def __init__(self,
model,
ema_network,
eval_dataset,
loss_fn,
decay=0.999,
save_epoch=1,
dataset_sink_mode=True,
start_epoch=0):
self.model = model
self.ema_network = ema_network
self.eval_dataset = eval_dataset
self.loss_fn = loss_fn
self.decay = decay
self.save_epoch = save_epoch
self.shadow = {}
self.ema_accuracy = {}
self.best_ema_accuracy = 0
self.best_accuracy = 0
self.best_ema_epoch = 0
self.best_epoch = 0
self._start_epoch = start_epoch
self.eval_metrics = {
'Validation-Loss': Loss(),
'Top1-Acc': Top1CategoricalAccuracy(),
'Top5-Acc': Top5CategoricalAccuracy()
}
self.dataset_sink_mode = dataset_sink_mode
def main():
"""Main entrance for training"""
args = parser.parse_args()
print(sys.argv)
#context.set_context(mode=context.GRAPH_MODE)
context.set_context(mode=context.PYNATIVE_MODE)
if args.GPU:
context.set_context(device_target='GPU')
# parse model argument
assert args.model.startswith(
"hournas"), "Only Tinynet models are supported."
#_, sub_name = args.model.split("_")
net = hournasnet(args.model,
num_classes=args.num_classes,
drop_rate=0.0,
drop_connect_rate=0.0,
global_pool="avg",
bn_tf=False,
bn_momentum=None,
bn_eps=None)
print(net)
print("Total number of parameters:", count_params(net))
cfg = edict({
'image_height': args.image_size,
'image_width': args.image_size,
})
cfg.batch_size = args.batch_size
print(cfg)
#input_size = net.default_cfg['input_size'][1]
val_data_url = args.data_path #os.path.join(args.data_path, 'val')
val_dataset = create_dataset_cifar10(val_data_url,
repeat_num=1,
training=False,
cifar_cfg=cfg)
loss = nn.SoftmaxCrossEntropyWithLogits(sparse=True, reduction='mean')
eval_metrics = {
'Validation-Loss': Loss(),
'Top1-Acc': Top1CategoricalAccuracy(),
'Top5-Acc': Top5CategoricalAccuracy()
}
ckpt = load_checkpoint(args.ckpt)
load_param_into_net(net, ckpt)
net.set_train(False)
model = Model(net, loss, metrics=eval_metrics)
metrics = model.eval(val_dataset, dataset_sink_mode=False)
print(metrics)
def main():
"""Main entrance for training"""
args = parser.parse_args()
print(sys.argv)
context.set_context(mode=context.GRAPH_MODE)
if args.GPU:
context.set_context(device_target='GPU')
# parse model argument
assert args.model.startswith(
"tinynet"), "Only Tinynet models are supported."
_, sub_name = args.model.split("_")
net = tinynet(sub_model=sub_name,
num_classes=args.num_classes,
drop_rate=0.0,
drop_connect_rate=0.0,
global_pool="avg",
bn_tf=False,
bn_momentum=None,
bn_eps=None)
print("Total number of parameters:", count_params(net))
input_size = net.default_cfg['input_size'][1]
val_data_url = os.path.join(args.data_path, 'val')
val_dataset = create_dataset_val(args.batch_size,
val_data_url,
workers=args.workers,
distributed=False,
input_size=input_size)
loss = LabelSmoothingCrossEntropy(smooth_factor=args.smoothing,
num_classes=args.num_classes)
loss.add_flags_recursive(fp32=True, fp16=False)
eval_metrics = {
'Validation-Loss': Loss(),
'Top1-Acc': Top1CategoricalAccuracy(),
'Top5-Acc': Top5CategoricalAccuracy()
}
ckpt = load_checkpoint(args.ckpt)
load_param_into_net(net, ckpt)
net.set_train(False)
model = Model(net, loss, metrics=eval_metrics)
metrics = model.eval(val_dataset, dataset_sink_mode=False)
print(metrics)
def main():
"""Main entrance for training"""
args = parser.parse_args()
print(sys.argv)
#context.set_context(mode=context.GRAPH_MODE)
context.set_context(mode=context.PYNATIVE_MODE)
if args.GPU:
context.set_context(device_target='GPU', device_id=args.device_id)
# parse model argument
assert args.model.startswith(
"hournas"), "Only Tinynet models are supported."
net = nasbenchnet()
cfg = edict({
'image_height': args.image_size,
'image_width': args.image_size,
})
cfg.batch_size = args.batch_size
val_data_url = args.data_path
val_dataset = create_dataset_cifar10(val_data_url,
repeat_num=1,
training=False,
cifar_cfg=cfg)
loss = LabelSmoothingCrossEntropy(smooth_factor=args.smoothing,
num_classes=args.num_classes)
loss.add_flags_recursive(fp32=True, fp16=False)
eval_metrics = {
'Validation-Loss': Loss(),
'Top1-Acc': Top1CategoricalAccuracy(),
'Top5-Acc': Top5CategoricalAccuracy()
}
ckpt = load_checkpoint(args.ckpt)
load_param_into_net(net, ckpt)
net.set_train(False)
model = Model(net, loss, metrics=eval_metrics)
metrics = model.eval(val_dataset, dataset_sink_mode=False)
print(metrics)
def main():
"""Main entrance for training"""
args = parser.parse_args()
print(sys.argv)
devid, args.rank_id, args.rank_size = 0, 0, 1
context.set_context(mode=context.GRAPH_MODE)
if args.distributed:
if args.GPU:
init("nccl")
context.set_context(device_target='GPU')
else:
init()
devid = int(os.getenv('DEVICE_ID'))
context.set_context(device_target='Ascend',
device_id=devid,
reserve_class_name_in_scope=False)
context.reset_auto_parallel_context()
args.rank_id = get_rank()
args.rank_size = get_group_size()
context.set_auto_parallel_context(
parallel_mode=ParallelMode.DATA_PARALLEL,
gradients_mean=True,
device_num=args.rank_size)
else:
if args.GPU:
context.set_context(device_target='GPU')
is_master = not args.distributed or (args.rank_id == 0)
# parse model argument
assert args.model.startswith(
"tinynet"), "Only Tinynet models are supported."
_, sub_name = args.model.split("_")
net = tinynet(sub_model=sub_name,
num_classes=args.num_classes,
drop_rate=args.drop,
drop_connect_rate=args.drop_connect,
global_pool="avg",
bn_tf=args.bn_tf,
bn_momentum=args.bn_momentum,
bn_eps=args.bn_eps)
if is_master:
print("Total number of parameters:", count_params(net))
# input image size of the network
input_size = net.default_cfg['input_size'][1]
train_dataset = val_dataset = None
train_data_url = os.path.join(args.data_path, 'train')
val_data_url = os.path.join(args.data_path, 'val')
val_dataset = create_dataset_val(args.batch_size,
val_data_url,
workers=args.workers,
distributed=False,
input_size=input_size)
if args.train:
train_dataset = create_dataset(args.batch_size,
train_data_url,
workers=args.workers,
distributed=args.distributed,
input_size=input_size)
batches_per_epoch = train_dataset.get_dataset_size()
loss = LabelSmoothingCrossEntropy(smooth_factor=args.smoothing,
num_classes=args.num_classes)
time_cb = TimeMonitor(data_size=batches_per_epoch)
loss_scale_manager = FixedLossScaleManager(args.loss_scale,
drop_overflow_update=False)
lr_array = get_lr(base_lr=args.lr,
total_epochs=args.epochs,
steps_per_epoch=batches_per_epoch,
decay_epochs=args.decay_epochs,
decay_rate=args.decay_rate,
warmup_epochs=args.warmup_epochs,
warmup_lr_init=args.warmup_lr,
global_epoch=0)
lr = Tensor(lr_array)
loss_cb = LossMonitor(lr_array,
args.epochs,
per_print_times=args.per_print_times,
start_epoch=0)
param_group = add_weight_decay(net, weight_decay=args.weight_decay)
if args.opt == 'sgd':
if is_master:
print('Using SGD optimizer')
optimizer = SGD(param_group,
learning_rate=lr,
momentum=args.momentum,
weight_decay=args.weight_decay,
loss_scale=args.loss_scale)
elif args.opt == 'rmsprop':
if is_master:
print('Using rmsprop optimizer')
optimizer = RMSProp(param_group,
learning_rate=lr,
decay=0.9,
weight_decay=args.weight_decay,
momentum=args.momentum,
epsilon=args.opt_eps,
loss_scale=args.loss_scale)
loss.add_flags_recursive(fp32=True, fp16=False)
eval_metrics = {
'Validation-Loss': Loss(),
'Top1-Acc': Top1CategoricalAccuracy(),
'Top5-Acc': Top5CategoricalAccuracy()
}
if args.ckpt:
ckpt = load_checkpoint(args.ckpt)
load_param_into_net(net, ckpt)
net.set_train(False)
model = Model(net,
loss,
optimizer,
metrics=eval_metrics,
loss_scale_manager=loss_scale_manager,
amp_level=args.amp_level)
net_ema = copy.deepcopy(net)
net_ema.set_train(False)
assert args.ema_decay > 0, "EMA should be used in tinynet training."
ema_cb = EmaEvalCallBack(network=net,
ema_network=net_ema,
loss_fn=loss,
eval_dataset=val_dataset,
decay=args.ema_decay,
save_epoch=args.ckpt_save_epoch,
dataset_sink_mode=args.dataset_sink,
start_epoch=0)
callbacks = [loss_cb, ema_cb, time_cb] if is_master else []
if is_master:
print("Training on " + args.model + " with " + str(args.num_classes) +
" classes")
model.train(args.epochs,
train_dataset,
callbacks=callbacks,
dataset_sink_mode=args.dataset_sink)