def ssd_eval(dataset_path, ckpt_path, anno_json):
"""SSD evaluation."""
batch_size = 1
ds = create_ssd_dataset(dataset_path,
batch_size=batch_size,
repeat_num=1,
is_training=False,
use_multiprocessing=False)
if config.model == "ssd300":
net = SSD300(ssd_mobilenet_v2(), config, is_training=False)
elif config.model == "ssd_vgg16":
net = ssd_vgg16(config=config)
elif config.model == "ssd_mobilenet_v1_fpn":
net = ssd_mobilenet_v1_fpn(config=config)
elif config.model == "ssd_resnet50_fpn":
net = ssd_resnet50_fpn(config=config)
else:
raise ValueError(f'config.model: {config.model} is not supported')
net = SsdInferWithDecoder(net, Tensor(default_boxes), config)
print("Load Checkpoint!")
param_dict = load_checkpoint(ckpt_path)
net.init_parameters_data()
load_param_into_net(net, param_dict)
net.set_train(False)
total = ds.get_dataset_size() * batch_size
print("\n========================================\n")
print("total images num: ", total)
print("Processing, please wait a moment.")
eval_param_dict = {"net": net, "dataset": ds, "anno_json": anno_json}
mAP = apply_eval(eval_param_dict)
print("\n========================================\n")
print(f"mAP: {mAP}")
def ssd_model_build(args_opt):
if config.model == "ssd300":
backbone = ssd_mobilenet_v2()
ssd = SSD300(backbone=backbone, config=config)
init_net_param(ssd)
if args_opt.freeze_layer == "backbone":
for param in backbone.feature_1.trainable_params():
param.requires_grad = False
elif config.model == "ssd_mobilenet_v1_fpn":
ssd = ssd_mobilenet_v1_fpn(config=config)
init_net_param(ssd)
if config.feature_extractor_base_param != "":
param_dict = load_checkpoint(config.feature_extractor_base_param)
for x in list(param_dict.keys()):
param_dict["network.feature_extractor.mobilenet_v1." + x] = param_dict[x]
del param_dict[x]
load_param_into_net(ssd.feature_extractor.mobilenet_v1.network, param_dict)
elif config.model == "ssd_resnet50_fpn":
ssd = ssd_resnet50_fpn(config=config)
init_net_param(ssd)
if config.feature_extractor_base_param != "":
param_dict = load_checkpoint(config.feature_extractor_base_param)
for x in list(param_dict.keys()):
param_dict["network.feature_extractor.resnet." + x] = param_dict[x]
del param_dict[x]
load_param_into_net(ssd.feature_extractor.resnet, param_dict)
else:
raise ValueError(f'config.model: {config.model} is not supported')
return ssd
def ssd_eval(dataset_path, ckpt_path, anno_json):
"""SSD evaluation."""
batch_size = 1
ds = create_ssd_dataset(dataset_path,
batch_size=batch_size,
repeat_num=1,
is_training=False,
use_multiprocessing=False)
if config.model == "ssd300":
net = SSD300(ssd_mobilenet_v2(), config, is_training=False)
elif config.model == "ssd_vgg16":
net = ssd_vgg16(config=config)
elif config.model == "ssd_mobilenet_v1_fpn":
net = ssd_mobilenet_v1_fpn(config=config)
elif config.model == "ssd_resnet50_fpn":
net = ssd_resnet50_fpn(config=config)
else:
raise ValueError(f'config.model: {config.model} is not supported')
net = SsdInferWithDecoder(net, Tensor(default_boxes), config)
print("Load Checkpoint!")
param_dict = load_checkpoint(ckpt_path)
net.init_parameters_data()
load_param_into_net(net, param_dict)
net.set_train(False)
i = batch_size
total = ds.get_dataset_size() * batch_size
start = time.time()
pred_data = []
print("\n========================================\n")
print("total images num: ", total)
print("Processing, please wait a moment.")
for data in ds.create_dict_iterator(output_numpy=True, num_epochs=1):
img_id = data['img_id']
img_np = data['image']
image_shape = data['image_shape']
output = net(Tensor(img_np))
for batch_idx in range(img_np.shape[0]):
pred_data.append({
"boxes": output[0].asnumpy()[batch_idx],
"box_scores": output[1].asnumpy()[batch_idx],
"img_id": int(np.squeeze(img_id[batch_idx])),
"image_shape": image_shape[batch_idx]
})
percent = round(i / total * 100., 2)
print(f' {str(percent)} [{i}/{total}]', end='\r')
i += batch_size
cost_time = int((time.time() - start) * 1000)
print(f' 100% [{total}/{total}] cost {cost_time} ms')
mAP = metrics(pred_data, anno_json)
print("\n========================================\n")
print(f"mAP: {mAP}")
def ssd_eval(dataset_path, ckpt_path):
"""SSD evaluation."""
batch_size = 1
ds = create_ssd_dataset(dataset_path,
batch_size=batch_size,
repeat_num=1,
is_training=False)
net = SSD300(ssd_mobilenet_v2(), config, is_training=False)
print("Load Checkpoint!")
param_dict = load_checkpoint(ckpt_path)
net.init_parameters_data()
load_param_into_net(net, param_dict)
net.set_train(False)
i = batch_size
total = ds.get_dataset_size() * batch_size
start = time.time()
pred_data = []
print("\n========================================\n")
print("total images num: ", total)
print("Processing, please wait a moment.")
for data in ds.create_dict_iterator(output_numpy=True):
img_id = data['img_id']
img_np = data['image']
image_shape = data['image_shape']
output = net(Tensor(img_np))
for batch_idx in range(img_np.shape[0]):
pred_data.append({
"boxes": output[0].asnumpy()[batch_idx],
"box_scores": output[1].asnumpy()[batch_idx],
"img_id": int(np.squeeze(img_id[batch_idx])),
"image_shape": image_shape[batch_idx]
})
percent = round(i / total * 100., 2)
print(f' {str(percent)} [{i}/{total}]', end='\r')
i += batch_size
cost_time = int((time.time() - start) * 1000)
print(f' 100% [{total}/{total}] cost {cost_time} ms')
mAP = metrics(pred_data)
print("\n========================================\n")
print(f"mAP: {mAP}")
parser.add_argument("--device_id", type=int, default=0, help="Device id")
parser.add_argument("--batch_size", type=int, default=1, help="batch size")
parser.add_argument("--ckpt_file", type=str, required=True, help="Checkpoint file path.")
parser.add_argument("--file_name", type=str, default="ssd", help="output file name.")
parser.add_argument('--file_format', type=str, choices=["AIR", "MINDIR"], default='AIR', help='file format')
parser.add_argument("--device_target", type=str, choices=["Ascend", "GPU", "CPU"], default="Ascend",
help="device target")
args = parser.parse_args()
context.set_context(mode=context.GRAPH_MODE, device_target=args.device_target)
if args.device_target == "Ascend":
context.set_context(device_id=args.device_id)
if __name__ == '__main__':
if config.model == "ssd300":
net = SSD300(ssd_mobilenet_v2(), config, is_training=False)
elif config.model == "ssd_vgg16":
net = ssd_vgg16(config=config)
elif config.model == "ssd_mobilenet_v1_fpn":
net = ssd_mobilenet_v1_fpn(config=config)
elif config.model == "ssd_resnet50_fpn":
net = ssd_resnet50_fpn(config=config)
else:
raise ValueError(f'config.model: {config.model} is not supported')
net = SsdInferWithDecoder(net, Tensor(default_boxes), config)
param_dict = load_checkpoint(args.ckpt_file)
net.init_parameters_data()
load_param_into_net(net, param_dict)
net.set_train(False)
def main():
parser = argparse.ArgumentParser(description="SSD training")
parser.add_argument(
"--only_create_dataset",
type=ast.literal_eval,
default=False,
help="If set it true, only create Mindrecord, default is False.")
parser.add_argument("--distribute",
type=ast.literal_eval,
default=False,
help="Run distribute, default is False.")
parser.add_argument("--device_id",
type=int,
default=0,
help="Device id, default is 0.")
parser.add_argument("--device_num",
type=int,
default=1,
help="Use device nums, default is 1.")
parser.add_argument("--lr",
type=float,
default=0.05,
help="Learning rate, default is 0.05.")
parser.add_argument("--mode",
type=str,
default="sink",
help="Run sink mode or not, default is sink.")
parser.add_argument("--dataset",
type=str,
default="coco",
help="Dataset, defalut is coco.")
parser.add_argument("--epoch_size",
type=int,
default=500,
help="Epoch size, default is 500.")
parser.add_argument("--batch_size",
type=int,
default=32,
help="Batch size, default is 32.")
parser.add_argument("--pre_trained",
type=str,
default=None,
help="Pretrained Checkpoint file path.")
parser.add_argument("--pre_trained_epoch_size",
type=int,
default=0,
help="Pretrained epoch size.")
parser.add_argument("--save_checkpoint_epochs",
type=int,
default=10,
help="Save checkpoint epochs, default is 10.")
parser.add_argument("--loss_scale",
type=int,
default=1024,
help="Loss scale, default is 1024.")
parser.add_argument("--filter_weight",
type=ast.literal_eval,
default=False,
help="Filter weight parameters, default is False.")
args_opt = parser.parse_args()
context.set_context(mode=context.GRAPH_MODE,
device_target="Ascend",
device_id=args_opt.device_id)
if args_opt.distribute:
device_num = args_opt.device_num
context.reset_auto_parallel_context()
context.set_auto_parallel_context(
parallel_mode=ParallelMode.DATA_PARALLEL,
mirror_mean=True,
device_num=device_num)
init()
rank = args_opt.device_id % device_num
else:
rank = 0
device_num = 1
print("Start create dataset!")
# It will generate mindrecord file in args_opt.mindrecord_dir,
# and the file name is ssd.mindrecord0, 1, ... file_num.
prefix = "ssd.mindrecord"
mindrecord_dir = config.mindrecord_dir
mindrecord_file = os.path.join(mindrecord_dir, prefix + "0")
if not os.path.exists(mindrecord_file):
if not os.path.isdir(mindrecord_dir):
os.makedirs(mindrecord_dir)
if args_opt.dataset == "coco":
if os.path.isdir(config.coco_root):
print("Create Mindrecord.")
data_to_mindrecord_byte_image("coco", True, prefix)
print("Create Mindrecord Done, at {}".format(mindrecord_dir))
else:
print("coco_root not exits.")
elif args_opt.dataset == "voc":
if os.path.isdir(config.voc_dir):
print("Create Mindrecord.")
voc_data_to_mindrecord(mindrecord_dir, True, prefix)
print("Create Mindrecord Done, at {}".format(mindrecord_dir))
else:
print("voc_dir not exits.")
else:
if os.path.isdir(config.image_dir) and os.path.exists(
config.anno_path):
print("Create Mindrecord.")
data_to_mindrecord_byte_image("other", True, prefix)
print("Create Mindrecord Done, at {}".format(mindrecord_dir))
else:
print("image_dir or anno_path not exits.")
if not args_opt.only_create_dataset:
loss_scale = float(args_opt.loss_scale)
# When create MindDataset, using the fitst mindrecord file, such as ssd.mindrecord0.
dataset = create_ssd_dataset(mindrecord_file,
repeat_num=1,
batch_size=args_opt.batch_size,
device_num=device_num,
rank=rank)
dataset_size = dataset.get_dataset_size()
print("Create dataset done!")
backbone = ssd_mobilenet_v2()
ssd = SSD300(backbone=backbone, config=config)
net = SSDWithLossCell(ssd, config)
init_net_param(net)
# checkpoint
ckpt_config = CheckpointConfig(save_checkpoint_steps=dataset_size *
args_opt.save_checkpoint_epochs)
ckpoint_cb = ModelCheckpoint(prefix="ssd",
directory=None,
config=ckpt_config)
if args_opt.pre_trained:
if args_opt.pre_trained_epoch_size <= 0:
raise KeyError(
"pre_trained_epoch_size must be greater than 0.")
param_dict = load_checkpoint(args_opt.pre_trained)
if args_opt.filter_weight:
filter_checkpoint_parameter(param_dict)
load_param_into_net(net, param_dict)
lr = Tensor(
get_lr(global_step=config.global_step,
lr_init=config.lr_init,
lr_end=config.lr_end_rate * args_opt.lr,
lr_max=args_opt.lr,
warmup_epochs=config.warmup_epochs,
total_epochs=args_opt.epoch_size,
steps_per_epoch=dataset_size))
opt = nn.Momentum(
filter(lambda x: x.requires_grad, net.get_parameters()), lr,
config.momentum, config.weight_decay, loss_scale)
net = TrainingWrapper(net, opt, loss_scale)
callback = [
TimeMonitor(data_size=dataset_size),
LossMonitor(), ckpoint_cb
]
model = Model(net)
dataset_sink_mode = False
if args_opt.mode == "sink":
print("In sink mode, one epoch return a loss.")
dataset_sink_mode = True
print(
"Start train SSD, the first epoch will be slower because of the graph compilation."
)
model.train(args_opt.epoch_size,
dataset,
callbacks=callback,
dataset_sink_mode=dataset_sink_mode)
def main():
args_opt = get_args()
rank = 0
device_num = 1
if args_opt.run_platform == "CPU":
context.set_context(mode=context.GRAPH_MODE, device_target="CPU")
else:
context.set_context(mode=context.GRAPH_MODE,
device_target=args_opt.run_platform,
device_id=args_opt.device_id)
if args_opt.distribute:
device_num = args_opt.device_num
context.reset_auto_parallel_context()
context.set_auto_parallel_context(
parallel_mode=ParallelMode.DATA_PARALLEL,
gradients_mean=True,
device_num=device_num)
init()
context.set_auto_parallel_context(
all_reduce_fusion_config=[29, 58, 89])
rank = get_rank()
mindrecord_file = create_mindrecord(args_opt.dataset, "ssd.mindrecord",
True)
if args_opt.only_create_dataset:
return
loss_scale = float(args_opt.loss_scale)
if args_opt.run_platform == "CPU":
loss_scale = 1.0
# When create MindDataset, using the fitst mindrecord file, such as ssd.mindrecord0.
use_multiprocessing = (args_opt.run_platform != "CPU")
dataset = create_ssd_dataset(mindrecord_file,
repeat_num=1,
batch_size=args_opt.batch_size,
device_num=device_num,
rank=rank,
use_multiprocessing=use_multiprocessing)
dataset_size = dataset.get_dataset_size()
print("Create dataset done!")
backbone = ssd_mobilenet_v2()
if config.model == "ssd300":
ssd = SSD300(backbone=backbone, config=config)
elif config.model == "ssd_mobilenet_v1_fpn":
ssd = ssd_mobilenet_v1_fpn(config=config)
else:
raise ValueError(f'config.model: {config.model} is not supported')
if args_opt.run_platform == "GPU":
ssd.to_float(dtype.float16)
net = SSDWithLossCell(ssd, config)
init_net_param(net)
if config.feature_extractor_base_param != "":
param_dict = load_checkpoint(config.feature_extractor_base_param)
for x in list(param_dict.keys()):
param_dict["network.feature_extractor.mobilenet_v1." +
x] = param_dict[x]
del param_dict[x]
load_param_into_net(ssd.feature_extractor.mobilenet_v1.network,
param_dict)
# checkpoint
ckpt_config = CheckpointConfig(save_checkpoint_steps=dataset_size *
args_opt.save_checkpoint_epochs)
save_ckpt_path = './ckpt_' + str(rank) + '/'
ckpoint_cb = ModelCheckpoint(prefix="ssd",
directory=save_ckpt_path,
config=ckpt_config)
if args_opt.pre_trained:
param_dict = load_checkpoint(args_opt.pre_trained)
if args_opt.filter_weight:
filter_checkpoint_parameter(param_dict)
load_param_into_net(net, param_dict)
if args_opt.freeze_layer == "backbone":
for param in backbone.feature_1.trainable_params():
param.requires_grad = False
lr = Tensor(
get_lr(global_step=args_opt.pre_trained_epoch_size * dataset_size,
lr_init=config.lr_init,
lr_end=config.lr_end_rate * args_opt.lr,
lr_max=args_opt.lr,
warmup_epochs=config.warmup_epochs,
total_epochs=args_opt.epoch_size,
steps_per_epoch=dataset_size))
if "use_global_norm" in config and config.use_global_norm:
opt = nn.Momentum(
filter(lambda x: x.requires_grad, net.get_parameters()), lr,
config.momentum, config.weight_decay, 1.0)
net = TrainingWrapper(net, opt, loss_scale, True)
else:
opt = nn.Momentum(
filter(lambda x: x.requires_grad, net.get_parameters()), lr,
config.momentum, config.weight_decay, loss_scale)
net = TrainingWrapper(net, opt, loss_scale)
callback = [TimeMonitor(data_size=dataset_size), LossMonitor(), ckpoint_cb]
model = Model(net)
dataset_sink_mode = False
if args_opt.mode == "sink" and args_opt.run_platform != "CPU":
print("In sink mode, one epoch return a loss.")
dataset_sink_mode = True
print(
"Start train SSD, the first epoch will be slower because of the graph compilation."
)
model.train(args_opt.epoch_size,
dataset,
callbacks=callback,
dataset_sink_mode=dataset_sink_mode)
def create_network(name, *args, **kwargs):
if name == "ssd300":
backbone = ssd_mobilenet_v2()
ssd = SSD300(backbone=backbone, config=config, *args, **kwargs)
return ssd
raise NotImplementedError(f"{name} is not implemented in the repo")
