def __init__(self,
models=None,
pretrained_list=None,
freeze_params_list=None,
**kargs):
super().__init__()
assert isinstance(models, list)
self.model_list = []
self.model_name_list = []
if pretrained_list is not None:
assert len(pretrained_list) == len(models)
if freeze_params_list is None:
freeze_params_list = [False] * len(models)
assert len(freeze_params_list) == len(models)
for idx, model_config in enumerate(models):
assert len(model_config) == 1
key = list(model_config.keys())[0]
model_config = model_config[key]
model_name = model_config.pop("name")
model = eval(model_name)(**model_config)
if freeze_params_list[idx]:
for param in model.parameters():
param.trainable = False
self.model_list.append(self.add_sublayer(key, model))
self.model_name_list.append(key)
if pretrained_list is not None:
for idx, pretrained in enumerate(pretrained_list):
if pretrained is not None:
load_dygraph_pretrain(
self.model_name_list[idx], path=pretrained)
def main():
args = parse_args()
operators = create_operators(args.interpolation)
# assign the place
place = 'gpu:{}'.format(ParallelEnv().dev_id) if args.use_gpu else 'cpu'
place = paddle.set_device(place)
net = ResNet50()
load_dygraph_pretrain(net, args.pretrained_model)
img = cv2.imread(args.image_file, cv2.IMREAD_COLOR)
data = preprocess(img, operators)
data = np.expand_dims(data, axis=0)
data = paddle.to_tensor(data)
net.eval()
_, fm = net(data)
assert args.channel_num >= 0 and args.channel_num <= fm.shape[
1], "the channel is out of the range, should be in {} but got {}".format(
[0, fm.shape[1]], args.channel_num)
fm = (np.squeeze(fm[0][args.channel_num].numpy()) * 255).astype(np.uint8)
fm = cv2.resize(fm, (img.shape[1], img.shape[0]))
if args.save_path is not None:
print("the feature map is saved in path: {}".format(args.save_path))
cv2.imwrite(args.save_path, fm)
def main():
args = parse_args()
net = architectures.__dict__[args.model]
model = Net(net, args.class_dim, args.model)
# get QAT model
quant_config = get_default_quant_config()
# TODO(littletomatodonkey): add PACT for export model
# quant_config["activation_preprocess_type"] = "PACT"
quanter = QAT(config=quant_config)
quanter.quantize(model)
load_dygraph_pretrain(model.pre_net,
path=args.pretrained_model,
load_static_weights=args.load_static_weights)
model.eval()
save_path = os.path.join(args.output_path, "inference")
quanter.save_quantized_model(model,
save_path,
input_spec=[
paddle.static.InputSpec(shape=[
None, 3, args.img_size, args.img_size
],
dtype='float32')
])
print('inference QAT model is saved to {}'.format(save_path))
def export(self):
assert self.mode == "export"
use_multilabel = self.config["Global"].get("use_multilabel", False)
model = ExportModel(self.config["Arch"], self.model, use_multilabel)
if self.config["Global"]["pretrained_model"] is not None:
load_dygraph_pretrain(model.base_model,
self.config["Global"]["pretrained_model"])
model.eval()
save_path = os.path.join(self.config["Global"]["save_inference_dir"],
"inference")
if model.quanter:
model.quanter.save_quantized_model(
model.base_model,
save_path,
input_spec=[
paddle.static.InputSpec(
shape=[None] + self.config["Global"]["image_shape"],
dtype='float32')
])
else:
model = paddle.jit.to_static(
model,
input_spec=[
paddle.static.InputSpec(
shape=[None] + self.config["Global"]["image_shape"],
dtype='float32')
])
paddle.jit.save(model, save_path)
def _load_pretrained(pretrained, model, model_url, use_ssld):
if pretrained is False:
pass
elif pretrained is True:
load_dygraph_pretrain_from_url(model, model_url, use_ssld=use_ssld)
elif isinstance(pretrained, str):
load_dygraph_pretrain(model, pretrained)
else:
raise RuntimeError(
"pretrained type is not available. Please use `string` or `boolean` type."
)
def main():
args = parse_args()
net = architectures.__dict__[args.model]
model = Net(net, args.class_dim, args.model)
load_dygraph_pretrain(model.pre_net,
path=args.pretrained_model,
load_static_weights=args.load_static_weights)
model.eval()
model = to_static(model,
input_spec=[
paddle.static.InputSpec(
shape=[None, 3, args.img_size, args.img_size],
dtype='float32')
])
paddle.jit.save(model, os.path.join(args.output_path, "inference"))
def export_fuse_model(configs):
slim_config = configs["Slim"].copy()
configs["Slim"] = None
fuse_model = build_model(configs)
fuse_model.head = GalleryLayer(configs)
configs["Slim"] = slim_config
quantize_model(configs, fuse_model)
load_dygraph_pretrain(fuse_model, configs["Global"]["pretrained_model"])
fuse_model.eval()
fuse_model.head.build_gallery_layer(fuse_model)
save_path = configs["Global"]["save_inference_dir"]
fuse_model.quanter.save_quantized_model(
fuse_model,
save_path,
input_spec=[
paddle.static.InputSpec(shape=[None] +
configs["Global"]["image_shape"],
dtype='float32')
])
def main():
args = utils.parse_args()
# assign the place
place = paddle.set_device('gpu' if args.use_gpu else 'cpu')
net = architectures.__dict__[args.model](class_dim=args.class_num)
load_dygraph_pretrain(net, args.pretrained_model, args.load_static_weights)
image_list = get_image_list(args.image_file)
for idx, filename in enumerate(image_list):
img = cv2.imread(filename)[:, :, ::-1]
data = utils.preprocess(img, args)
data = np.expand_dims(data, axis=0)
data = paddle.to_tensor(data)
net.eval()
outputs = net(data)
if args.model == "GoogLeNet":
outputs = outputs[0]
outputs = F.softmax(outputs)
outputs = outputs.numpy()
probs = postprocess(outputs)
top1_class_id = 0
rank = 1
print("Current image file: {}".format(filename))
for idx, prob in probs:
print("\ttop{:d}, class id: {:d}, probability: {:.4f}".format(
rank, idx, prob))
if rank == 1:
top1_class_id = idx
rank += 1
if args.pre_label_image:
save_prelabel_results(top1_class_id, filename,
args.pre_label_out_idr)
return
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 __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"])
