def prune_and_eval(model, sorted_bn, prune_idx,percent,cfg): #不是最终准确的结果
print(f'mAP of the original model is:')
with torch.no_grad():
eval = do_evaluation(cfg, model, distributed=False)
print(eval[0]['metrics'])
model_copy = deepcopy(model)
thre_index = int(len(sorted_bn) * percent)
# 获得α参数的阈值，小于该值的α参数对应的通道，全部裁剪掉
thre = sorted_bn[thre_index]
thre = thre.cuda()
print(f'Channels with Gamma value less than {thre:.4f} are pruned!')
remain_num = 0
for idx in prune_idx:
bn_module = model_copy.backbone.module_list[idx][1]
mask = bn_module.weight.data.abs().ge(thre).float()
remain_num += int(mask.sum())
bn_module.weight.data.mul_(mask)
print(f'Number of channels has been reduced from {len(sorted_bn)} to {remain_num}')
print(f'Prune ratio: {1 - remain_num / len(sorted_bn):.3f}')
print('快速看剪枝效果----》')
print(f'mAP of the pruned model is:')
with torch.no_grad():
eval=do_evaluation(cfg, model_copy, distributed=False)
print(eval[0]['metrics'])
return thre
def evaluation(cfg, ckpt):
logger = logging.getLogger("SSD.inference")
model = SSDDetector(cfg)
checkpointer = CheckPointer(model, save_dir=cfg.OUTPUT_DIR, logger=logger)
model = torch_utils.to_cuda(model)
checkpointer.load(ckpt, use_latest=ckpt is None)
do_evaluation(cfg, model)
def evaluation(cfg, ckpt, distributed):
logger = logging.getLogger("SSD.inference")
model = build_detection_model(cfg)
checkpointer = CheckPointer(model, save_dir=cfg.OUTPUT_DIR, logger=logger)
device = torch.device(cfg.MODEL.DEVICE)
model.to(device)
checkpointer.load(ckpt, use_latest=ckpt is None)
do_evaluation(cfg, model, distributed)
def main():
parser = argparse.ArgumentParser(description='Single Shot MultiBox Detector Training With PyTorch')
parser.add_argument(
"--config-file",
default="",
metavar="FILE",
help="path to config file",
type=str,
)
parser.add_argument("--local_rank", type=int, default=0)
parser.add_argument('--vgg', help='Pre-trained vgg model path, download from https://s3.amazonaws.com/amdegroot-models/vgg16_reducedfc.pth')
parser.add_argument('--resume', default=None, type=str, help='Checkpoint state_dict file to resume training from')
parser.add_argument('--log_step', default=50, type=int, help='Print logs every log_step')
parser.add_argument('--save_step', default=5000, type=int, help='Save checkpoint every save_step')
parser.add_argument('--use_tensorboard', default=True, type=str2bool)
parser.add_argument(
"--skip-test",
dest="skip_test",
help="Do not test the final model",
action="store_true",
)
parser.add_argument(
"opts",
help="Modify config options using the command-line",
default=None,
nargs=argparse.REMAINDER,
)
args = parser.parse_args()
num_gpus = int(os.environ["WORLD_SIZE"]) if "WORLD_SIZE" in os.environ else 1
args.distributed = num_gpus > 1
args.num_gpus = num_gpus
if args.distributed:
torch.cuda.set_device(args.local_rank)
torch.distributed.init_process_group(backend="nccl", init_method="env://")
logger = setup_logger("SSD", distributed_util.get_rank())
logger.info("Using {} GPUs".format(num_gpus))
logger.info(args)
cfg.merge_from_file(args.config_file)
cfg.merge_from_list(args.opts)
cfg.freeze()
logger.info("Loaded configuration file {}".format(args.config_file))
with open(args.config_file, "r") as cf:
config_str = "\n" + cf.read()
logger.info(config_str)
logger.info("Running with config:\n{}".format(cfg))
model = train(cfg, args)
if not args.skip_test:
logger.info('Start evaluating...')
torch.cuda.empty_cache() # speed up evaluating after training finished
do_evaluation(cfg, model, cfg.OUTPUT_DIR, distributed=args.distributed)
def evaluation(cfg, ckpt, distributed):
logger = logging.getLogger("SSD.inference")
model = build_detection_model(cfg)
checkpointer = CheckPointer(model, save_dir=cfg.OUTPUT_DIR, logger=logger)
device = torch.device(cfg.MODEL.DEVICE)
#model.load_state_dict(torch.load('outputs/vgg_ssd300_voc0712.pth'), strict=False)
model.to(device)
checkpointer.load(ckpt, use_latest=ckpt is None)
do_evaluation(cfg, model, distributed)
def evaluation(cfg, weights_file, output_dir, distributed):
if not os.path.exists(output_dir):
os.makedirs(output_dir)
device = torch.device(cfg.MODEL.DEVICE)
model = build_ssd_model(cfg)
model.load(weights_file)
logger = logging.getLogger("SSD.inference")
logger.info('Loaded weights from {}.'.format(weights_file))
model.to(device)
do_evaluation(cfg, model, output_dir, distributed)
def prune_and_eval(model, CBL_idx, CBLidx2mask, cfg):
print(f'mAP of the original model is:')
with torch.no_grad():
eval = do_evaluation(cfg, model, distributed=False)
print(eval[0]['metrics'])
model_copy = deepcopy(model)
for idx in CBL_idx:
bn_module = model_copy.backbone.module_list[idx][1]
mask = CBLidx2mask[idx].cuda()
bn_module.weight.data.mul_(mask)
print('快速看剪枝效果----》')
print(f'mAP of the pruned model is:')
with torch.no_grad():
eval = do_evaluation(cfg, model_copy, distributed=False)
print(eval[0]['metrics'])
def score(self):
torch.cuda.empty_cache()
eval_results = do_evaluation(self.cfg,
self.model,
distributed=self.args.distributed)
mAP = eval_results[0]['metrics']['mAP']
return mAP
def evaluation(cfg, ckpt, distributed, model_path=None):
logger = logging.getLogger("SSD.inference")
model = build_detection_model(cfg)
logger.info("Model :\n{}".format(model)) #如果用print,多gpu会打印两便
checkpointer = CheckPointer(model, save_dir=cfg.OUTPUT_DIR, logger=logger)
device = torch.device(cfg.MODEL.DEVICE)
model.to(device)
if model_path is None:
checkpointer.load(ckpt, use_latest=ckpt is None)
else:
model.load_state_dict(torch.load(model_path))
if cfg.TEST.BN_FUSE is True:
print('BN_FUSE.')
model.backbone.bn_fuse()
model.to(device)
do_evaluation(cfg, model, distributed)
def evaluation(cfg, ckpt, distributed):
logger: logging.RootLogger = logging.getLogger("SSD.inference")
model = build_detection_model(cfg)
checkpointer = CheckPointer(model, save_dir=cfg.OUTPUT_DIR, logger=logger)
device = torch.device(cfg.MODEL.DEVICE)
model.to(device)
checkpointer.load(ckpt, use_latest=ckpt is None)
for scale in np.linspace(0.5, 1.0, 5):
logger.info(f"Running eval with rescale factor: {scale}")
eval_result = do_evaluation(cfg, model, distributed, rescale=scale)
def evaluation(cfg, args, weights_file, output_dir, distributed):
if not os.path.exists(output_dir):
os.makedirs(output_dir)
device = torch.device(cfg.MODEL.DEVICE)
model = build_ssd_model(cfg)
model.load(open(weights_file, 'rb'))
logger = logging.getLogger("SSD.inference")
logger.info('Loaded weights from {}.'.format(weights_file))
model.to(device)
if args.eval_mode == "test":
do_evaluation(cfg, model, output_dir, distributed)
else:
dataset_metrics = do_evaluation(cfg, model, cfg.OUTPUT_DIR, distributed, datasets_dict=_create_val_datasets(args, cfg, logger))
count = len(dataset_metrics)
map_sum = 0
for k,v in dataset_metrics.items():
#logger.info("mAP on {}: {:.3f}".format(k, v.info["mAP"]))
map_sum += v.info["mAP"]
avg_map = map_sum/count
print("'Model': '{}', 'Avg_mAP': {}".format(weights_file, avg_map))
def main():
args = get_parser().parse_args()
cfg.merge_from_file(args.config_file)
cfg.merge_from_list(args.opts)
cfg.freeze()
output_dir = pathlib.Path(cfg.OUTPUT_DIR)
output_dir.mkdir(exist_ok=True, parents=True)
logger = setup_logger("SSD", output_dir)
logger.info(args)
logger.info("Loaded configuration file {}".format(args.config_file))
with open(args.config_file, "r") as cf:
config_str = "\n" + cf.read()
logger.info(config_str)
logger.info("Running with config:\n{}".format(cfg))
model = start_train(cfg)
logger.info('Start evaluating...')
torch.cuda.empty_cache() # speed up evaluating after training finished
do_evaluation(cfg, model)
def main():
parser = argparse.ArgumentParser(
description='Single Shot MultiBox Detector Training With PyTorch')
parser.add_argument(
"config_file",
default="",
metavar="FILE",
help="path to config file",
type=str,
)
parser.add_argument(
"opts",
help="Modify config options using the command-line",
default=None,
nargs=argparse.REMAINDER,
)
args = parser.parse_args()
cfg.merge_from_file(args.config_file)
cfg.merge_from_list(args.opts)
cfg.freeze()
output_dir = pathlib.Path(cfg.OUTPUT_DIR)
output_dir.mkdir(exist_ok=True, parents=True)
logger = setup_logger("SSD", output_dir)
logger.info(args)
logger.info("Loaded configuration file {}".format(args.config_file))
with open(args.config_file, "r") as cf:
config_str = "\n" + cf.read()
logger.info(config_str)
logger.info("Running with config:\n{}".format(cfg))
model = start_train(cfg)
logger.info('Start evaluating...')
torch.cuda.empty_cache() # speed up evaluating after training finished
do_evaluation(cfg, model)
def do_train(cfg, model,
data_loader,
optimizer,
scheduler,
device,
args):
logger = logging.getLogger("SSD.trainer")
logger.info("Start training")
model.train()
save_to_disk = distributed_util.get_rank() == 0
if args.use_tensorboard and save_to_disk:
import tensorboardX
summary_writer = tensorboardX.SummaryWriter(log_dir=cfg.OUTPUT_DIR)
else:
summary_writer = None
max_iter = len(data_loader)
start_training_time = time.time()
trained_time = 0
tic = time.time()
end = time.time()
for iteration, (images, boxes, labels) in enumerate(data_loader):
iteration = iteration + 1
scheduler.step()
images = images.to(device)
boxes = boxes.to(device)
labels = labels.to(device)
#print(images.shape)
#print(labels.shape)
#print(boxes.shape)
optimizer.zero_grad()
loss_dict = model(images, targets=(boxes, labels))
# reduce losses over all GPUs for logging purposes
loss_dict_reduced = reduce_loss_dict(loss_dict)
losses_reduced = sum(loss for loss in loss_dict_reduced.values())
loss = sum(loss for loss in loss_dict.values())
loss.backward()
optimizer.step()
trained_time += time.time() - end
end = time.time()
if iteration % args.log_step == 0:
eta_seconds = int((trained_time / iteration) * (max_iter - iteration))
log_str = [
"Iter: {:06d}, Lr: {:.5f}, Cost: {:.2f}s, Eta: {}".format(iteration,
optimizer.param_groups[0]['lr'],
time.time() - tic, str(datetime.timedelta(seconds=eta_seconds))),
"total_loss: {:.3f}".format(losses_reduced.item())
]
for loss_name, loss_item in loss_dict_reduced.items():
log_str.append("{}: {:.3f}".format(loss_name, loss_item.item()))
log_str = ', '.join(log_str)
logger.info(log_str)
if summary_writer:
global_step = iteration
summary_writer.add_scalar('losses/total_loss', losses_reduced, global_step=global_step)
for loss_name, loss_item in loss_dict_reduced.items():
summary_writer.add_scalar('losses/{}'.format(loss_name), loss_item, global_step=global_step)
summary_writer.add_scalar('lr', optimizer.param_groups[0]['lr'], global_step=global_step)
tic = time.time()
if save_to_disk and iteration % args.save_step == 0:
model_path = os.path.join(cfg.OUTPUT_DIR, "ssd{}_vgg_iteration_{:06d}.pth".format(cfg.INPUT.IMAGE_SIZE, iteration))
_save_model(logger, model, model_path)
# Do eval when training, to trace the mAP changes and see performance improved whether or nor
if args.eval_step > 0 and iteration % args.eval_step == 0 and not iteration == max_iter:
do_evaluation(cfg, model, cfg.OUTPUT_DIR, distributed=args.distributed)
model.train()
if save_to_disk:
model_path = os.path.join(cfg.OUTPUT_DIR, "ssd{}_vgg_final.pth".format(cfg.INPUT.IMAGE_SIZE))
_save_model(logger, model, model_path)
# compute training time
total_training_time = int(time.time() - start_training_time)
total_time_str = str(datetime.timedelta(seconds=total_training_time))
logger.info("Total training time: {} ({:.4f} s / it)".format(total_time_str, total_training_time / max_iter))
return model
def main():
# 解析命令行 读取配置文件
'''
规定了模型的基本参数，训练的类，一共是20类加上背景所以是21
模型的输入大小，为了不对原图造成影响，一般是填充为300*300的图像
训练的文件夹路径2007和2012，测试的文件夹路径2007
最大迭代次数为120000.学习率还有gamma的值，总之就是一系列的超参数
输出的文件目录
MODEL:
NUM_CLASSES: 21
INPUT:
IMAGE_SIZE: 300
DATASETS:
TRAIN: ("voc_2007_trainval", "voc_2012_trainval")
TEST: ("voc_2007_test", )
SOLVER:
MAX_ITER: 120000
LR_STEPS: [80000, 100000]
GAMMA: 0.1
BATCH_SIZE: 32
LR: 1e-3
OUTPUT_DIR: 'outputs/vgg_ssd300_voc0712'
Returns:
'''
parser = argparse.ArgumentParser(description='Single Shot MultiBox Detector Training With PyTorch')
parser.add_argument(
"--config-file",
default="configs/vgg_ssd300_voc0712.yaml",
# default="configs/vgg_ssd300_visdrone0413.yaml",
metavar="FILE",
help="path to config file",
type=str,
)
# 每2500步保存一次文件，并且验证一次文件，记录是每10次记录一次，然后如果不想看tensor的记录的话，可以关闭，使用的是tensorboardX
parser.add_argument("--local_rank", type=int, default=0)
parser.add_argument('--log_step', default=10, type=int, help='Print logs every log_step')
parser.add_argument('--save_step', default=2500, type=int, help='Save checkpoint every save_step')
parser.add_argument('--eval_step', default=2500, type=int, help='Evaluate dataset every eval_step, disabled when eval_step < 0')
parser.add_argument('--use_tensorboard', default=True, type=str2bool)
parser.add_argument(
"--skip-test",
dest="skip_test",
help="Do not test the final model",
action="store_true",
)
parser.add_argument(
"opts",
help="Modify config options using the command-line",
default=None,
nargs=argparse.REMAINDER,
)
# 参数解析，可以使用多GPU进行训练
args = parser.parse_args()
num_gpus = int(os.environ["WORLD_SIZE"]) if "WORLD_SIZE" in os.environ else 1
args.distributed = num_gpus > 1
args.num_gpus = num_gpus
# 做一些启动前必要的检查
if torch.cuda.is_available():
# This flag allows you to enable the inbuilt cudnn auto-tuner to
# find the best algorithm to use for your hardware.
torch.backends.cudnn.benchmark = True
if args.distributed:
torch.cuda.set_device(args.local_rank)
torch.distributed.init_process_group(backend="nccl", init_method="env://")
synchronize()
cfg.merge_from_file(args.config_file)
cfg.merge_from_list(args.opts)
cfg.freeze()
# 创建模型输出文件夹
if cfg.OUTPUT_DIR:
mkdir(cfg.OUTPUT_DIR)
# 使用logger来进行记录
logger = setup_logger("SSD", dist_util.get_rank(), cfg.OUTPUT_DIR)
logger.info("Using {} GPUs".format(num_gpus))
logger.info(args)
# 加载配置文件
logger.info("Loaded configuration file {}".format(args.config_file))
with open(args.config_file, "r") as cf:
config_str = "\n" + cf.read()
logger.info(config_str)
logger.info("Running with config:\n{}".format(cfg))
# 模型训练
# model = train(cfg, args)
model = train(cfg, args)
# 开始进行验证
if not args.skip_test:
logger.info('Start evaluating...')
torch.cuda.empty_cache() # speed up evaluating after training finished
do_evaluation(cfg, model, distributed=args.distributed)
def do_train(cfg, model, data_loader, optimizer, scheduler, checkpointer,
arguments):
logger = logging.getLogger("SSD.trainer")
logger.info("Start training ...")
meters = MetricLogger()
model.train()
summary_writer = torch.utils.tensorboard.SummaryWriter(
log_dir=os.path.join(cfg.OUTPUT_DIR, 'tf_logs'))
max_iter = len(data_loader)
start_iter = arguments["iteration"]
start_training_time = time.time()
end = time.time()
for iteration, (images, targets, _) in enumerate(data_loader, start_iter):
iteration = iteration + 1
arguments["iteration"] = iteration
images = torch_utils.to_cuda(images)
targets = torch_utils.to_cuda(targets)
loss_dict = model(images, targets=targets)
loss = sum(loss for loss in loss_dict.values())
meters.update(total_loss=loss, **loss_dict)
optimizer.zero_grad()
loss.backward()
optimizer.step()
scheduler.step()
batch_time = time.time() - end
end = time.time()
meters.update(time=batch_time)
if iteration % cfg.LOG_STEP == 0:
eta_seconds = meters.time.global_avg * (max_iter - iteration)
eta_string = str(datetime.timedelta(seconds=int(eta_seconds)))
logger.info(
meters.delimiter.join([
"iter: {iter:06d}",
"lr: {lr:.5f}",
'{meters}',
"eta: {eta}",
'mem: {mem}M',
]).format(iter=iteration,
lr=optimizer.param_groups[0]['lr'],
meters=str(meters),
eta=eta_string,
mem=round(torch.cuda.max_memory_allocated() /
1024.0 / 1024.0)))
global_step = iteration
summary_writer.add_scalar('losses/total_loss',
loss,
global_step=global_step)
for loss_name, loss_item in loss_dict.items():
summary_writer.add_scalar('losses/{}'.format(loss_name),
loss_item,
global_step=global_step)
summary_writer.add_scalar('lr',
optimizer.param_groups[0]['lr'],
global_step=global_step)
if iteration % cfg.MODEL_SAVE_STEP == 0:
checkpointer.save("model_{:06d}".format(iteration), **arguments)
if cfg.EVAL_STEP > 0 and iteration % cfg.EVAL_STEP == 0:
eval_results = do_evaluation(cfg, model, iteration=iteration)
for eval_result, dataset in zip(eval_results, cfg.DATASETS.TEST):
write_metric(eval_result['metrics'], 'metrics/' + dataset,
summary_writer, iteration)
model.train() # *IMPORTANT*: change to train mode after eval.
checkpointer.save("model_final", **arguments)
# compute training time
total_training_time = int(time.time() - start_training_time)
total_time_str = str(datetime.timedelta(seconds=total_training_time))
logger.info("Total training time: {} ({:.4f} s / it)".format(
total_time_str, total_training_time / max_iter))
return model
def main():
parser = argparse.ArgumentParser(
description='Single Shot MultiBox Detector Training With PyTorch')
parser.add_argument(
"--config-file",
default="",
metavar="FILE",
help="path to config file",
type=str,
)
parser.add_argument("--local_rank", type=int, default=0)
parser.add_argument(
'--vgg',
help=
'Pre-trained vgg model path, download from https://s3.amazonaws.com/amdegroot-models/vgg16_reducedfc.pth'
)
parser.add_argument(
'--resume',
default=None,
type=str,
help='Checkpoint state_dict file to resume training from')
parser.add_argument('--log_step',
default=50,
type=int,
help='Print logs every log_step')
parser.add_argument('--save_step',
default=5000,
type=int,
help='Save checkpoint every save_step')
parser.add_argument(
'--eval_step',
default=0,
type=int,
help=
'Evaluate dataset every eval_step, disabled when eval_step <= 0. Default: disabled'
)
parser.add_argument('--use_tensorboard', default=True, type=str2bool)
parser.add_argument("--num_workers",
default=4,
type=int,
help="Number of workers to use for data loaders")
parser.add_argument(
"--eval_mode",
default="test",
type=str,
help=
'Use defined test datasets for periodic evaluation or use a validation split. Default: "test", alternative "val"'
)
parser.add_argument(
"--return_best",
default=False,
type=str2bool,
help=
"If false (default) tests on the target the last model. If true tests on the target the model with the best performance on the validation set"
)
parser.add_argument(
"--skip-test",
dest="skip_test",
help="Do not test the final model",
action="store_true",
)
parser.add_argument(
"opts",
help="Modify config options using the command-line",
default=None,
nargs=argparse.REMAINDER,
)
args = parser.parse_args()
num_gpus = int(
os.environ["WORLD_SIZE"]) if "WORLD_SIZE" in os.environ else 1
args.distributed = num_gpus > 1
args.num_gpus = num_gpus
if torch.cuda.is_available():
# This flag allows you to enable the inbuilt cudnn auto-tuner to
# find the best algorithm to use for your hardware.
torch.backends.cudnn.benchmark = True
if args.distributed:
torch.cuda.set_device(args.local_rank)
torch.distributed.init_process_group(backend="nccl",
init_method="env://")
logger = setup_logger("SSD", distributed_util.get_rank())
logger.info("Using {} GPUs".format(num_gpus))
logger.info(args)
cfg.merge_from_file(args.config_file)
cfg.merge_from_list(args.opts)
cfg.freeze()
logger.info("Loaded configuration file {}".format(args.config_file))
with open(args.config_file, "r") as cf:
config_str = "\n" + cf.read()
logger.info(config_str)
logger.info("Running with config:\n{}".format(cfg))
if not os.path.exists(cfg.OUTPUT_DIR):
if not args.distributed or (args.distributed
and distributed_util.is_main_process()):
os.makedirs(cfg.OUTPUT_DIR)
model = train(cfg, args)
if not args.skip_test:
logger.info('Start evaluating...')
torch.cuda.empty_cache() # speed up evaluating after training finished
do_evaluation(cfg, model, cfg.OUTPUT_DIR, distributed=args.distributed)
def do_train(cfg, model, data_loader, optimizer, scheduler, checkpointer,
device, arguments, args):
logger = logging.getLogger("SSD.trainer")
logger.info("Start training ...")
meters = MetricLogger()
model.train()
save_to_disk = dist_util.get_rank() == 0
if args.use_tensorboard and save_to_disk:
import tensorboardX
summary_writer = tensorboardX.SummaryWriter(
log_dir=os.path.join(cfg.OUTPUT_DIR, 'tf_logs'))
else:
summary_writer = None
max_iter = len(data_loader)
start_iter = arguments["iteration"]
start_training_time = time.time()
end = time.time()
max_epoch = 10
for epoch in range(max_epoch):
logger.info('epoch: {}'.format(epoch))
for iteration, (images, targets,
_) in enumerate(data_loader, start_iter):
# print("imgs shape: ",images.shape,iteration)
# continue
# iteration = iteration + 1
arguments["iteration"] = iteration
scheduler.step()
images = images.to(device)
targets = targets.to(device)
loss_dict = model(images, targets=targets)
loss = sum(loss for loss in loss_dict.values())
# reduce losses over all GPUs for logging purposes
loss_dict_reduced = reduce_loss_dict(loss_dict)
losses_reduced = sum(loss for loss in loss_dict_reduced.values())
meters.update(total_loss=losses_reduced, **loss_dict_reduced)
optimizer.zero_grad()
loss.backward()
optimizer.step()
batch_time = time.time() - end
end = time.time()
meters.update(time=batch_time)
# log step
if iteration % args.log_step == 0:
eta_seconds = meters.time.global_avg * (max_iter - iteration)
eta_string = str(datetime.timedelta(seconds=int(eta_seconds)))
logger.info(
meters.delimiter.join([
"iter: {iter:06d}",
"lr: {lr:.5f}",
'{meters}',
"eta: {eta}",
'mem: {mem}M',
]).format(
iter=iteration,
lr=optimizer.param_groups[0]['lr'],
meters=str(meters),
eta=eta_string,
mem=round(torch.cuda.max_memory_allocated() / 1024.0 /
1024.0),
))
if summary_writer:
global_step = iteration
summary_writer.add_scalar('losses/total_loss',
losses_reduced,
global_step=global_step)
for loss_name, loss_item in loss_dict_reduced.items():
summary_writer.add_scalar(
'losses/{}'.format(loss_name),
loss_item,
global_step=global_step)
summary_writer.add_scalar('lr',
optimizer.param_groups[0]['lr'],
global_step=global_step)
# save step
if iteration % args.save_step == 0:
checkpointer.save("model_{:06d}".format(iteration),
**arguments)
# eval step
if args.eval_step > 0 and iteration % args.eval_step == 0 and not iteration == max_iter:
# if True:
eval_results = do_evaluation(cfg,
model,
distributed=args.distributed,
iteration=iteration)
if dist_util.get_rank() == 0 and summary_writer:
for eval_result, dataset in zip(eval_results,
cfg.DATASETS.TEST):
write_metric(eval_result['metrics'],
'metrics/' + dataset, summary_writer,
iteration)
model.train() # *IMPORTANT*: change to train mode after eval.
checkpointer.save("model_final", **arguments)
# compute training time
total_training_time = int(time.time() - start_training_time)
total_time_str = str(datetime.timedelta(seconds=total_training_time))
logger.info("Total training time: {} ({:.4f} s / it)".format(
total_time_str, total_training_time / max_iter))
return model
def shortcut_prune(cfg, model, pruned_cfg, file, max, percent, quick,
weight_path):
obtain_num_parameters = lambda model: sum(
[param.nelement() for param in model.parameters()])
origin_nparameters = obtain_num_parameters(model)
origin_size = model_size(model)
#这里采用的shortcut方法，是https://github.com/SpursLipu/YOLOv3-ModelCompression-MultidatasetTraining-Multibackbone/blob/4516d76ba89b561983babd679543135484e7e9ac/slim_prune.py的方法
CBL_idx, Conv_idx, prune_idx, _, _ = parse_module_defs(
model.backbone.module_defs)
# 将所有要剪枝的BN层的α参数，拷贝到bn_weights列表
bn_weights = gather_bn_weights(model.backbone.module_list, prune_idx)
# torch.sort返回二维列表，第一维是排序后的值列表，第二维是排序后的值列表对应的索引
sorted_bn = torch.sort(bn_weights)[0]
thresh_index = int(len(bn_weights) * percent)
thresh = sorted_bn[thresh_index].cuda()
print(f'Global Threshold should be less than {thresh:.9f}.')
predictor_channels = list(cfg.MODEL.BACKBONE.OUT_CHANNELS)
# 获得保留的卷积核的个数和每层对应的mask,以及对应的head通道数
num_filters, filters_mask, predictor_channels = obtain_filters_mask(
model.backbone, thresh, CBL_idx, prune_idx, predictor_channels, max)
# CBLidx2mask存储CBL_idx中，每一层BN层对应的mask
CBLidx2mask = {idx: mask for idx, mask in zip(CBL_idx, filters_mask)}
CBLidx2filters = {
idx: filters
for idx, filters in zip(CBL_idx, num_filters)
}
for i in model.backbone.module_defs:
if i['type'] == 'shortcut':
i['is_access'] = False
print('merge the mask of layers connected to shortcut!')
merge_mask(model.backbone, CBLidx2mask, CBLidx2filters)
prune_and_eval(model, CBL_idx, CBLidx2mask, cfg)
for i in CBLidx2mask:
CBLidx2mask[i] = CBLidx2mask[i].clone().cpu().numpy()
if quick == 0:
print('实际剪枝---》')
pruned_model = prune_model_keep_size(cfg, model, prune_idx, CBL_idx,
CBLidx2mask)
if max == 0:
with torch.no_grad():
eval = do_evaluation(cfg, pruned_model, distributed=False)
print('after prune_model_keep_size mAP is {}'.format(
eval[0]['metrics'])) #对于最大剪枝，这里是不准的 相当于还没有截掉后面层
# 获得原始模型的module_defs，并修改该defs中的卷积核数量
compact_module_defs = deepcopy(model.backbone.module_defs)
prune_after = -1 # cbl_idx索引号，后面的层都不要了（实际上看，是第一个BN偏置全零层最近的预测层之后的都被剪掉） 针对max
if max == 1:
new_predictor_channels = []
for idx in CBL_idx:
if model.backbone.module_defs[idx][
'feature'] == 'linear' or model.backbone.module_defs[
idx]['feature'] == 'l2_norm':
i = int(model.backbone.module_defs[idx]['feature_idx'])
if predictor_channels[i] != -1:
new_predictor_channels.append(predictor_channels[i])
if i + 1 < len(predictor_channels):
if predictor_channels[i + 1] == -1:
prune_after = idx
break
if i + 1 == len(predictor_channels):
break
elif model.backbone.module_defs[idx + 1][
'type'] == 'shortcut' and model.backbone.module_defs[
idx + 1]['feature'] == 'linear':
i = int(model.backbone.module_defs[idx + 1]['feature_idx'])
new_predictor_channels.append(
predictor_channels[i]) # 第一个short_cut连接head不会被裁掉
predictor_channels = new_predictor_channels
for idx, num in zip(CBL_idx, num_filters):
assert compact_module_defs[idx]['type'] == 'convolutional'
if idx == prune_after + 1 and prune_after != -1:
compact_module_defs[idx]['filters'] = '-1' #这一层连同之后都不要
break
else:
compact_module_defs[idx]['filters'] = str(num)
write_cfg(pruned_cfg, compact_module_defs)
print(f'Config file has been saved: {pruned_cfg}')
cfg.MODEL.BACKBONE.OUT_CHANNELS = tuple(predictor_channels)
print(
f'PRUNED_MODEL.BACKBONE.OUT_CHANNELS:{cfg.MODEL.BACKBONE.OUT_CHANNELS}'
)
cfg.MODEL.BACKBONE.CFG = pruned_cfg
cfg.MODEL.BACKBONE.PRETRAINED = False #定义模型时会加载预训练权重，这里不需要，因为之前的权重不匹配现在的通道数
compact_model = build_detection_model(cfg)
# print(compact_model)
device = torch.device(cfg.MODEL.DEVICE)
compact_model.to(device)
init_weights_from_loose_model(compact_model, pruned_model, CBL_idx,
Conv_idx, CBLidx2mask, prune_after)
compact_nparameters = obtain_num_parameters(compact_model)
compact_size = model_size(compact_model)
random_input = torch.rand(
(16, 3, cfg.INPUT.IMAGE_SIZE, cfg.INPUT.IMAGE_SIZE)).to(device)
pruned_forward_time = obtain_avg_forward_time(random_input,
pruned_model)
compact_forward_time = obtain_avg_forward_time(random_input,
compact_model)
# print(compact_model)
# print(compact_model)
with torch.no_grad():
eval = do_evaluation(cfg, compact_model, distributed=False)
print('Final pruned model mAP is {}'.format(eval[0]['metrics']))
metric_table = [
["Metric", "Before", "After"],
[
"Parameters(M)", f"{origin_nparameters/(1024*1024)}",
f"{compact_nparameters/(1024*1024)}"
],
["模型体积(MB)", f"{origin_size}", f"{compact_size}"],
[
"Inference(ms)", f'{pruned_forward_time*1000/16:.4f}',
f'{compact_forward_time*1000/16:.4f}'
] #bs=16
]
print(AsciiTable(metric_table).table)
print(
f'压缩率：{(origin_nparameters-compact_nparameters)/origin_nparameters}'
)
file.write(
f'PRUNED_MODEL.BACKBONE.OUT_CHANNELS:{cfg.MODEL.BACKBONE.OUT_CHANNELS}'
+ '\n')
file.write(AsciiTable(metric_table).table + '\n')
file.write(
f'压缩率：{(origin_nparameters-compact_nparameters)/origin_nparameters}'
+ '\n')
file.close()
torch.save(compact_model.state_dict(), weight_path)
print(f'Compact model has been saved.')
def do_train(
cfg: CfgNode,
model: SSDDetector,
data_loader: DataLoader,
optimizer: SGD,
scheduler: MultiStepLR,
checkpointer,
device: device,
arguments,
args: Namespace,
output_dir: Path,
model_manager: Dict[str, Any],
) -> SSDDetector:
logger = logging.getLogger("SSD.trainer")
logger.info("Start training ...")
meters = MetricLogger()
model.train()
save_to_disk = dist_util.get_rank() == 0
if args.use_tensorboard and save_to_disk:
import tensorboardX
summary_writer = tensorboardX.SummaryWriter(logdir=output_dir / "logs")
else:
summary_writer = None
max_iter = len(data_loader)
start_iter = arguments["iteration"]
start_training_time = time.time()
end = time.time()
logger.info("MAX_ITER: {}".format(max_iter))
# GB: 2019-09-08:
# For rescaling tests, do an eval before fine-tuning-training, so we know what
# the eval results are before any weights are updated:
# do_evaluation(
# cfg,
# model,
# distributed=args.distributed,
# iteration=0,
# )
# model.train() # *IMPORTANT*: change to train mode after eval.
for iteration, (images, targets, _) in enumerate(data_loader, start_iter):
# TODO: Print learning rate:
iteration = iteration + 1
arguments["iteration"] = iteration
scheduler.step()
images = images.to(device)
targets = targets.to(device)
loss_dict = model(images, targets=targets)
loss = sum(loss for loss in loss_dict.values())
# reduce losses over all GPUs for logging purposes
loss_dict_reduced = reduce_loss_dict(loss_dict)
losses_reduced = sum(loss for loss in loss_dict_reduced.values())
loss = sum(loss for loss in loss_dict.values())
meters.update(total_loss=losses_reduced, **loss_dict_reduced)
optimizer.zero_grad()
loss.backward()
optimizer.step()
batch_time = time.time() - end
end = time.time()
meters.update(time=batch_time)
if iteration % args.log_step == 0:
eta_seconds = meters.time.global_avg * (max_iter - iteration)
eta_string = str(datetime.timedelta(seconds=int(eta_seconds)))
logger.info(
meters.delimiter.join([
"iter: {iter:06d}",
"lr: {lr:.5f}",
"{meters}",
"eta: {eta}",
"mem: {mem}M",
]).format(
iter=iteration,
lr=optimizer.param_groups[0]["lr"],
meters=str(meters),
eta=eta_string,
mem=round(torch.cuda.max_memory_allocated() / 1024.0 /
1024.0),
))
if summary_writer:
global_step = iteration
summary_writer.add_scalar("losses/total_loss",
losses_reduced,
global_step=global_step)
for loss_name, loss_item in loss_dict_reduced.items():
summary_writer.add_scalar(
"losses/{}".format(loss_name),
loss_item,
global_step=global_step,
)
summary_writer.add_scalar("lr",
optimizer.param_groups[0]["lr"],
global_step=global_step)
# This project doesn't use epochs, it does something with batch samplers
# instead, so there is only a concept of "iteration". For now hardcode epoch as
# zero to put into file name:
epoch = 0
save_name = f"ssd{cfg.INPUT.IMAGE_SIZE}-vgg_{cfg.DATASETS.TRAIN[0]}_0_{epoch}_{iteration:06d}"
model_path = Path(output_dir) / f"{save_name}.pth"
# Above if block would be replaced by this:
if iteration % args.save_step == 0:
checkpointer.save(save_name, **arguments)
# Do eval when training, to trace the mAP changes and see performance improved
# whether or nor
if (args.eval_step > 0 and iteration % args.eval_step == 0
and not iteration == max_iter):
eval_results = do_evaluation(
cfg,
model,
distributed=args.distributed,
iteration=iteration,
)
do_best_model_checkpointing(cfg, model_path, eval_results,
model_manager, logger)
if dist_util.get_rank() == 0 and summary_writer:
for eval_result, dataset in zip(eval_results,
cfg.DATASETS.TEST):
write_metric(
eval_result["metrics"],
"metrics/" + dataset,
summary_writer,
iteration,
)
model.train() # *IMPORTANT*: change to train mode after eval.
if iteration % args.save_step == 0:
remove_extra_checkpoints(output_dir, [model_path], logger)
checkpointer.save("model_final", **arguments)
# compute training time
total_training_time = int(time.time() - start_training_time)
total_time_str = str(datetime.timedelta(seconds=total_training_time))
logger.info("Total training time: {} ({:.4f} s / it)".format(
total_time_str, total_training_time / max_iter))
return model
def do_train(cfg, model,
data_loader,
optimizer,
scheduler,
device,
args,
val_sets_dict=None):
logger = logging.getLogger("SSD.trainer")
logger.info("Start training")
model.train()
save_to_disk = distributed_util.get_rank() == 0
if args.use_tensorboard and save_to_disk:
import tensorboardX
summary_writer = tensorboardX.SummaryWriter(log_dir=cfg.OUTPUT_DIR)
tf_writer = tf.compat.v1.summary.FileWriter(cfg.OUTPUT_DIR)
else:
summary_writer = None
if cfg.DATASETS.DG:
dataloaders = data_loader
max_iter = len(data_loader[0])
dataiters = [iter(dataloader) for dataloader in dataloaders]
else:
max_iter = len(data_loader)
data_loader = iter(data_loader)
start_training_time = time.time()
trained_time = 0
tic = time.time()
end = time.time()
if args.return_best:
best_map = 0
for iteration in range(scheduler.last_epoch, max_iter):
if cfg.DATASETS.DG:
# domain generalization settings
# we need to read images from different sources
images = torch.ones(cfg.SOLVER.BATCH_SIZE * len(dataloaders), 3, cfg.INPUT.IMAGE_SIZE, cfg.INPUT.IMAGE_SIZE)
for j in range(len(dataloaders)):
if cfg.MODEL.SELF_SUPERVISED:
d_images, d_boxes, d_labels, d_j_images, d_j_index, d_orig_boxes, d_orig_labels = next(dataiters[j])
else:
d_images, d_boxes, d_labels, d_orig_boxes, d_orig_labels = next(dataiters[j])
start_bs = cfg.SOLVER.BATCH_SIZE * j
end_bs = start_bs + cfg.SOLVER.BATCH_SIZE
images[start_bs:end_bs, :, :, :] = d_images
if j == 0:
boxes = d_boxes
labels = d_labels
orig_boxes = d_orig_boxes
orig_labels = d_orig_labels
if cfg.MODEL.SELF_SUPERVISED:
j_images = d_j_images
j_index = d_j_index
else:
boxes = torch.cat((boxes, d_boxes))
labels = torch.cat((labels, d_labels))
orig_boxes = torch.cat((orig_boxes, d_orig_boxes))
orig_labels = torch.cat((orig_labels, d_orig_labels))
if cfg.MODEL.SELF_SUPERVISED:
j_images = torch.cat((j_images, d_j_images))
j_index = torch.cat((j_index, d_j_index))
else:
if cfg.MODEL.SELF_SUPERVISED:
images, boxes, labels, j_images, j_index, orig_boxes, orig_labels = next(data_loader)
else:
images, boxes, labels, orig_boxes, orig_labels = next(data_loader)
# it is not a problem if we increment iteration because it will be reset in the loop
iteration = iteration + 1
images = images.to(device)
boxes = boxes.to(device)
labels = labels.to(device)
optimizer.zero_grad()
loss_dict = model(images, targets=(boxes, labels))
# reduce losses over all GPUs for logging purposes
loss_dict_reduced = reduce_loss_dict(loss_dict)
losses_reduced = sum(loss for loss in loss_dict_reduced.values())
loss = sum(loss for loss in loss_dict.values())
# loss.backward() becomes:
with amp.scale_loss(loss, optimizer) as scaled_loss:
scaled_loss.backward()
if cfg.MODEL.SELF_SUPERVISED:
j_images = j_images.to(device)
j_index = j_index.to(device)
loss_dict_j = model(j_images, targets=j_index, auxiliary_task=True)
loss_dict_reduced_j = reduce_loss_dict(loss_dict_j)
losses_reduced_j = sum(loss for loss in loss_dict_reduced_j.values())
loss_j = sum(loss for loss in loss_dict_j.values())
# apply reduction factor for auxiliary loss
loss_j = loss_j * cfg.MODEL.SELF_SUPERVISOR.WEIGHT
# loss.backward() becomes:
with amp.scale_loss(loss_j, optimizer) as scaled_loss:
scaled_loss.backward()
# append this loss to the dictionary of losses
loss_dict.update(loss_dict_j)
losses_reduced += losses_reduced_j
optimizer.step()
scheduler.step()
trained_time += time.time() - end
end = time.time()
if iteration % args.log_step == 0:
eta_seconds = int((trained_time / iteration) * (max_iter - iteration))
log_str = [
"Iter: {:06d}, Lr: {:.5f}, Cost: {:.2f}s, Eta: {}".format(iteration,
optimizer.param_groups[0]['lr'],
time.time() - tic,
str(datetime.timedelta(seconds=eta_seconds))),
"total_loss: {:.3f}".format(losses_reduced.item())
]
for loss_name, loss_item in loss_dict_reduced.items():
log_str.append("{}: {:.3f}".format(loss_name, loss_item.item()))
log_str = ', '.join(log_str)
logger.info(log_str)
if summary_writer:
global_step = iteration
summary_writer.add_scalar('losses/total_loss', losses_reduced, global_step=global_step)
for loss_name, loss_item in loss_dict_reduced.items():
summary_writer.add_scalar('losses/{}'.format(loss_name), loss_item, global_step=global_step)
summary_writer.add_scalar('lr', optimizer.param_groups[0]['lr'], global_step=global_step)
if cfg.MODEL.SELF_SUPERVISED:
_log_images_tensorboard(cfg, global_step, images, orig_boxes, orig_labels, summary_writer, j_images=j_images)
else:
_log_images_tensorboard(cfg, global_step, images, orig_boxes, orig_labels, summary_writer)
#for tag, value in model.named_parameters():
# tag = tag.replace('.', '/')
# if 'ss_classifier' in tag:
# print(tag, value)
#_log_network_params(tf_writer, model, global_step)
tic = time.time()
if save_to_disk and iteration % args.save_step == 0:
model_path = os.path.join(cfg.OUTPUT_DIR,
"ssd{}_vgg_iteration_{:06d}.pth".format(cfg.INPUT.IMAGE_SIZE, iteration))
save_training_checkpoint(logger, model, scheduler, optimizer, model_path)
# Do eval when training, to trace the mAP changes and see whether or not performance improved
# if args.return_best = True the model returned should be the one that gave best performances on the val set
if args.eval_step > 0 and iteration % args.eval_step == 0 and (not iteration == max_iter or args.return_best):
dataset_metrics = do_evaluation(cfg, model, cfg.OUTPUT_DIR, distributed=args.distributed, datasets_dict=val_sets_dict)
model.train()
if args.distributed and not distributed_util.is_main_process():
continue
avg_map = _compute_avg_map(dataset_metrics)
if args.return_best:
if avg_map > best_map:
best_map = avg_map
logger.info("With iteration {} passed the best! New best avg map: {:4f}".format(iteration, best_map))
model_path = os.path.join(cfg.OUTPUT_DIR, "ssd{}_vgg_best.pth".format(cfg.INPUT.IMAGE_SIZE))
_save_model(logger, model, model_path)
else:
logger.info("With iteration {} the best has not been reached. Best avg map: {:4f}, Current avg mAP: {:4f}".format(iteration, best_map, avg_map))
# logging
if summary_writer:
global_step = iteration
summary_writer.add_scalar("val_avg_map", avg_map, global_step=global_step)
for dataset_name, metrics in dataset_metrics.items():
for metric_name, metric_value in metrics.get_printable_metrics().items():
summary_writer.add_scalar('/'.join(['val', dataset_name, metric_name]), metric_value,
global_step=global_step)
if save_to_disk:
model_path = os.path.join(cfg.OUTPUT_DIR, "ssd{}_vgg_final.pth".format(cfg.INPUT.IMAGE_SIZE))
_save_model(logger, model, model_path)
# compute training time
total_training_time = int(time.time() - start_training_time)
total_time_str = str(datetime.timedelta(seconds=total_training_time))
logger.info("Total training time: {} ({:.4f} s / it)".format(total_time_str, total_training_time / max_iter))
if args.return_best:
model.load(os.path.join(cfg.OUTPUT_DIR, "ssd{}_vgg_best.pth".format(cfg.INPUT.IMAGE_SIZE)))
return model
def main():
parser = ArgumentParser(
description="Single Shot MultiBox Detector Training With PyTorch")
parser.add_argument(
"--config-file",
default="",
metavar="FILE",
help="config file name or path (relative to the configs/ folder) ",
type=str,
)
parser.add_argument("--local_rank", type=int, default=0)
parser.add_argument("--log_step",
default=50,
type=int,
help="Print logs every log_step")
parser.add_argument("--save_step",
default=5000,
type=int,
help="Save checkpoint every save_step")
parser.add_argument(
"--eval_step",
default=5000,
type=int,
help="Evaluate dataset every eval_step, disabled when eval_step < 0",
)
parser.add_argument("--use_tensorboard", default=True, type=str2bool)
parser.add_argument(
"--skip-test",
dest="skip_test",
help="Do not test the final model",
action="store_true",
)
parser.add_argument(
"opts",
help="Modify config options using the command-line",
default=None,
nargs=REMAINDER,
)
parser.add_argument(
"--resume_experiment",
default="None",
dest="resume",
type=str,
help="Checkpoint state_dict file to resume training from",
)
args = parser.parse_args()
num_gpus = int(
os.environ["WORLD_SIZE"]) if "WORLD_SIZE" in os.environ else 1
args.distributed = num_gpus > 1
args.num_gpus = num_gpus
if torch.cuda.is_available():
# This flag allows you to enable the inbuilt cudnn auto-tuner to
# find the best algorithm to use for your hardware.
torch.backends.cudnn.benchmark = True
else:
cfg.MODEL.DEVICE = "cpu"
if args.distributed:
torch.cuda.set_device(args.local_rank)
torch.distributed.init_process_group(backend="nccl",
init_method="env://")
synchronize()
eman = ExperimentManager("ssd")
output_dir = eman.get_output_dir()
args.config_file = str(
Path(__file__).parent / "configs" / args.config_file)
cfg.merge_from_file(args.config_file)
cfg.merge_from_list(args.opts)
cfg.OUTPUT_DIR = str(output_dir)
cfg.freeze()
eman.start({"cfg": cfg, "args": vars(args)})
# We use our own output dir, set by ExperimentManager:
# if cfg.OUTPUT_DIR:
# mkdir(cfg.OUTPUT_DIR)
logger = setup_logger("SSD", dist_util.get_rank(), output_dir / "logs")
logger.info("Using {} GPUs".format(num_gpus))
logger.info(args)
logger.info("Loaded configuration file {}".format(args.config_file))
with open(args.config_file, "r") as cf:
config_str = "\n" + cf.read()
logger.info(config_str)
logger.info("Running with config:\n{}".format(cfg))
logger.info(f"Output dir: {output_dir}")
model_manager = {"best": None, "new": None}
model = train(cfg, args, output_dir, model_manager)
if not args.skip_test:
logger.info("Start evaluating...")
torch.cuda.empty_cache() # speed up evaluating after training finished
eval_results = do_evaluation(
cfg,
model,
distributed=args.distributed,
)
do_best_model_checkpointing(
cfg,
output_dir / "model_final.pth",
eval_results,
model_manager,
logger,
is_final=True,
)
eman.mark_dir_if_complete()
def main():
parser = argparse.ArgumentParser(
description='Single Shot MultiBox Detector Training With PyTorch')
parser.add_argument(
"--config-file",
default="",
metavar="FILE",
help="path to config file",
type=str,
)
parser.add_argument("--local_rank", type=int, default=0)
parser.add_argument('--log_step',
default=10,
type=int,
help='Print logs every log_step')
parser.add_argument('--save_step',
default=2500,
type=int,
help='Save checkpoint every save_step')
parser.add_argument(
'--eval_step',
default=2500,
type=int,
help='Evaluate dataset every eval_step, disabled when eval_step < 0')
parser.add_argument('--use_tensorboard', default=True, type=str2bool)
parser.add_argument(
"--skip-test",
dest="skip_test",
help="Do not test the final model",
action="store_true",
)
parser.add_argument(
"opts",
help="Modify config options using the command-line",
default=None,
nargs=argparse.REMAINDER,
)
args = parser.parse_args()
num_gpus = int(
os.environ["WORLD_SIZE"]) if "WORLD_SIZE" in os.environ else 1
args.distributed = num_gpus > 1
args.num_gpus = num_gpus
if torch.cuda.is_available():
# This flag allows you to enable the inbuilt cudnn auto-tuner to
# find the best algorithm to use for your hardware.
torch.backends.cudnn.benchmark = True
if args.distributed:
torch.cuda.set_device(args.local_rank)
torch.distributed.init_process_group(backend="nccl",
init_method="env://")
synchronize()
# Train distance regression network
train_distance_regr()
cfg.merge_from_file(args.config_file)
cfg.merge_from_list(args.opts)
cfg.freeze()
if cfg.OUTPUT_DIR:
mkdir(cfg.OUTPUT_DIR)
logger = setup_logger("SSD", dist_util.get_rank(), cfg.OUTPUT_DIR)
logger.info("Using {} GPUs".format(num_gpus))
logger.info(args)
logger.info("Loaded configuration file {}".format(args.config_file))
with open(args.config_file, "r") as cf:
config_str = "\n" + cf.read()
logger.info(config_str)
logger.info("Running with config:\n{}".format(cfg))
model = train(cfg, args)
if not args.skip_test:
logger.info('Start evaluating...')
torch.cuda.empty_cache() # speed up evaluating after training finished
do_evaluation(cfg, model, distributed=args.distributed)
def do_train(cfg, model, data_loader, optimizer, scheduler, checkpointer,
device, arguments, args):
logger = logging.getLogger("SSD.trainer")
logger.info("Start training ...")
meters = MetricLogger()
# #获得要剪枝的层
if cfg.PRUNE.TYPE != 'no':
if hasattr(model, 'module'):
backbone = model.module.backbone
else:
backbone = model.backbone
if cfg.PRUNE.TYPE == 'normal':
logger.info('normal sparse training')
_, _, prune_idx = normal_prune.parse_module_defs(
backbone.module_defs)
elif cfg.PRUNE.TYPE == 'shortcut':
logger.info('shortcut sparse training')
_, _, prune_idx, _, _ = shortcut_prune.parse_module_defs(
backbone.module_defs)
model.train()
save_to_disk = dist_util.get_rank() == 0
if args.use_tensorboard and save_to_disk:
try:
from torch.utils.tensorboard import SummaryWriter
except ImportError:
from tensorboardX import SummaryWriter
summary_writer = SummaryWriter(
log_dir=os.path.join(cfg.OUTPUT_DIR, 'tf_logs'))
else:
summary_writer = None
max_iter = len(data_loader)
start_iter = arguments["iteration"]
start_training_time = time.time()
end = time.time()
for iteration, (images, targets, _) in enumerate(data_loader, start_iter):
iteration = iteration + 1
arguments["iteration"] = iteration
images = images.to(device)
targets = targets.to(device)
loss_dict = model(images, targets=targets)
loss = sum(loss for loss in loss_dict.values())
# reduce losses over all GPUs for logging purposes
loss_dict_reduced = reduce_loss_dict(loss_dict)
losses_reduced = sum(loss for loss in loss_dict_reduced.values())
meters.update(total_loss=losses_reduced, **loss_dict_reduced)
optimizer.zero_grad()
loss.backward()
# 对要剪枝层的γ参数稀疏化
if cfg.PRUNE.TYPE != 'no':
if hasattr(model, 'module'):
bn_sparse.updateBN(model.module.backbone.module_list,
cfg.PRUNE.SR, prune_idx)
else:
# print(model.backbone.module_list)
bn_sparse.updateBN(model.backbone.module_list, cfg.PRUNE.SR,
prune_idx)
optimizer.step()
scheduler.step()
batch_time = time.time() - end
end = time.time()
meters.update(time=batch_time)
if iteration % args.log_step == 0:
eta_seconds = meters.time.global_avg * (max_iter - iteration)
eta_string = str(datetime.timedelta(seconds=int(eta_seconds)))
logger.info(
meters.delimiter.join([
"iter: {iter:06d}",
"lr: {lr:.5f}",
'{meters}',
"eta: {eta}",
'mem: {mem}M',
]).format(
iter=iteration,
lr=optimizer.param_groups[0]['lr'],
meters=str(meters),
eta=eta_string,
mem=round(torch.cuda.max_memory_allocated() / 1024.0 /
1024.0),
))
if summary_writer:
global_step = iteration
summary_writer.add_scalar('losses/total_loss',
losses_reduced,
global_step=global_step)
for loss_name, loss_item in loss_dict_reduced.items():
summary_writer.add_scalar('losses/{}'.format(loss_name),
loss_item,
global_step=global_step)
summary_writer.add_scalar('lr',
optimizer.param_groups[0]['lr'],
global_step=global_step)
if iteration % args.save_step == 0:
checkpointer.save("model_{:06d}".format(iteration), **arguments)
if args.eval_step > 0 and iteration % args.eval_step == 0 and not iteration == max_iter:
eval_results = do_evaluation(cfg,
model,
distributed=False,
iteration=iteration) #单gpu测试
if dist_util.get_rank() == 0 and summary_writer:
for eval_result, dataset in zip(eval_results,
cfg.DATASETS.TEST):
write_metric(eval_result['metrics'], 'metrics/' + dataset,
summary_writer, iteration)
model.train() # *IMPORTANT*: change to train mode after eval.
checkpointer.save("model_final", **arguments)
# compute training time
total_training_time = int(time.time() - start_training_time)
total_time_str = str(datetime.timedelta(seconds=total_training_time))
logger.info("Total training time: {} ({:.4f} s / it)".format(
total_time_str, total_training_time / max_iter))
return model
def do_train(cfg, model, data_loader, optimizer, checkpointer, arguments,
scheduler):
logger = logging.getLogger("SSD.trainer")
logger.info("Start training ...")
meters = MetricLogger()
model.train()
summary_writer = torch.utils.tensorboard.SummaryWriter(
log_dir=os.path.join(cfg.OUTPUT_DIR, 'tf_logs'))
max_iter = len(data_loader)
start_iter = arguments["iteration"]
start_training_time = time.time()
end = time.time()
scaler = torch.cuda.amp.GradScaler()
print(model)
for iteration, (images, targets, _) in enumerate(data_loader, start_iter):
iteration = iteration + 1
arguments["iteration"] = iteration
images = torch_utils.to_cuda(images)
targets = torch_utils.to_cuda(targets)
# Casts operations to mixed precision
with torch.cuda.amp.autocast():
loss_dict = model(images.half(), targets=targets)
loss = sum(loss for loss in loss_dict.values())
meters.update(total_loss=loss, **loss_dict)
optimizer.zero_grad()
# Scales the loss, and calls backward()
# to create scaled gradients
scaler.scale(loss).backward()
# loss.backward()
# Unscales gradients and calls
# or skips optimizer.step()
scaler.step(optimizer)
# optimizer.step(iteration)
# Updates the scale for next iteration
scaler.update()
if iteration > 5000:
scheduler.step()
batch_time = time.time() - end
end = time.time()
meters.update(time=batch_time)
if iteration % cfg.LOG_STEP == 0:
eta_seconds = meters.time.global_avg * (max_iter - iteration)
eta_string = str(datetime.timedelta(seconds=int(eta_seconds)))
lr = optimizer.param_groups[0]['lr']
to_log = [
f"iter: {iteration:06d}",
f"lr: {lr:.5f}",
f'{meters}',
f"eta: {eta_string}",
]
if torch.cuda.is_available():
mem = round(torch.cuda.max_memory_allocated() / 1024.0 /
1024.0)
to_log.append(f'mem: {mem}M')
logger.info(meters.delimiter.join(to_log))
global_step = iteration
summary_writer.add_scalar('losses/total_loss',
loss,
global_step=global_step)
for loss_name, loss_item in loss_dict.items():
summary_writer.add_scalar('losses/{}'.format(loss_name),
loss_item,
global_step=global_step)
summary_writer.add_scalar('lr',
optimizer.param_groups[0]['lr'],
global_step=global_step)
if iteration % cfg.MODEL_SAVE_STEP == 0:
checkpointer.save("model_{:06d}".format(iteration), **arguments)
if cfg.EVAL_STEP > 0 and iteration % cfg.EVAL_STEP == 0:
eval_results = do_evaluation(cfg, model, iteration=iteration)
for eval_result, dataset in zip(eval_results, cfg.DATASETS.TEST):
write_metric(eval_result['metrics'], 'metrics/' + dataset,
summary_writer, iteration)
model.train() # *IMPORTANT*: change to train mode after eval.
if iteration >= cfg.SOLVER.MAX_ITER:
break
checkpointer.save("model_final", **arguments)
# compute training time
total_training_time = int(time.time() - start_training_time)
total_time_str = str(datetime.timedelta(seconds=total_training_time))
logger.info("Total training time: {} ({:.4f} s / it)".format(
total_time_str, total_training_time / max_iter))
return model
def do_train(cfg, model, data_loader, optimizer, scheduler, checkpointer,
device, arguments, args):
logger = logging.getLogger("SSD.trainer")
logger.info("Start training ...")
meters = MetricLogger()
model.train()
save_to_disk = dist_util.get_rank() == 0
if args.use_tensorboard and save_to_disk:
import tensorboardX
summary_writer = tensorboardX.SummaryWriter(
log_dir=os.path.join(cfg.OUTPUT_DIR, 'tf_logs'))
else:
summary_writer = None
max_iter = len(data_loader)
start_iter = arguments["iteration"]
start_training_time = time.time()
end = time.time()
for iteration, (images, targets, _, boxes_norm,
labels_norm) in enumerate(data_loader, start_iter):
iteration = iteration + 1
arguments["iteration"] = iteration
scheduler.step()
images = images.to(device)
targets = targets.to(device)
#+++++++++++++++++++++++++++++++++++++++++++++++ Mask GT ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
mask_t = np.zeros((images.shape[0], 81, 64, 64))
mask_t[:, 0, :, :] = np.ones((1, 1, 64, 64))
for i in range(images.shape[0]):
for L, B_norm in zip(labels_norm[i], boxes_norm[i]):
xmin = int(B_norm[0] * 64)
ymin = int(B_norm[1] * 64)
xmax = int(B_norm[2] * 64)
ymax = int(B_norm[3] * 64)
lab = int(L)
mask_t[i, 0, ymin:ymax, xmin:xmax] = 0.0
mask_t[i, lab, ymin:ymax, xmin:xmax] = 1.0
mask_t = Variable(torch.from_numpy((mask_t).astype(np.float32))).cuda()
#+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
loss_dict = model(images, targets=(targets, mask_t))
loss = sum(loss for loss in loss_dict.values())
# reduce losses over all GPUs for logging purposes
loss_dict_reduced = reduce_loss_dict(loss_dict)
losses_reduced = sum(loss for loss in loss_dict_reduced.values())
meters.update(total_loss=losses_reduced, **loss_dict_reduced)
optimizer.zero_grad()
loss.backward()
optimizer.step()
batch_time = time.time() - end
end = time.time()
meters.update(time=batch_time)
if iteration % args.log_step == 0:
eta_seconds = meters.time.global_avg * (max_iter - iteration)
eta_string = str(datetime.timedelta(seconds=int(eta_seconds)))
logger.info(
meters.delimiter.join([
"iter: {iter:06d}",
"lr: {lr:.5f}",
'{meters}',
"eta: {eta}",
'mem: {mem}M',
]).format(
iter=iteration,
lr=optimizer.param_groups[0]['lr'],
meters=str(meters),
eta=eta_string,
mem=round(torch.cuda.max_memory_allocated() / 1024.0 /
1024.0),
))
if summary_writer:
global_step = iteration
summary_writer.add_scalar('losses/total_loss',
losses_reduced,
global_step=global_step)
for loss_name, loss_item in loss_dict_reduced.items():
summary_writer.add_scalar('losses/{}'.format(loss_name),
loss_item,
global_step=global_step)
summary_writer.add_scalar('lr',
optimizer.param_groups[0]['lr'],
global_step=global_step)
if iteration % args.save_step == 0:
checkpointer.save("model_{:06d}".format(iteration), **arguments)
if args.eval_step > 0 and iteration % args.eval_step == 0 and not iteration == max_iter:
eval_results = do_evaluation(cfg,
model,
distributed=args.distributed,
iteration=iteration)
if dist_util.get_rank() == 0 and summary_writer:
for eval_result, dataset in zip(eval_results,
cfg.DATASETS.TEST):
write_metric(eval_result['metrics'], 'metrics/' + dataset,
summary_writer, iteration)
model.train() # *IMPORTANT*: change to train mode after eval.
checkpointer.save("model_final", **arguments)
# compute training time
total_training_time = int(time.time() - start_training_time)
total_time_str = str(datetime.timedelta(seconds=total_training_time))
logger.info("Total training time: {} ({:.4f} s / it)".format(
total_time_str, total_training_time / max_iter))
return model
def do_train(cfg, model, data_loader, optimizer, scheduler, checkpointer,
device, arguments, args):
logger = logging.getLogger("SSD.trainer")
logger.info("Start training ...")
meters = MetricLogger()
# 模型设置为train()模式，表示参数是可以进行更新的
model.train()
save_to_disk = dist_util.get_rank() == 0
# 这个是关于模型训练过程中的过程记录
if args.use_tensorboard and save_to_disk:
import tensorboardX
summary_writer = tensorboardX.SummaryWriter(
log_dir=os.path.join(cfg.OUTPUT_DIR, 'tf_logs'))
else:
summary_writer = None
# dataloader的大小，根据配置文件中的iteration进行训练
# arguments = {"iteration": 0}，按照目前的理解是按照断点进行训练，这个表示的是当前的迭代次数这样
max_iter = len(data_loader)
start_iter = arguments["iteration"]
# 开始计时
start_training_time = time.time()
end = time.time()
# 一次训练中，数据长度应该是dataloader的大小，也就是按照batchsize进行分割之后的大小
# 数据集会返回图像和图像对应的标签，也就是(类别数目) （c+4）k，k个先验框、c个类别，然后加一个框的坐标位置
for iteration, (images, targets, _) in enumerate(data_loader, start_iter):
# print(iteration)
# print(targets)
iteration = iteration + 1
arguments["iteration"] = iteration
images = images.to(device)
targets = targets.to(device)
# 把输入和目标输出传入模型，模型就会返回loss
loss_dict = model(images, targets=targets)
loss = sum(loss for loss in loss_dict.values())
# reduce losses over all GPUs for logging purposes
# 这里是多GPU的操作，暂时先不用去理会
loss_dict_reduced = reduce_loss_dict(loss_dict)
losses_reduced = sum(loss for loss in loss_dict_reduced.values())
meters.update(total_loss=losses_reduced, **loss_dict_reduced)
# 这里是标准的反向传播的过程，传播就完事了
optimizer.zero_grad()
loss.backward()
optimizer.step()
scheduler.step()
# 记录时间、写日志、写模型然后保存训练中的过程记录之类的，这里也基本是死的，主要找到模型就完事了
batch_time = time.time() - end
end = time.time()
meters.update(time=batch_time)
if iteration % args.log_step == 0:
eta_seconds = meters.time.global_avg * (max_iter - iteration)
eta_string = str(datetime.timedelta(seconds=int(eta_seconds)))
logger.info(
meters.delimiter.join([
"iter: {iter:06d}",
"lr: {lr:.5f}",
'{meters}',
"eta: {eta}",
'mem: {mem}M',
]).format(
iter=iteration,
lr=optimizer.param_groups[0]['lr'],
meters=str(meters),
eta=eta_string,
mem=round(torch.cuda.max_memory_allocated() / 1024.0 /
1024.0),
))
if summary_writer:
global_step = iteration
summary_writer.add_scalar('losses/total_loss',
losses_reduced,
global_step=global_step)
for loss_name, loss_item in loss_dict_reduced.items():
summary_writer.add_scalar('losses/{}'.format(loss_name),
loss_item,
global_step=global_step)
summary_writer.add_scalar('lr',
optimizer.param_groups[0]['lr'],
global_step=global_step)
if iteration % args.save_step == 0:
checkpointer.save("model_{:06d}".format(iteration), **arguments)
# 目前问题主要存在这个部分，就是利用模型进行验证的过程中会报错，验证的文件有错误
if args.eval_step > 0 and iteration % args.eval_step == 0 and not iteration == max_iter:
eval_results = do_evaluation(cfg,
model,
distributed=args.distributed,
iteration=iteration)
if dist_util.get_rank() == 0 and summary_writer:
for eval_result, dataset in zip(eval_results,
cfg.DATASETS.TEST):
write_metric(eval_result['metrics'], 'metrics/' + dataset,
summary_writer, iteration)
model.train() # *IMPORTANT*: change to train mode after eval.
checkpointer.save("model_final", **arguments)
# compute training time
total_training_time = int(time.time() - start_training_time)
total_time_str = str(datetime.timedelta(seconds=total_training_time))
logger.info("Total training time: {} ({:.4f} s / it)".format(
total_time_str, total_training_time / max_iter))
return model
def do_train_with_style(cfg, model, data_loader, style_loader, optimizer,
scheduler, checkpointer, device, arguments, args):
logger = logging.getLogger("SSD.trainer")
logger.info("Start training ...")
meters = MetricLogger()
model.train()
save_to_disk = dist_util.get_rank() == 0
if args.use_tensorboard and save_to_disk:
try:
from torch.utils.tensorboard import SummaryWriter
except ImportError:
from tensorboardX import SummaryWriter
summary_writer = SummaryWriter(
log_dir=os.path.join(cfg.OUTPUT_DIR, 'tf_logs'))
else:
summary_writer = None
max_iter = len(data_loader)
start_iter = arguments["iteration"]
start_training_time = time.time()
end = time.time()
# prepare AdaIN models
default_path = '/content/drive/MyDrive/DA_detection/models/'
vgg_path = default_path + 'vgg_normalized.pth'
if 'VGG_PATH' in os.environ:
vgg_path = os.environ['VGG_PATH']
decoder_path = default_path + 'decoder.pth'
if 'DECODER_PATH' in os.environ:
decoder_path = os.environ['DECODER_PATH']
# DEBUG: print('AdaIN > models loaded')
for iteration, (images, targets, ids) in enumerate(data_loader,
start_iter):
iteration = iteration + 1
arguments["iteration"] = iteration
# AdaIN routine
random.seed()
styles = next(iter(style_loader))
# DEBUG: print('AdaIN > begin new batch')
if random.random() > args.p:
apply_style_transfer(vgg_path, decoder_path, images, styles[0],
args.p)
# DEBUG: print('AdaIN > end batch')
images = images.to(device)
targets = targets.to(device)
loss_dict = model(images, targets=targets)
loss = sum(loss for loss in loss_dict.values())
# reduce losses over all GPUs for logging purposes
loss_dict_reduced = reduce_loss_dict(loss_dict)
losses_reduced = sum(loss for loss in loss_dict_reduced.values())
meters.update(total_loss=losses_reduced, **loss_dict_reduced)
optimizer.zero_grad()
loss.backward()
optimizer.step()
scheduler.step()
batch_time = time.time() - end
end = time.time()
meters.update(time=batch_time)
if iteration % args.log_step == 0:
eta_seconds = meters.time.global_avg * (max_iter - iteration)
eta_string = str(datetime.timedelta(seconds=int(eta_seconds)))
if device == "cuda":
logger.info(
meters.delimiter.join([
"iter: {iter:06d}",
"lr: {lr:.5f}",
'{meters}',
"eta: {eta}",
'mem: {mem}M',
]).format(
iter=iteration,
lr=optimizer.param_groups[0]['lr'],
meters=str(meters),
eta=eta_string,
mem=round(torch.cuda.max_memory_allocated() / 1024.0 /
1024.0),
))
else:
logger.info(
meters.delimiter.join([
"iter: {iter:06d}",
"lr: {lr:.5f}",
'{meters}',
"eta: {eta}",
]).format(
iter=iteration,
lr=optimizer.param_groups[0]['lr'],
meters=str(meters),
eta=eta_string,
))
if summary_writer:
global_step = iteration
summary_writer.add_scalar('losses/total_loss',
losses_reduced,
global_step=global_step)
for loss_name, loss_item in loss_dict_reduced.items():
summary_writer.add_scalar('losses/{}'.format(loss_name),
loss_item,
global_step=global_step)
summary_writer.add_scalar('lr',
optimizer.param_groups[0]['lr'],
global_step=global_step)
if iteration % args.save_step == 0:
checkpointer.save("model_{:06d}".format(iteration), **arguments)
if args.eval_step > 0 and iteration % args.eval_step == 0 and not iteration == max_iter:
eval_results = do_evaluation(cfg,
model,
distributed=args.distributed,
iteration=iteration)
if dist_util.get_rank() == 0 and summary_writer:
for eval_result, dataset in zip(eval_results,
cfg.DATASETS.TEST):
write_metric(eval_result['metrics'], 'metrics/' + dataset,
summary_writer, iteration)
model.train() # *IMPORTANT*: change to train mode after eval.
checkpointer.save("model_final", **arguments)
# compute training time
total_training_time = int(time.time() - start_training_time)
total_time_str = str(datetime.timedelta(seconds=total_training_time))
logger.info("Total training time: {} ({:.4f} s / it)".format(
total_time_str, total_training_time / max_iter))
return model
def normal_prune(cfg,model,pruned_cfg,file,regular,max,percent,quick,weight_path):
obtain_num_parameters = lambda model: sum([param.nelement() for param in model.parameters()])
origin_nparameters = obtain_num_parameters(model)
origin_size=model_size(model)
CBL_idx, Conv_idx, prune_idx=parse_module_defs(model.backbone.module_defs)
# 将所有要剪枝的BN层的α参数，拷贝到bn_weights列表
bn_weights = gather_bn_weights(model.backbone.module_list, prune_idx)
# torch.sort返回二维列表，第一维是排序后的值列表，第二维是排序后的值列表对应的索引
sorted_bn = torch.sort(bn_weights)[0]
# 避免剪掉所有channel的最高阈值(每个BN层的gamma的最大值的最小值即为阈值上限)
# highest_thre = []
# for idx in prune_idx:
# # .item()可以得到张量里的元素值
# highest_thre.append(model.backbone.module_list[idx][1].weight.data.abs().max().item())
# highest_thre = min(highest_thre)
# 找到highest_thre对应的下标对应的百分比
# percent_limit = (sorted_bn==highest_thre).nonzero().item()/len(bn_weights)
# print(f'Threshold should be less than {highest_thre:.4f}.')
# print(f'The corresponding prune ratio should less than {percent_limit:.3f}.') #这一行的限制只是为了防止normal剪枝某一层减空，如果保留这一层的一个，则没有这个限制了
thre=prune_and_eval(model,sorted_bn,prune_idx,percent,cfg)
if quick ==0:
print('实际剪枝---》')
predictor_channels=list(cfg.MODEL.BACKBONE.OUT_CHANNELS)
# 获得保留的卷积核的个数和每层对应的mask,以及对应的head通道数
num_filters, filters_mask,predictor_channels=obtain_filters_mask(model.backbone, thre, CBL_idx, prune_idx,predictor_channels,regular,max)
# CBLidx2mask存储CBL_idx中，每一层BN层对应的mask
CBLidx2mask = {idx: mask for idx, mask in zip(CBL_idx, filters_mask)}
pruned_model = prune_model_keep_size(cfg,model, prune_idx, CBL_idx, CBLidx2mask,Conv_idx)
if max==0:
with torch.no_grad():
eval = do_evaluation(cfg,pruned_model, distributed=False)
print('after prune_model_keep_size mAP is {}'.format(eval[0]['metrics'])) #对于最大剪枝，这里是不准的 相当于还没有截掉后面层
# 获得原始模型的module_defs，并修改该defs中的卷积核数量
compact_module_defs = deepcopy(model.backbone.module_defs)
prune_after = -1 # cbl_idx索引号，后面的层都不要了（实际上看，是第一个BN偏置全零层最近的预测层之后的都被剪掉） 针对max
if max==1:
new_predictor_channels=[]
for idx in CBL_idx:
if model.backbone.module_defs[idx]['feature'] == 'linear' or model.backbone.module_defs[idx]['feature'] =='l2_norm':
i = int(model.backbone.module_defs[idx]['feature_idx'])
if predictor_channels[i] != -1:
new_predictor_channels.append(predictor_channels[i])
if i + 1 < len(predictor_channels):
if predictor_channels[i + 1] == -1:
prune_after = idx
break
if i+1==len(predictor_channels):
break
elif model.backbone.module_defs[idx+1]['type']=='shortcut' and model.backbone.module_defs[idx+1]['feature']=='linear':
i = int(model.backbone.module_defs[idx+1]['feature_idx'])
new_predictor_channels.append(predictor_channels[i])#第一个short_cut连接head不会被裁掉
predictor_channels=new_predictor_channels
for idx, num in zip(CBL_idx, num_filters):
assert compact_module_defs[idx]['type'] == 'convolutional'
if idx==prune_after+1 and prune_after!=-1:
compact_module_defs[idx]['filters']='-1'#这一层连同之后都不要
break
else:
compact_module_defs[idx]['filters'] = str(num)
write_cfg(pruned_cfg, compact_module_defs)
print(f'Config file has been saved: {pruned_cfg}')
cfg.MODEL.BACKBONE.OUT_CHANNELS=tuple(predictor_channels)
print(f'PRUNED_MODEL.BACKBONE.OUT_CHANNELS:{cfg.MODEL.BACKBONE.OUT_CHANNELS}')
cfg.MODEL.BACKBONE.CFG=pruned_cfg
cfg.MODEL.BACKBONE.PRETRAINED=False #定义模型时会加载预训练权重，这里不需要，因为之前的权重不匹配现在的通道数
compact_model = build_detection_model(cfg)
# print(compact_model)
device = torch.device(cfg.MODEL.DEVICE)
compact_model.to(device)
init_weights_from_loose_model(compact_model, pruned_model, CBL_idx, Conv_idx, CBLidx2mask,prune_after)
compact_nparameters = obtain_num_parameters(compact_model)
compact_size = model_size(compact_model)
random_input = torch.rand((16, 3, cfg.INPUT.IMAGE_SIZE, cfg.INPUT.IMAGE_SIZE)).to(device)
pruned_forward_time = obtain_avg_forward_time(random_input, pruned_model)
compact_forward_time = obtain_avg_forward_time(random_input, compact_model)
# print(compact_model)
# print(compact_model)
with torch.no_grad():
eval = do_evaluation(cfg, compact_model, distributed=False)
print('Final pruned model mAP is {}'.format(eval[0]['metrics']))
metric_table = [
["Metric", "Before", "After"],
["Parameters(M)", f"{origin_nparameters/(1024*1024)}", f"{compact_nparameters/(1024*1024)}"],
["模型体积(MB)", f"{origin_size}", f"{compact_size}"],
["Inference(ms)", f'{pruned_forward_time*1000/16:.4f}', f'{compact_forward_time*1000/16:.4f}'] #bs=16
]
print(AsciiTable(metric_table).table)
print(f'压缩率：{(origin_nparameters-compact_nparameters)/origin_nparameters}')
file.write(f'PRUNED_MODEL.BACKBONE.OUT_CHANNELS:{cfg.MODEL.BACKBONE.OUT_CHANNELS}' + '\n')
file.write(AsciiTable(metric_table).table + '\n')
file.write(f'压缩率：{(origin_nparameters-compact_nparameters)/origin_nparameters}' + '\n')
file.close()
torch.save(compact_model.state_dict(),weight_path)
print(f'Compact model has been saved.')