def main():
state_dump = os.path.join(MODEL_OUT, "state.pth")
model_dump = os.path.join(MODEL_OUT, "model.pth")
if os.path.exists(model_dump):
model = torch.load(model_dump)
elif os.path.exists(state_dump):
model = Net()
model.load_state_dict(torch.load(state_dump))
else:
model = Net()
optimizer = torch.optim.SGD(model.parameters(), lr=LEARNING_RATE)
loss_fn = nn.MSELoss()
data_loader = dataloader.get_train_loader()
train_model(model, data_loader, optimizer, loss_fn)
print("Saving model state dict into {}".format(state_dump))
torch.save(model.state_dict(), state_dump)
print("Saving model into {}".format(model_dump))
torch.save(model, model_dump)
model.eval()
image, expected = dataloader.get_random_image()
predicted = test_inference(model, image)
print("Predicted {}, expected {}".format(predicted, expected))
plt.imshow(image[0].reshape(28, 28), cmap="gray")
plt.show()
def main(args):
train_dl, style_img = get_train_loader(4)
model = StyleTransfer(style_img, ['relu2_2'],
['relu1_2', 'relu2_2', 'relu3_3', 'relu4_3'],
alpha=1e3)
x, y = next(iter(train_dl))
print(model.train_batch(x))
def run():
seed_everything(args.seed)
df = pd.read_csv(os.path.join(args.data_dir, "train.csv"))
df = process_data(df, args.subset)
df_folds = create_folds(df, args.n_folds)
train_image_ids = df_folds[df_folds["fold"] != args.fold].index.values
valid_image_ids = df_folds[df_folds["fold"] == args.fold].index.values
train_loader = get_train_loader(
args.data_dir,
df,
train_image_ids,
transforms=get_train_augs(args),
do_cutmix=args.cutmix,
batch_size=args.bs,
num_workers=args.num_workers,
)
valid_loader = get_valid_loader(
args.data_dir,
df,
valid_image_ids,
transforms=get_valid_augs(args),
batch_size=args.bs,
num_workers=args.num_workers,
)
model = get_model(args.model_variant,
model_dir=args.model_dir,
checkpoint_path=args.load_path).cuda()
if args.scheduler == "one_cycle":
args.steps_per_epoch = len(train_image_ids) // args.bs
scheduler_class, scheduler_params = get_scheduler(args)
else:
scheduler_class, scheduler_params = get_scheduler(args)
learner = Learner(model, scheduler_class, scheduler_params, hparams=args)
learner.fit(train_loader, valid_loader)
torch.manual_seed(config.seed)
if torch.cuda.is_available():
torch.cuda.manual_seed(config.seed)
parser = argparse.ArgumentParser()
with Engine(custom_parser=parser) as engine:
args = parser.parse_args()
cudnn.benchmark = True
if engine.distributed:
torch.cuda.set_device(engine.local_rank)
# data loader
train_loader, train_sampler = get_train_loader(engine, ADE)
# config network and criterion
criterion = nn.CrossEntropyLoss(reduction='mean', ignore_index=-1)
if engine.distributed:
logger.info('Use the Multi-Process-SyncBatchNorm')
BatchNorm2d = SyncBatchNorm
# else:
# BatchNorm2d = BatchNorm2d
model = CPNet(config.num_classes,
criterion=criterion,
pretrained_model=config.pretrained_model,
norm_layer=BatchNorm2d)
init_weight(model.business_layer,
nn.init.kaiming_normal_,
from models import CVAE
from utils import trainEpochs, plot_results
from dataloader import get_train_loader
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
# ----------Hyper Parameters---------- #
hidden_size = 256
cond_embedding_size = 8
latent_size = 32
n_epochs = 900
batch_size = 16
learning_rate = 0.001
train_loader = get_train_loader(batch_size=batch_size)
for KL_annealing_method in ['cyclical', 'monotonic']:
cvae = CVAE(hidden_size, latent_size, cond_embedding_size).to(device)
history = trainEpochs(train_loader=train_loader,
cvae=cvae,
n_epochs=n_epochs,
learning_rate=learning_rate,
KL_annealing_method=KL_annealing_method)
plot_results(history, KL_annealing_method)
filename = f'./histories/{KL_annealing_method}.pkl'
with open(filename, "wb") as fp:
pickle.dump(history, fp)
torch.manual_seed(config.seed)
if torch.cuda.is_available():
torch.cuda.manual_seed(config.seed)
parser = argparse.ArgumentParser()
with Engine(custom_parser=parser) as engine:
args = parser.parse_args()
cudnn.benchmark = True
if engine.distributed:
torch.cuda.set_device(engine.local_rank)
# data loader
train_loader, train_sampler = get_train_loader(engine, DAVIS)
# config network and criterion
criterion = nn.BCEWithLogitsLoss(reduction='none')
#criterion = BinaryDiceLoss()
#criterion = sigmoid_focal_loss_star_jit
if engine.distributed:
logger.info('Use the Multi-Process-SyncBatchNorm')
cfg = get_cfg()
cfg.merge_from_file('./fpn_config/semantic_R_50_FPN_1x.yaml')
model = AnchorDiffNet(cfg,
embedding=cfg.MODEL.SEM_SEG_HEAD.CONVS_DIM,
batch_mode='sync')
parser = argparse.ArgumentParser()
with Engine(custom_parser=parser) as engine:
args = parser.parse_args()
cudnn.benchmark = True
seed = config.seed
if engine.distributed:
seed = engine.local_rank
torch.manual_seed(seed)
if torch.cuda.is_available():
torch.cuda.manual_seed(seed)
# data loader
train_loader, train_sampler = get_train_loader(engine, Cityscapes)
# config network and criterion
criterion = nn.CrossEntropyLoss(reduction='mean', ignore_index=255)
ohem_criterion = ProbOhemCrossEntropy2d(
ignore_label=255,
thresh=0.7,
min_kept=int(config.batch_size // len(engine.devices) *
config.image_height * config.image_width //
(16 * config.gt_down_sampling**2)),
use_weight=False)
if engine.distributed:
BatchNorm2d = SyncBatchNorm
model = BiSeNet(config.num_classes,
torch.manual_seed(config.seed)
if torch.cuda.is_available():
torch.cuda.manual_seed(config.seed)
parser = argparse.ArgumentParser()
with Engine(custom_parser=parser) as engine:
args = parser.parse_args()
cudnn.benchmark = True
if engine.distributed:
torch.cuda.set_device(engine.local_rank)
# data loader
train_loader, train_sampler = get_train_loader(engine, PascalContext)
# config network and criterion
criterion = nn.CrossEntropyLoss(reduction='mean', ignore_index=-1)
if engine.distributed:
logger.info('Use the Multi-Process-SyncBatchNorm')
BatchNorm2d = SyncBatchNorm
model = CPNet(config.num_classes,
criterion=criterion,
pretrained_model=config.pretrained_model,
norm_layer=BatchNorm2d)
init_weight(model.business_layer,
nn.init.kaiming_normal_,
BatchNorm2d,
def main():
create_exp_dir(config.save,
scripts_to_save=glob.glob('*.py') + glob.glob('*.sh'))
logger = SummaryWriter(config.save)
log_format = '%(asctime)s %(message)s'
logging.basicConfig(stream=sys.stdout,
level=logging.INFO,
format=log_format,
datefmt='%m/%d %I:%M:%S %p')
fh = logging.FileHandler(os.path.join(config.save, 'log.txt'))
fh.setFormatter(logging.Formatter(log_format))
logging.getLogger().addHandler(fh)
logging.info("args = %s", str(config))
# preparation ################
torch.backends.cudnn.enabled = True
torch.backends.cudnn.benchmark = True
seed = config.seed
np.random.seed(seed)
torch.manual_seed(seed)
if torch.cuda.is_available():
torch.cuda.manual_seed(seed)
# config network and criterion ################
min_kept = int(config.batch_size * config.image_height *
config.image_width // (16 * config.gt_down_sampling**2))
ohem_criterion = ProbOhemCrossEntropy2d(ignore_label=255,
thresh=0.7,
min_kept=min_kept,
use_weight=False)
distill_criterion = nn.KLDivLoss()
# data loader ###########################
if config.is_test:
data_setting = {
'img_root': config.img_root_folder,
'gt_root': config.gt_root_folder,
'train_source': config.train_eval_source,
'eval_source': config.eval_source,
'test_source': config.test_source,
'down_sampling': config.down_sampling
}
else:
data_setting = {
'img_root': config.img_root_folder,
'gt_root': config.gt_root_folder,
'train_source': config.train_source,
'eval_source': config.eval_source,
'test_source': config.test_source,
'down_sampling': config.down_sampling
}
train_loader = get_train_loader(config, Cityscapes, test=config.is_test)
# Model #######################################
models = []
evaluators = []
testers = []
lasts = []
for idx, arch_idx in enumerate(config.arch_idx):
if config.load_epoch == "last":
state = torch.load(
os.path.join(config.load_path, "arch_%d.pt" % arch_idx))
else:
state = torch.load(
os.path.join(
config.load_path,
"arch_%d_%d.pt" % (arch_idx, int(config.load_epoch))))
model = Network([
state["alpha_%d_0" % arch_idx].detach(),
state["alpha_%d_1" % arch_idx].detach(),
state["alpha_%d_2" % arch_idx].detach()
], [
None, state["beta_%d_1" % arch_idx].detach(),
state["beta_%d_2" % arch_idx].detach()
], [
state["ratio_%d_0" % arch_idx].detach(),
state["ratio_%d_1" % arch_idx].detach(),
state["ratio_%d_2" % arch_idx].detach()
],
num_classes=config.num_classes,
layers=config.layers,
Fch=config.Fch,
width_mult_list=config.width_mult_list,
stem_head_width=config.stem_head_width[idx],
ignore_skip=arch_idx == 0)
mIoU02 = state["mIoU02"]
latency02 = state["latency02"]
obj02 = objective_acc_lat(mIoU02, latency02)
mIoU12 = state["mIoU12"]
latency12 = state["latency12"]
obj12 = objective_acc_lat(mIoU12, latency12)
if obj02 > obj12: last = [2, 0]
else: last = [2, 1]
lasts.append(last)
model.build_structure(last)
logging.info("net: " + str(model))
for b in last:
if len(config.width_mult_list) > 1:
plot_op(getattr(model, "ops%d" % b),
getattr(model, "path%d" % b),
width=getattr(model, "widths%d" % b),
head_width=config.stem_head_width[idx][1],
F_base=config.Fch).savefig(os.path.join(
config.save, "ops_%d_%d.png" % (arch_idx, b)),
bbox_inches="tight")
else:
plot_op(getattr(model, "ops%d" % b),
getattr(model, "path%d" % b),
F_base=config.Fch).savefig(os.path.join(
config.save, "ops_%d_%d.png" % (arch_idx, b)),
bbox_inches="tight")
plot_path_width(model.lasts, model.paths, model.widths).savefig(
os.path.join(config.save, "path_width%d.png" % arch_idx))
plot_path_width([2, 1, 0], [model.path2, model.path1, model.path0],
[model.widths2, model.widths1, model.widths0]).savefig(
os.path.join(config.save,
"path_width_all%d.png" % arch_idx))
flops, params = profile(model,
inputs=(torch.randn(1, 3, 1024, 2048), ))
logging.info("params = %fMB, FLOPs = %fGB", params / 1e6, flops / 1e9)
logging.info("ops:" + str(model.ops))
logging.info("path:" + str(model.paths))
logging.info("last:" + str(model.lasts))
model = model.cuda()
init_weight(model,
nn.init.kaiming_normal_,
torch.nn.BatchNorm2d,
config.bn_eps,
config.bn_momentum,
mode='fan_in',
nonlinearity='relu')
if arch_idx == 0 and len(config.arch_idx) > 1:
partial = torch.load(
os.path.join(config.teacher_path, "weights%d.pt" % arch_idx))
state = model.state_dict()
pretrained_dict = {k: v for k, v in partial.items() if k in state}
state.update(pretrained_dict)
model.load_state_dict(state)
elif config.is_eval:
partial = torch.load(
os.path.join(config.eval_path, "weights%d.pt" % arch_idx))
state = model.state_dict()
pretrained_dict = {k: v for k, v in partial.items() if k in state}
state.update(pretrained_dict)
model.load_state_dict(state)
evaluator = SegEvaluator(Cityscapes(data_setting, 'val', None),
config.num_classes,
config.image_mean,
config.image_std,
model,
config.eval_scale_array,
config.eval_flip,
0,
out_idx=0,
config=config,
verbose=False,
save_path=None,
show_image=False)
evaluators.append(evaluator)
tester = SegTester(Cityscapes(data_setting, 'test', None),
config.num_classes,
config.image_mean,
config.image_std,
model,
config.eval_scale_array,
config.eval_flip,
0,
out_idx=0,
config=config,
verbose=False,
save_path=None,
show_image=False)
testers.append(tester)
# Optimizer ###################################
base_lr = config.lr
if arch_idx == 1 or len(config.arch_idx) == 1:
# optimize teacher solo OR student (w. distill from teacher)
optimizer = torch.optim.SGD(model.parameters(),
lr=base_lr,
momentum=config.momentum,
weight_decay=config.weight_decay)
models.append(model)
# Cityscapes ###########################################
if config.is_eval:
logging.info(config.load_path)
logging.info(config.eval_path)
logging.info(config.save)
# validation
print("[validation...]")
with torch.no_grad():
valid_mIoUs = infer(models, evaluators, logger)
for idx, arch_idx in enumerate(config.arch_idx):
if arch_idx == 0:
logger.add_scalar("mIoU/val_teacher", valid_mIoUs[idx], 0)
logging.info("teacher's valid_mIoU %.3f" %
(valid_mIoUs[idx]))
else:
logger.add_scalar("mIoU/val_student", valid_mIoUs[idx], 0)
logging.info("student's valid_mIoU %.3f" %
(valid_mIoUs[idx]))
exit(0)
tbar = tqdm(range(config.nepochs), ncols=80)
for epoch in tbar:
logging.info(config.load_path)
logging.info(config.save)
logging.info("lr: " + str(optimizer.param_groups[0]['lr']))
# training
tbar.set_description("[Epoch %d/%d][train...]" %
(epoch + 1, config.nepochs))
train_mIoUs = train(train_loader, models, ohem_criterion,
distill_criterion, optimizer, logger, epoch)
torch.cuda.empty_cache()
for idx, arch_idx in enumerate(config.arch_idx):
if arch_idx == 0:
logger.add_scalar("mIoU/train_teacher", train_mIoUs[idx],
epoch)
logging.info("teacher's train_mIoU %.3f" % (train_mIoUs[idx]))
else:
logger.add_scalar("mIoU/train_student", train_mIoUs[idx],
epoch)
logging.info("student's train_mIoU %.3f" % (train_mIoUs[idx]))
adjust_learning_rate(base_lr, 0.992, optimizer, epoch + 1,
config.nepochs)
# validation
if not config.is_test and ((epoch + 1) % 10 == 0 or epoch == 0):
tbar.set_description("[Epoch %d/%d][validation...]" %
(epoch + 1, config.nepochs))
with torch.no_grad():
valid_mIoUs = infer(models, evaluators, logger)
for idx, arch_idx in enumerate(config.arch_idx):
if arch_idx == 0:
logger.add_scalar("mIoU/val_teacher", valid_mIoUs[idx],
epoch)
logging.info("teacher's valid_mIoU %.3f" %
(valid_mIoUs[idx]))
else:
logger.add_scalar("mIoU/val_student", valid_mIoUs[idx],
epoch)
logging.info("student's valid_mIoU %.3f" %
(valid_mIoUs[idx]))
save(models[idx],
os.path.join(config.save, "weights%d.pt" % arch_idx))
# test
if config.is_test and (epoch + 1) >= 250 and (epoch + 1) % 10 == 0:
tbar.set_description("[Epoch %d/%d][test...]" %
(epoch + 1, config.nepochs))
with torch.no_grad():
test(epoch, models, testers, logger)
for idx, arch_idx in enumerate(config.arch_idx):
save(models[idx],
os.path.join(config.save, "weights%d.pt" % arch_idx))
input = torch.mul(input, mask)
target = torch.mul(target, mask)
loss = F.mse_loss(input, target, reduction=self.reduction)
return loss
parser = argparse.ArgumentParser()
with Engine(custom_parser=parser) as engine:
args = parser.parse_args()
seed = config.seed
torch.manual_seed(seed)
train_loader, train_sampler = get_train_loader(engine, NYUDataset)
criterion = Mseloss()
BatchNorm2d = nn.BatchNorm2d
model = Displacement_Field()
init_weight(model.displacement_net, nn.init.xavier_normal_, BatchNorm2d,
config.bn_eps, config.bn_momentum)
base_lr = config.lr
total_iteration = config.nepochs * config.niters_per_epoch
lr_policy = PolyLR(base_lr, config.lr_power, total_iteration)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model.to(device)
def main(pretrain=True):
config.save = 'search-{}-{}'.format(config.save,
time.strftime("%Y%m%d-%H%M%S"))
create_exp_dir(config.save,
scripts_to_save=glob.glob('*.py') + glob.glob('*.sh'))
logger = SummaryWriter(config.save)
log_format = '%(asctime)s %(message)s'
logging.basicConfig(stream=sys.stdout,
level=logging.INFO,
format=log_format,
datefmt='%m/%d %I:%M:%S %p')
fh = logging.FileHandler(os.path.join(config.save, 'log.txt'))
fh.setFormatter(logging.Formatter(log_format))
logging.getLogger().addHandler(fh)
assert type(pretrain) == bool or type(pretrain) == str
update_arch = True
if pretrain == True:
update_arch = False
logging.info("args = %s", str(config))
# preparation ################
torch.backends.cudnn.enabled = True
torch.backends.cudnn.benchmark = True
seed = config.seed
np.random.seed(seed)
torch.manual_seed(seed)
if torch.cuda.is_available():
torch.cuda.manual_seed(seed)
# config network and criterion ################
min_kept = int(config.batch_size * config.image_height *
config.image_width // (16 * config.gt_down_sampling**2))
ohem_criterion = ProbOhemCrossEntropy2d(ignore_label=255,
thresh=0.7,
min_kept=min_kept,
use_weight=False)
# Model #######################################
model = Network(config.num_classes,
config.layers,
ohem_criterion,
Fch=config.Fch,
width_mult_list=config.width_mult_list,
prun_modes=config.prun_modes,
stem_head_width=config.stem_head_width)
flops, params = profile(model,
inputs=(torch.randn(1, 3, 1024, 2048), ),
verbose=False)
logging.info("params = %fMB, FLOPs = %fGB", params / 1e6, flops / 1e9)
model = model.cuda()
if type(pretrain) == str:
partial = torch.load(pretrain + "/weights.pt", map_location='cuda:0')
state = model.state_dict()
pretrained_dict = {
k: v
for k, v in partial.items()
if k in state and state[k].size() == partial[k].size()
}
state.update(pretrained_dict)
model.load_state_dict(state)
else:
init_weight(model,
nn.init.kaiming_normal_,
nn.BatchNorm2d,
config.bn_eps,
config.bn_momentum,
mode='fan_in',
nonlinearity='relu')
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model.to(device)
architect = Architect(model, config)
# Optimizer ###################################
base_lr = config.lr
parameters = []
parameters += list(model.stem.parameters())
parameters += list(model.cells.parameters())
parameters += list(model.refine32.parameters())
parameters += list(model.refine16.parameters())
parameters += list(model.head0.parameters())
parameters += list(model.head1.parameters())
parameters += list(model.head2.parameters())
parameters += list(model.head02.parameters())
parameters += list(model.head12.parameters())
optimizer = torch.optim.SGD(parameters,
lr=base_lr,
momentum=config.momentum,
weight_decay=config.weight_decay)
# lr policy ##############################
lr_policy = torch.optim.lr_scheduler.ExponentialLR(optimizer, 0.978)
# data loader ###########################
data_setting = {
'img_root': config.img_root_folder,
'gt_root': config.gt_root_folder,
'train_source': config.train_source,
'eval_source': config.eval_source,
'down_sampling': config.down_sampling
}
train_loader_model = get_train_loader(config,
EGTEA,
portion=config.train_portion)
train_loader_arch = get_train_loader(config,
EGTEA,
portion=config.train_portion - 1)
evaluator = SegEvaluator(EGTEA(data_setting, 'val', None),
config.num_classes,
config.image_mean,
config.image_std,
model,
config.eval_scale_array,
config.eval_flip,
0,
config=config,
verbose=False,
save_path=None,
show_image=False)
if update_arch:
for idx in range(len(config.latency_weight)):
logger.add_scalar("arch/latency_weight%d" % idx,
config.latency_weight[idx], 0)
logging.info("arch_latency_weight%d = " % idx +
str(config.latency_weight[idx]))
tbar = tqdm(range(config.nepochs), ncols=80)
valid_mIoU_history = []
FPSs_history = []
latency_supernet_history = []
latency_weight_history = []
valid_names = ["8s", "16s", "32s", "8s_32s", "16s_32s"]
arch_names = {0: "teacher", 1: "student"}
for epoch in tbar:
logging.info(pretrain)
logging.info(config.save)
logging.info("lr: " + str(optimizer.param_groups[0]['lr']))
logging.info("update arch: " + str(update_arch))
# training
tbar.set_description("[Epoch %d/%d][train...]" %
(epoch + 1, config.nepochs))
train(pretrain,
train_loader_model,
train_loader_arch,
model,
architect,
ohem_criterion,
optimizer,
lr_policy,
logger,
epoch,
update_arch=update_arch)
torch.cuda.empty_cache()
lr_policy.step()
# validation
tbar.set_description("[Epoch %d/%d][validation...]" %
(epoch + 1, config.nepochs))
with torch.no_grad():
if pretrain == True:
model.prun_mode = "min"
valid_mIoUs = infer(epoch, model, evaluator, logger, FPS=False)
for i in range(5):
logger.add_scalar('mIoU/val_min_%s' % valid_names[i],
valid_mIoUs[i], epoch)
logging.info("Epoch %d: valid_mIoU_min_%s %.3f" %
(epoch, valid_names[i], valid_mIoUs[i]))
if len(model._width_mult_list) > 1:
model.prun_mode = "max"
valid_mIoUs = infer(epoch,
model,
evaluator,
logger,
FPS=False)
for i in range(5):
logger.add_scalar('mIoU/val_max_%s' % valid_names[i],
valid_mIoUs[i], epoch)
logging.info("Epoch %d: valid_mIoU_max_%s %.3f" %
(epoch, valid_names[i], valid_mIoUs[i]))
model.prun_mode = "random"
valid_mIoUs = infer(epoch,
model,
evaluator,
logger,
FPS=False)
for i in range(5):
logger.add_scalar(
'mIoU/val_random_%s' % valid_names[i],
valid_mIoUs[i], epoch)
logging.info("Epoch %d: valid_mIoU_random_%s %.3f" %
(epoch, valid_names[i], valid_mIoUs[i]))
else:
valid_mIoUss = []
FPSs = []
model.prun_mode = None
for idx in range(len(model._arch_names)):
# arch_idx
model.arch_idx = idx
valid_mIoUs, fps0, fps1 = infer(epoch, model, evaluator,
logger)
valid_mIoUss.append(valid_mIoUs)
FPSs.append([fps0, fps1])
for i in range(5):
# preds
logger.add_scalar(
'mIoU/val_%s_%s' %
(arch_names[idx], valid_names[i]), valid_mIoUs[i],
epoch)
logging.info("Epoch %d: valid_mIoU_%s_%s %.3f" %
(epoch, arch_names[idx], valid_names[i],
valid_mIoUs[i]))
if config.latency_weight[idx] > 0:
logger.add_scalar(
'Objective/val_%s_8s_32s' % arch_names[idx],
objective_acc_lat(valid_mIoUs[3], 1000. / fps0),
epoch)
logging.info(
"Epoch %d: Objective_%s_8s_32s %.3f" %
(epoch, arch_names[idx],
objective_acc_lat(valid_mIoUs[3], 1000. / fps0)))
logger.add_scalar(
'Objective/val_%s_16s_32s' % arch_names[idx],
objective_acc_lat(valid_mIoUs[4], 1000. / fps1),
epoch)
logging.info(
"Epoch %d: Objective_%s_16s_32s %.3f" %
(epoch, arch_names[idx],
objective_acc_lat(valid_mIoUs[4], 1000. / fps1)))
valid_mIoU_history.append(valid_mIoUss)
FPSs_history.append(FPSs)
if update_arch:
latency_supernet_history.append(architect.latency_supernet)
latency_weight_history.append(architect.latency_weight)
save(model, os.path.join(config.save, 'weights.pt'))
if type(pretrain) == str:
# contains arch_param names: {"alphas": alphas, "betas": betas, "gammas": gammas, "ratios": ratios}
for idx, arch_name in enumerate(model._arch_names):
state = {}
for name in arch_name['alphas']:
state[name] = getattr(model, name)
for name in arch_name['betas']:
state[name] = getattr(model, name)
for name in arch_name['ratios']:
state[name] = getattr(model, name)
state["mIoU02"] = valid_mIoUs[3]
state["mIoU12"] = valid_mIoUs[4]
if pretrain is not True:
state["latency02"] = 1000. / fps0
state["latency12"] = 1000. / fps1
torch.save(
state,
os.path.join(config.save, "arch_%d_%d.pt" % (idx, epoch)))
torch.save(state,
os.path.join(config.save, "arch_%d.pt" % (idx)))
if update_arch:
for idx in range(len(config.latency_weight)):
if config.latency_weight[idx] > 0:
if (int(FPSs[idx][0] >= config.FPS_max[idx]) +
int(FPSs[idx][1] >= config.FPS_max[idx])) >= 1:
architect.latency_weight[idx] /= 2
elif (int(FPSs[idx][0] <= config.FPS_min[idx]) +
int(FPSs[idx][1] <= config.FPS_min[idx])) > 0:
architect.latency_weight[idx] *= 2
logger.add_scalar(
"arch/latency_weight_%s" % arch_names[idx],
architect.latency_weight[idx], epoch + 1)
logging.info("arch_latency_weight_%s = " %
arch_names[idx] +
str(architect.latency_weight[idx]))
port = str(int(float(time.time())) % 20)
os.environ['MASTER_PORT'] = str(10097 + int(port))
with Engine(custom_parser=parser) as engine:
args = parser.parse_args()
cudnn.benchmark = True
seed = config.seed
if engine.distributed:
seed = engine.local_rank
torch.manual_seed(seed)
if torch.cuda.is_available():
torch.cuda.manual_seed(seed)
# data loader
train_loader, train_sampler = get_train_loader(engine, NYUv2)
if engine.distributed and (engine.local_rank == 0):
tb_dir = config.tb_dir + '/{}'.format(time.strftime("%b%d_%d-%H-%M", time.localtime()))
generate_tb_dir = config.tb_dir + '/tb'
logger = SummaryWriter(log_dir=tb_dir)
engine.link_tb(tb_dir, generate_tb_dir)
# config network and criterion
criterion = nn.CrossEntropyLoss(reduction='mean',
ignore_index=255)
if engine.distributed:
BatchNorm2d = SyncBatchNorm
model = Network(class_num=config.num_classes, feature=128, bn_momentum=config.bn_momentum,
torch.manual_seed(config.seed)
if torch.cuda.is_available():
torch.cuda.manual_seed(config.seed)
parser = argparse.ArgumentParser()
with Engine(custom_parser=parser) as engine:
args = parser.parse_args()
cudnn.benchmark = True
if engine.distributed:
torch.cuda.set_device(engine.local_rank)
# data loader
train_loader, train_sampler = get_train_loader(engine, Scut)
# config network and criterion
criterion = nn.CrossEntropyLoss(reduction='mean',
ignore_index=255) # 默认使用交叉熵损失
edge_criterion = EdgeLoss(ignore_label=255)
'''
min_kept = int(config.batch_size // len(
engine.devices) * config.image_height * config.image_width // 16)
criterion = ProbOhemCrossEntropy2d(ignore_label=255, thresh=0.7,
min_kept=min_kept,
use_weight=False)
'''
if engine.distributed:
logger.info('Use the Multi-Process-SyncBatchNorm')
cudnn.benchmark = True
torch.manual_seed(123)
if torch.cuda.is_available():
torch.cuda.manual_seed(123)
device = torch.device('cuda') if CFG.cuda else torch.device('cpu')
for path in [CFG.result_dir, CFG.model_dir]:
if not os.path.isdir(path):
os.mkdir(path)
print('Setup ImageLoader')
loader = get_train_loader(CFG.color_path,
CFG.batch_size,
resize=True,
size=(512, 512))
print('Define Network')
netG = define_U(device, in_channel=1, out_channel=3).to(device)
netD = define_D(device, 3).to(device)
print('Define Optimizer')
optG = optim.Adam(netG.parameters(), lr=CFG.lr, betas=CFG.betas)
optD = optim.Adam(netD.parameters(), lr=CFG.lr, betas=CFG.betas)
if CFG.is_load_model:
try:
saved = torch.load(os.path.join(CFG.model_dir, CFG.model_name))
start_epoch = saved['epoch']
netG.load_state_dict(saved['netG'])
def process(args, train_id, test_loader):
model = GazeNet()
logger.info(json.dumps(vars(args), indent=2))
# TensorBoard SummaryWriter
# writer = SummaryWriter() if args.tensorboard else None
writer = None
# set random seed
seed = args.seed
torch.manual_seed(seed)
np.random.seed(seed)
random.seed(seed)
# create output directory
outdir = args.outdir
if not os.path.exists(outdir):
os.makedirs(outdir)
outpath = os.path.join(outdir, 'config.json')
with open(outpath, 'w') as fout:
json.dump(vars(args), fout, indent=2)
# data loaders
train_loader = get_train_loader(args.dataset, train_id, args.batch_size,
args.num_workers, True)
print('data collected')
# model
model_path = os.path.join(outdir, 'model_state.pth')
model = GazeNet()
model = nn.DataParallel(model)
model.cuda()
if os.path.exists(os.path.join(outdir, 'model_state.pth')):
model.load_state_dict(torch.load(model_path))
model.eval()
else:
None
criterion = nn.MSELoss(size_average=True)
# optimizer
optimizer = torch.optim.SGD(model.parameters(),
lr=args.base_lr,
momentum=args.momentum,
weight_decay=args.weight_decay,
nesterov=args.nesterov)
scheduler = torch.optim.lr_scheduler.MultiStepLR(optimizer,
milestones=[20, 30],
gamma=args.lr_decay)
# Tensorboard Settings
# config = {
# 'tensorboard': args.tensorboard,
# 'tensorboard_images': args.tensorboard_images,
# 'tensorboard_parameters': args.tensorboard_parameters,
# }
config = None
# run test before start training
test(0, model, criterion, test_loader, config, writer)
for epoch in range(1, args.epochs + 1):
scheduler.step()
train(epoch, model, optimizer, criterion, train_loader, config, writer)
angle_error = test(epoch, model, criterion, test_loader, config,
writer)
state = OrderedDict([
('args', vars(args)),
('state_dict', model.state_dict()),
('optimizer', optimizer.state_dict()),
('epoch', epoch),
('angle_error', angle_error),
])
torch.save(state, model_path)