_mix,
config=config,
mode=mode)
lr_scheduler = None
optimizer = None
if is_train:
optimizer, lr_scheduler = build_optimizer(
config["Optimizer"], config["Global"]["epochs"],
step_each_epoch)
optimizer = mixed_precision_optimizer(config, optimizer)
if is_distributed:
optimizer = dist_optimizer(config, optimizer)
optimizer.minimize(fetchs['loss'][0])
return fetchs, lr_scheduler, feeds, optimizer
def __init__(self, config, mode="train"):
assert mode in ["train", "eval", "infer", "export"]
self.mode = mode
self.config = config
self.eval_mode = self.config["Global"].get("eval_mode",
"classification")
if "Head" in self.config["Arch"] or self.config["Arch"].get(
"is_rec", False):
self.is_rec = True
else:
self.is_rec = False
# set seed
seed = self.config["Global"].get("seed", False)
if seed or seed == 0:
assert isinstance(seed, int), "The 'seed' must be a integer!"
paddle.seed(seed)
np.random.seed(seed)
random.seed(seed)
# init logger
self.output_dir = self.config['Global']['output_dir']
log_file = os.path.join(self.output_dir, self.config["Arch"]["name"],
f"{mode}.log")
init_logger(log_file=log_file)
print_config(config)
# init train_func and eval_func
assert self.eval_mode in ["classification", "retrieval"], logger.error(
"Invalid eval mode: {}".format(self.eval_mode))
self.train_epoch_func = train_epoch
self.eval_func = getattr(evaluation, self.eval_mode + "_eval")
self.use_dali = self.config['Global'].get("use_dali", False)
# for visualdl
self.vdl_writer = None
if self.config['Global'][
'use_visualdl'] and mode == "train" and dist.get_rank() == 0:
vdl_writer_path = os.path.join(self.output_dir, "vdl")
if not os.path.exists(vdl_writer_path):
os.makedirs(vdl_writer_path)
self.vdl_writer = LogWriter(logdir=vdl_writer_path)
# set device
assert self.config["Global"]["device"] in [
"cpu", "gpu", "xpu", "npu", "mlu"
]
self.device = paddle.set_device(self.config["Global"]["device"])
logger.info('train with paddle {} and device {}'.format(
paddle.__version__, self.device))
# AMP training
self.amp = True if "AMP" in self.config and self.mode == "train" else False
if self.amp and self.config["AMP"] is not None:
self.scale_loss = self.config["AMP"].get("scale_loss", 1.0)
self.use_dynamic_loss_scaling = self.config["AMP"].get(
"use_dynamic_loss_scaling", False)
else:
self.scale_loss = 1.0
self.use_dynamic_loss_scaling = False
if self.amp:
AMP_RELATED_FLAGS_SETTING = {
'FLAGS_max_inplace_grad_add': 8,
}
if paddle.is_compiled_with_cuda():
AMP_RELATED_FLAGS_SETTING.update(
{'FLAGS_cudnn_batchnorm_spatial_persistent': 1})
paddle.fluid.set_flags(AMP_RELATED_FLAGS_SETTING)
if "class_num" in config["Global"]:
global_class_num = config["Global"]["class_num"]
if "class_num" not in config["Arch"]:
config["Arch"]["class_num"] = global_class_num
msg = f"The Global.class_num will be deprecated. Please use Arch.class_num instead. Arch.class_num has been set to {global_class_num}."
else:
msg = "The Global.class_num will be deprecated. Please use Arch.class_num instead. The Global.class_num has been ignored."
logger.warning(msg)
#TODO(gaotingquan): support rec
class_num = config["Arch"].get("class_num", None)
self.config["DataLoader"].update({"class_num": class_num})
# build dataloader
if self.mode == 'train':
self.train_dataloader = build_dataloader(self.config["DataLoader"],
"Train", self.device,
self.use_dali)
if self.mode == "eval" or (self.mode == "train" and
self.config["Global"]["eval_during_train"]):
if self.eval_mode == "classification":
self.eval_dataloader = build_dataloader(
self.config["DataLoader"], "Eval", self.device,
self.use_dali)
elif self.eval_mode == "retrieval":
self.gallery_query_dataloader = None
if len(self.config["DataLoader"]["Eval"].keys()) == 1:
key = list(self.config["DataLoader"]["Eval"].keys())[0]
self.gallery_query_dataloader = build_dataloader(
self.config["DataLoader"]["Eval"], key, self.device,
self.use_dali)
else:
self.gallery_dataloader = build_dataloader(
self.config["DataLoader"]["Eval"], "Gallery",
self.device, self.use_dali)
self.query_dataloader = build_dataloader(
self.config["DataLoader"]["Eval"], "Query",
self.device, self.use_dali)
# build loss
if self.mode == "train":
loss_info = self.config["Loss"]["Train"]
self.train_loss_func = build_loss(loss_info)
if self.mode == "eval" or (self.mode == "train" and
self.config["Global"]["eval_during_train"]):
loss_config = self.config.get("Loss", None)
if loss_config is not None:
loss_config = loss_config.get("Eval")
if loss_config is not None:
self.eval_loss_func = build_loss(loss_config)
else:
self.eval_loss_func = None
else:
self.eval_loss_func = None
# build metric
if self.mode == 'train':
metric_config = self.config.get("Metric")
if metric_config is not None:
metric_config = metric_config.get("Train")
if metric_config is not None:
if hasattr(
self.train_dataloader, "collate_fn"
) and self.train_dataloader.collate_fn is not None:
for m_idx, m in enumerate(metric_config):
if "TopkAcc" in m:
msg = f"'TopkAcc' metric can not be used when setting 'batch_transform_ops' in config. The 'TopkAcc' metric has been removed."
logger.warning(msg)
break
metric_config.pop(m_idx)
self.train_metric_func = build_metrics(metric_config)
else:
self.train_metric_func = None
else:
self.train_metric_func = None
if self.mode == "eval" or (self.mode == "train" and
self.config["Global"]["eval_during_train"]):
metric_config = self.config.get("Metric")
if self.eval_mode == "classification":
if metric_config is not None:
metric_config = metric_config.get("Eval")
if metric_config is not None:
self.eval_metric_func = build_metrics(metric_config)
elif self.eval_mode == "retrieval":
if metric_config is None:
metric_config = [{"name": "Recallk", "topk": (1, 5)}]
else:
metric_config = metric_config["Eval"]
self.eval_metric_func = build_metrics(metric_config)
else:
self.eval_metric_func = None
# build model
self.model = build_model(self.config)
# set @to_static for benchmark, skip this by default.
apply_to_static(self.config, self.model)
# load_pretrain
if self.config["Global"]["pretrained_model"] is not None:
if self.config["Global"]["pretrained_model"].startswith("http"):
load_dygraph_pretrain_from_url(
self.model, self.config["Global"]["pretrained_model"])
else:
load_dygraph_pretrain(
self.model, self.config["Global"]["pretrained_model"])
# build optimizer
if self.mode == 'train':
self.optimizer, self.lr_sch = build_optimizer(
self.config["Optimizer"], self.config["Global"]["epochs"],
len(self.train_dataloader), [self.model])
# for amp training
if self.amp:
self.scaler = paddle.amp.GradScaler(
init_loss_scaling=self.scale_loss,
use_dynamic_loss_scaling=self.use_dynamic_loss_scaling)
amp_level = self.config['AMP'].get("level", "O1")
if amp_level not in ["O1", "O2"]:
msg = "[Parameter Error]: The optimize level of AMP only support 'O1' and 'O2'. The level has been set 'O1'."
logger.warning(msg)
self.config['AMP']["level"] = "O1"
amp_level = "O1"
self.model, self.optimizer = paddle.amp.decorate(
models=self.model,
optimizers=self.optimizer,
level=amp_level,
save_dtype='float32')
# for distributed
world_size = dist.get_world_size()
self.config["Global"]["distributed"] = world_size != 1
if world_size != 4 and self.mode == "train":
msg = f"The training strategy in config files provided by PaddleClas is based on 4 gpus. But the number of gpus is {world_size} in current training. Please modify the stategy (learning rate, batch size and so on) if use config files in PaddleClas to train."
logger.warning(msg)
if self.config["Global"]["distributed"]:
dist.init_parallel_env()
self.model = paddle.DataParallel(self.model)
# build postprocess for infer
if self.mode == 'infer':
self.preprocess_func = create_operators(
self.config["Infer"]["transforms"])
self.postprocess_func = build_postprocess(
self.config["Infer"]["PostProcess"])
def create_optimizer(config, step_each_epoch):
# create learning_rate instance
optimizer, lr_sch = build_optimizer(config["Optimizer"],
config["Global"]["epochs"],
step_each_epoch)
return optimizer, lr_sch
def __init__(self, config, mode="train"):
assert mode in ["train", "eval", "infer", "export"]
self.mode = mode
self.config = config
self.eval_mode = self.config["Global"].get("eval_mode",
"classification")
if "Head" in self.config["Arch"]:
self.is_rec = True
else:
self.is_rec = False
# set seed
seed = self.config["Global"].get("seed", False)
if seed or seed == 0:
assert isinstance(seed, int), "The 'seed' must be a integer!"
paddle.seed(seed)
np.random.seed(seed)
random.seed(seed)
# init logger
self.output_dir = self.config['Global']['output_dir']
log_file = os.path.join(self.output_dir, self.config["Arch"]["name"],
f"{mode}.log")
init_logger(name='root', log_file=log_file)
print_config(config)
# init train_func and eval_func
assert self.eval_mode in ["classification", "retrieval"], logger.error(
"Invalid eval mode: {}".format(self.eval_mode))
self.train_epoch_func = train_epoch
self.eval_func = getattr(evaluation, self.eval_mode + "_eval")
self.use_dali = self.config['Global'].get("use_dali", False)
# for visualdl
self.vdl_writer = None
if self.config['Global']['use_visualdl'] and mode == "train":
vdl_writer_path = os.path.join(self.output_dir, "vdl")
if not os.path.exists(vdl_writer_path):
os.makedirs(vdl_writer_path)
self.vdl_writer = LogWriter(logdir=vdl_writer_path)
# set device
assert self.config["Global"]["device"] in ["cpu", "gpu", "xpu", "npu"]
self.device = paddle.set_device(self.config["Global"]["device"])
logger.info('train with paddle {} and device {}'.format(
paddle.__version__, self.device))
# AMP training
self.amp = True if "AMP" in self.config else False
if self.amp and self.config["AMP"] is not None:
self.scale_loss = self.config["AMP"].get("scale_loss", 1.0)
self.use_dynamic_loss_scaling = self.config["AMP"].get(
"use_dynamic_loss_scaling", False)
else:
self.scale_loss = 1.0
self.use_dynamic_loss_scaling = False
if self.amp:
AMP_RELATED_FLAGS_SETTING = {
'FLAGS_cudnn_batchnorm_spatial_persistent': 1,
'FLAGS_max_inplace_grad_add': 8,
}
paddle.fluid.set_flags(AMP_RELATED_FLAGS_SETTING)
#TODO(gaotingquan): support rec
class_num = config["Arch"].get("class_num", None)
self.config["DataLoader"].update({"class_num": class_num})
# build dataloader
if self.mode == 'train':
self.train_dataloader = build_dataloader(self.config["DataLoader"],
"Train", self.device,
self.use_dali)
if self.mode == "eval" or (self.mode == "train" and
self.config["Global"]["eval_during_train"]):
if self.eval_mode == "classification":
self.eval_dataloader = build_dataloader(
self.config["DataLoader"], "Eval", self.device,
self.use_dali)
elif self.eval_mode == "retrieval":
self.gallery_query_dataloader = None
if len(self.config["DataLoader"]["Eval"].keys()) == 1:
key = list(self.config["DataLoader"]["Eval"].keys())[0]
self.gallery_query_dataloader = build_dataloader(
self.config["DataLoader"]["Eval"], key, self.device,
self.use_dali)
else:
self.gallery_dataloader = build_dataloader(
self.config["DataLoader"]["Eval"], "Gallery",
self.device, self.use_dali)
self.query_dataloader = build_dataloader(
self.config["DataLoader"]["Eval"], "Query",
self.device, self.use_dali)
# build loss
if self.mode == "train":
loss_info = self.config["Loss"]["Train"]
self.train_loss_func = build_loss(loss_info)
if self.mode == "eval" or (self.mode == "train" and
self.config["Global"]["eval_during_train"]):
loss_config = self.config.get("Loss", None)
if loss_config is not None:
loss_config = loss_config.get("Eval")
if loss_config is not None:
self.eval_loss_func = build_loss(loss_config)
else:
self.eval_loss_func = None
else:
self.eval_loss_func = None
# build metric
if self.mode == 'train':
metric_config = self.config.get("Metric")
if metric_config is not None:
metric_config = metric_config.get("Train")
if metric_config is not None:
self.train_metric_func = build_metrics(metric_config)
else:
self.train_metric_func = None
else:
self.train_metric_func = None
if self.mode == "eval" or (self.mode == "train" and
self.config["Global"]["eval_during_train"]):
metric_config = self.config.get("Metric")
if self.eval_mode == "classification":
if metric_config is not None:
metric_config = metric_config.get("Eval")
if metric_config is not None:
self.eval_metric_func = build_metrics(metric_config)
elif self.eval_mode == "retrieval":
if metric_config is None:
metric_config = [{"name": "Recallk", "topk": (1, 5)}]
else:
metric_config = metric_config["Eval"]
self.eval_metric_func = build_metrics(metric_config)
else:
self.eval_metric_func = None
# build model
self.model = build_model(self.config["Arch"])
# set @to_static for benchmark, skip this by default.
apply_to_static(self.config, self.model)
# for slim
self.pruner = get_pruner(self.config, self.model)
self.quanter = get_quaner(self.config, self.model)
# load_pretrain
if self.config["Global"]["pretrained_model"] is not None:
if self.config["Global"]["pretrained_model"].startswith("http"):
load_dygraph_pretrain_from_url(
self.model, self.config["Global"]["pretrained_model"])
else:
load_dygraph_pretrain(
self.model, self.config["Global"]["pretrained_model"])
# build optimizer
if self.mode == 'train':
self.optimizer, self.lr_sch = build_optimizer(
self.config["Optimizer"], self.config["Global"]["epochs"],
len(self.train_dataloader), [self.model])
# for distributed
self.config["Global"][
"distributed"] = paddle.distributed.get_world_size() != 1
if self.config["Global"]["distributed"]:
dist.init_parallel_env()
if self.config["Global"]["distributed"]:
self.model = paddle.DataParallel(self.model)
# build postprocess for infer
if self.mode == 'infer':
self.preprocess_func = create_operators(
self.config["Infer"]["transforms"])
self.postprocess_func = build_postprocess(
self.config["Infer"]["PostProcess"])