def test_resetnet50_build():
X = me.tensor()
Y = me.tensor()
X.set_dtype(dtype.float32)
Y.set_dtype(dtype.float32)
network = resnet50()
model = Model(network=network, loss_fn=loss_func, optimizer=optimizer)
def test_amp_resnet50_loss():
inputs = Tensor(np.ones([2, 3, 224, 224]).astype(np.float32))
label = Tensor(np.zeros([2, 10]).astype(np.float32))
net = resnet50()
loss = nn.SoftmaxCrossEntropyWithLogits(reduction='mean')
optimizer = nn.Momentum(net.trainable_params(), learning_rate=0.1, momentum=0.9)
train_network = amp.build_train_network(net, optimizer, loss, level="O2")
train_network(inputs, label)
def resnet50_train(args_opt):
epoch_size = args_opt.epoch_size
batch_size = 32
class_num = 10
loss_scale_num = 1024
local_data_path = '/cache/data'
# set graph mode and parallel mode
context.set_context(mode=context.GRAPH_MODE, device_target="Ascend", save_graphs=False)
context.set_context(enable_task_sink=True, device_id=device_id)
context.set_context(enable_loop_sink=True)
context.set_context(enable_mem_reuse=True)
if device_num > 1:
context.set_auto_parallel_context(device_num=device_num,
parallel_mode=ParallelMode.DATA_PARALLEL,
mirror_mean=True)
local_data_path = os.path.join(local_data_path, str(device_id))
# data download
print('Download data.')
mox.file.copy_parallel(src_url=args_opt.data_url, dst_url=local_data_path)
# create dataset
print('Create train and evaluate dataset.')
train_dataset = create_dataset(dataset_path=local_data_path, do_train=True,
repeat_num=epoch_size, batch_size=batch_size)
eval_dataset = create_dataset(dataset_path=local_data_path, do_train=False,
repeat_num=1, batch_size=batch_size)
train_step_size = train_dataset.get_dataset_size()
print('Create dataset success.')
# create model
net = resnet50(class_num = class_num)
loss = SoftmaxCrossEntropyWithLogits(sparse=True)
lr = Tensor(get_lr(global_step=0, total_epochs=epoch_size, steps_per_epoch=train_step_size))
opt = Momentum(net.trainable_params(), lr, momentum=0.9, weight_decay=1e-4, loss_scale=loss_scale_num)
loss_scale = FixedLossScaleManager(loss_scale_num, False)
model = Model(net, loss_fn=loss, optimizer=opt, loss_scale_manager=loss_scale, metrics={'acc'})
# define performance callback to show ips and loss callback to show loss for every epoch
performance_cb = PerformanceCallback(batch_size)
loss_cb = LossMonitor()
cb = [performance_cb, loss_cb]
print(f'Start run training, total epoch: {epoch_size}.')
model.train(epoch_size, train_dataset, callbacks=cb)
if device_num == 1 or device_id == 0:
print(f'Start run evaluation.')
output = model.eval(eval_dataset)
print(f'Evaluation result: {output}.')
save_graphs=False)
context.set_context(enable_task_sink=True, device_id=device_id)
context.set_context(enable_loop_sink=True)
context.set_context(enable_mem_reuse=True)
if __name__ == '__main__':
if not args_opt.do_eval and args_opt.run_distribute:
context.set_auto_parallel_context(
device_num=args_opt.device_num,
parallel_mode=ParallelMode.DATA_PARALLEL,
mirror_mean=True)
auto_parallel_context().set_all_reduce_fusion_split_indices([140])
init()
epoch_size = config.epoch_size
net = resnet50(class_num=config.class_num)
loss = SoftmaxCrossEntropyWithLogits(sparse=True)
if args_opt.do_train:
dataset = create_dataset(dataset_path=args_opt.dataset_path,
do_train=True,
repeat_num=epoch_size,
batch_size=config.batch_size)
step_size = dataset.get_dataset_size()
loss_scale = FixedLossScaleManager(config.loss_scale,
drop_overflow_update=False)
lr = Tensor(
get_lr(global_step=0,
lr_init=config.lr_init,
lr_end=config.lr_end,
def train_process(q, device_id, epoch_size, device_num, enable_hccl):
os.system("mkdir " + str(device_id))
os.chdir(str(device_id))
context.set_context(mode=context.GRAPH_MODE,
device_target="Ascend",
save_graphs=False)
context.set_context(device_id=device_id)
os.environ['MINDSPORE_HCCL_CONFIG_PATH'] = MINDSPORE_HCCL_CONFIG_PATH
os.environ['RANK_ID'] = str(device_id)
os.environ['RANK_SIZE'] = str(device_num)
if enable_hccl:
context.set_auto_parallel_context(
device_num=device_num,
parallel_mode=ParallelMode.DATA_PARALLEL,
mirror_mean=True,
parameter_broadcast=True)
auto_parallel_context().set_all_reduce_fusion_split_indices([107, 160])
init()
# network
net = resnet50(class_num=config.class_num)
# evaluation network
dist_eval_network = ClassifyCorrectCell(net)
if not config.use_label_smooth:
config.label_smooth_factor = 0.0
# loss
loss = nn.SoftmaxCrossEntropyWithLogits(
sparse=True,
reduction="mean",
smooth_factor=config.label_smooth_factor,
num_classes=config.class_num)
# train dataset
dataset = create_dataset(dataset_path=dataset_path,
do_train=True,
repeat_num=epoch_size,
batch_size=config.batch_size)
step_size = dataset.get_dataset_size()
eval_interval = config.eval_interval
dataset.__loop_size__ = step_size * eval_interval
# evalutation dataset
eval_dataset = create_dataset(dataset_path=eval_path,
do_train=False,
repeat_num=epoch_size,
batch_size=config.eval_batch_size)
# loss scale
loss_scale = FixedLossScaleManager(config.loss_scale,
drop_overflow_update=False)
# learning rate
lr = Tensor(
get_learning_rate(lr_init=config.lr_init,
lr_end=0.0,
lr_max=config.lr_max,
warmup_epochs=config.warmup_epochs,
total_epochs=config.epoch_size,
steps_per_epoch=step_size,
lr_decay_mode=config.lr_decay_mode))
# optimizer
decayed_params = list(
filter(
lambda x: 'beta' not in x.name and 'gamma' not in x.name and 'bias'
not in x.name, net.trainable_params()))
no_decayed_params = [
param for param in net.trainable_params()
if param not in decayed_params
]
group_params = [{
'params': decayed_params,
'weight_decay': config.weight_decay
}, {
'params': no_decayed_params,
'weight_decay': 0.0
}, {
'order_params': net.trainable_params()
}]
if config.use_lars:
momentum = nn.Momentum(group_params,
lr,
config.momentum,
loss_scale=config.loss_scale,
use_nesterov=config.use_nesterov)
opt = nn.LARS(momentum,
epsilon=config.lars_epsilon,
coefficient=config.lars_coefficient,
lars_filter=lambda x: 'beta' not in x.name and 'gamma'
not in x.name and 'bias' not in x.name)
else:
opt = nn.Momentum(group_params,
lr,
config.momentum,
loss_scale=config.loss_scale,
use_nesterov=config.use_nesterov)
# model
model = Model(net,
loss_fn=loss,
optimizer=opt,
loss_scale_manager=loss_scale,
amp_level="O2",
keep_batchnorm_fp32=False,
metrics={
'acc':
DistAccuracy(batch_size=config.eval_batch_size,
device_num=device_num)
},
eval_network=dist_eval_network)
# model init
print("init_start", device_id)
model.init(dataset, eval_dataset)
print("init_stop", device_id)
# callbacks
loss_cb = LossGet(1, step_size)
# train and eval
print("run_start", device_id)
acc = 0.0
time_cost = 0.0
for epoch_idx in range(0, int(epoch_size / eval_interval)):
model.train(1, dataset, callbacks=loss_cb)
eval_start = time.time()
output = model.eval(eval_dataset)
eval_cost = (time.time() - eval_start) * 1000
acc = float(output["acc"])
time_cost = loss_cb.get_per_step_time()
loss = loss_cb.get_loss()
print(
"the {} epoch's resnet result:\n "
"device{}, training loss {}, acc {}, "
"training per step cost {:.2f} ms, eval cost {:.2f} ms, total_cost {:.2f} ms"
.format(epoch_idx, device_id, loss, acc, time_cost, eval_cost,
time_cost * step_size + eval_cost))
q.put({'acc': acc, 'cost': time_cost})
def test_resnet_construct(x):
# not right model to import
network = resnet50()
return network.construct(x)
