def train_process_thor(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_2
os.environ['RANK_ID'] = str(device_id - 4)
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], "hccl_world_groupsum1")
auto_parallel_context().set_all_reduce_fusion_split_indices(
[27], "hccl_world_groupsum2")
auto_parallel_context().set_all_reduce_fusion_split_indices(
[27], "hccl_world_groupsum3")
auto_parallel_context().set_all_reduce_fusion_split_indices(
[27], "hccl_world_groupsum4")
auto_parallel_context().set_all_reduce_fusion_split_indices(
[27], "hccl_world_groupsum5")
init()
# network
damping = get_model_damping(0, 0.03, 0.87, 50, 5004)
net = resnet50_thor(class_num=thor_config.class_num,
damping=damping,
loss_scale=thor_config.loss_scale,
frequency=thor_config.frequency)
# evaluation network
dist_eval_network = ClassifyCorrectCell(net)
if not thor_config.label_smooth:
thor_config.label_smooth_factor = 0.0
# loss
loss = nn.SoftmaxCrossEntropyWithLogits(
sparse=True,
reduction="mean",
smooth_factor=thor_config.label_smooth_factor,
num_classes=thor_config.class_num)
# train dataset
dataset = create_dataset(dataset_path=dataset_path,
do_train=True,
repeat_num=epoch_size,
batch_size=thor_config.batch_size)
step_size = dataset.get_dataset_size()
eval_interval = thor_config.eval_interval
# evalutation dataset
eval_dataset = create_dataset(dataset_path=eval_path,
do_train=False,
repeat_num=epoch_size,
batch_size=thor_config.eval_batch_size)
# loss scale
loss_scale = FixedLossScaleManager(thor_config.loss_scale,
drop_overflow_update=False)
# learning rate
lr = Tensor(get_model_lr(0, 0.045, 6, 70, 5004))
# optimizer
opt = THOR(filter(lambda x: x.requires_grad,
net.get_parameters()), lr, thor_config.momentum,
filter(lambda x: 'matrix_A' in x.name, net.get_parameters()),
filter(lambda x: 'matrix_G' in x.name, net.get_parameters()),
filter(lambda x: 'A_inv_max' in x.name, net.get_parameters()),
filter(lambda x: 'G_inv_max' in x.name, net.get_parameters()),
thor_config.weight_decay, thor_config.loss_scale)
# model
model = THOR_Model(net,
loss_fn=loss,
optimizer=opt,
loss_scale_manager=loss_scale,
amp_level="O2",
keep_batchnorm_fp32=False,
metrics={
'acc':
DistAccuracy(batch_size=thor_config.eval_batch_size,
device_num=device_num)
},
eval_network=dist_eval_network,
frequency=thor_config.frequency)
# 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
acc = 0.0
time_cost = 0.0
print("run_start", device_id)
for epoch_idx in range(0, int(epoch_size / eval_interval)):
model.train(eval_interval, 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 train_process_thor(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_2
os.environ['RANK_ID'] = str(device_id - 4)
os.environ['RANK_SIZE'] = str(device_num)
if enable_hccl:
context.set_auto_parallel_context(
device_num=device_num,
parallel_mode=ParallelMode.DATA_PARALLEL,
gradients_mean=True,
all_reduce_fusion_config=[85, 160])
init()
# network
net = resnet50_thor(thor_config.class_num)
if not thor_config.label_smooth:
thor_config.label_smooth_factor = 0.0
# loss
loss = CrossEntropySmooth(sparse=True,
reduction="mean",
smooth_factor=thor_config.label_smooth_factor,
num_classes=thor_config.class_num)
# train dataset
dataset = create_dataset_thor(dataset_path=dataset_path,
do_train=True,
repeat_num=1,
batch_size=thor_config.batch_size)
step_size = dataset.get_dataset_size()
eval_interval = thor_config.eval_interval
# evaluation dataset
eval_dataset = create_dataset(dataset_path=eval_path,
do_train=False,
repeat_num=1,
batch_size=thor_config.eval_batch_size)
# loss scale
loss_scale = FixedLossScaleManager(thor_config.loss_scale,
drop_overflow_update=False)
# learning rate
lr = get_thor_lr(0, 0.05803, 4.04839, 53, 5004, decay_epochs=39)
damping = get_thor_damping(0, 0.02714, 0.50036, 70, 5004)
# optimizer
split_indices = [26, 53]
opt = THOR(net,
Tensor(lr),
Tensor(damping),
thor_config.momentum,
thor_config.weight_decay,
thor_config.loss_scale,
thor_config.batch_size,
split_indices=split_indices)
# evaluation network
dist_eval_network = ClassifyCorrectCell(net)
# model
model = THOR_Model(net,
loss_fn=loss,
optimizer=opt,
loss_scale_manager=loss_scale,
amp_level="O2",
keep_batchnorm_fp32=False,
metrics={
'acc':
DistAccuracy(batch_size=thor_config.eval_batch_size,
device_num=device_num)
},
eval_network=dist_eval_network,
frequency=thor_config.frequency)
# 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
acc = 0.0
time_cost = 0.0
print("run_start", device_id)
for epoch_idx in range(0, int(epoch_size / eval_interval)):
model.train(eval_interval, 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})