def test_two_matmul_transpose():
class Net(nn.Cell):
def __init__(self):
super().__init__()
self.matmul1 = P.MatMul()
self.matmul2 = P.MatMul()
self.transpose1 = P.Transpose()
self.transpose2 = P.Transpose()
def construct(self, x, y, b):
out = self.matmul1(x, y)
out = self.matmul2(out, b)
out = self.transpose1(out, (1, 0))
out = self.transpose2(out, (1, 0))
return out
size = 16
context.set_auto_parallel_context(device_num=size, global_rank=0)
x = Tensor(np.ones([128, 32]), dtype=ms.float32)
y = Tensor(np.ones([32, 64]), dtype=ms.float32)
b = Tensor(np.ones([64, 64]), dtype=ms.float32)
net = NetWithLoss(Net())
context.set_auto_parallel_context(parallel_mode="auto_parallel")
net.set_auto_parallel()
reset_op_id()
_executor.compile(net, x, y, b, phase='train')
strategies = _executor._get_strategy(net)
expected_strategies = {'Default/network-Net/Transpose-op0': [[1, 16]],
'Default/network-Net/Transpose-op1': [[16, 1]],
'Default/network-Net/MatMul-op2': [[16, 1], [1, 1]],
'Default/network-Net/MatMul-op3': [[16, 1], [1, 1]]}
assert strategies == expected_strategies
def test_train_64k_8p(epoch_size=3,
batch_size=32,
num_classes=65536): #1048576 #131072 #32768 #8192
dev_num = 8
context.set_auto_parallel_context(parallel_mode=ParallelMode.AUTO_PARALLEL,
device_num=dev_num)
cost_model_context.set_cost_model_context(costmodel_gamma=0.001,
costmodel_beta=400.0)
set_algo_parameters(elementwise_op_strategy_follow=True)
resset_op_id()
np.random.seed(6)
input_np = np.ones([batch_size, 3, 224, 224]).astype(np.float32)
label_np = np.zeros([batch_size]).astype(np.int32)
for i in range(0, batch_size):
label_np[i] = i % num_classes
dataset = DatasetLenet(Tensor(input_np), Tensor(label_np), 1)
net = resnet50(num_classes)
loss = SoftmaxCrossEntropyExpand(sparse=True)
opt = Momentum(filter(lambda x: x.requires_grad, net.get_parameters()),
0.01, 0.9)
model = Model(net, loss_fn=loss, optimizer=opt)
model.train(5, dataset, dataset_sink_mode=False)
strategies = _executor._get_strategy(model._train_network)
for (k, v) in strategies.items():
if re.search('Conv2D-op', k) is not None:
assert v[0][0] == dev_num
elif re.search('MatMul-op', k) is not None:
assert v == [[1, 1], [dev_num, 1]]
elif re.search('ReduceSum-op', k) is not None:
assert v == [[1, dev_num]]
def test_auto_parallel_arithmetic_broadcast_both():
class Net(nn.Cell):
def __init__(self):
super().__init__()
self.matmul = P.MatMul()
self.floordiv = P.FloorDiv()
def construct(self, x, y, b):
out = self.matmul(x, y)
out = self.floordiv(out, b)
return out
context.set_auto_parallel_context(device_num=8, global_rank=0)
net = NetWithLoss(Net())
context.set_auto_parallel_context(parallel_mode="auto_parallel")
reset_op_id()
x = Tensor(np.ones([64, 32]), dtype=ms.float32)
y = Tensor(np.ones([32, 1]), dtype=ms.float32)
b = Tensor(np.ones([1, 64]), dtype=ms.float32)
compile_net(net, x, y, b, phase='train')
strategies = _executor._get_strategy(net)
expected_strategies = {
'Default/network-Net/FloorDiv-op0': [[8, 1], [1, 1]],
'Default/network-Net/MatMul-op1': [[8, 1], [1, 1]]
}
assert strategies == expected_strategies
def test_matmul_prelu():
class Net(nn.Cell):
def __init__(self):
super().__init__()
self.mul1 = P.Mul()
self.prelu = P.PReLU()
def construct(self, x, y, b):
out = self.mul1(x, y)
out = self.prelu(out, b)
return out
size = 16
context.set_auto_parallel_context(device_num=size, global_rank=0)
x = Tensor(np.ones([16, 3, 128, 32]), dtype=ms.float32)
y = Tensor(np.ones([16, 3, 128, 32]), dtype=ms.float32)
b = Tensor(np.array([0.01, 0.02, 0.03]), dtype=ms.float32)
net = NetWithLoss(Net())
context.set_auto_parallel_context(parallel_mode="auto_parallel")
net.set_auto_parallel()
reset_op_id()
_executor.compile(net, x, y, b, phase='train')
strategies = _executor._get_strategy(net)
for (k, v) in strategies.items():
if re.search('PReLU-op', k) is not None:
assert v == [[16, 1, 1, 1], [1]]
elif re.search('Mul-op', k) is not None:
assert v == [[16, 1, 1, 1], [16, 1, 1, 1]]
def test_matmul_prelu():
class Net(nn.Cell):
def __init__(self):
super().__init__()
self.mul1 = P.Mul()
self.prelu = P.PReLU()
def construct(self, x, y, b):
out = self.mul1(x, y)
out = self.prelu(out, b)
return out
size = 16
context.set_auto_parallel_context(device_num=size, global_rank=0)
x = Tensor(np.ones([16, 3, 128, 32]), dtype=ms.float32)
y = Tensor(np.ones([16, 3, 128, 32]), dtype=ms.float32)
b = Tensor(np.array([0.01, 0.02, 0.03]), dtype=ms.float32)
net = NetWithLoss(Net())
context.set_auto_parallel_context(parallel_mode="auto_parallel")
reset_op_id()
_executor.compile(net, x, y, b, phase='train')
strategies = _executor._get_strategy(net)
assert strategies['Default/network-Net/PReLU-op2'] == [[16, 1, 1, 1], [1]]
assert strategies['Default/network-Net/Mul-op3'] == [[16, 1, 1, 1],
[16, 1, 1, 1]]
def test_common_parameter():
class Net(nn.Cell):
def __init__(self):
super().__init__()
self.matmul1 = P.MatMul()
self.matmul2 = P.MatMul()
self.matmul3 = P.MatMul()
self.weight1 = Parameter(Tensor(
np.ones([64, 64]).astype(np.float16) * 0.01),
"w",
requires_grad=True)
self.cast1 = P.Cast()
self.cast2 = P.Cast()
def construct(self, x, y, z, w):
m1_result = self.matmul1(x, self.cast1(self.weight1,
mstype.float32))
m2_result = self.matmul2(z, self.cast2(self.weight1,
mstype.float32))
m3_result = self.matmul3(m2_result, m1_result)
return m3_result
size = 8
context.set_auto_parallel_context(device_num=size, global_rank=0)
set_algo_parameters(elementwise_op_strategy_follow=True)
x = Tensor(np.ones([64, 64]), dtype=ms.float32)
y = Tensor(np.ones([64, 64]), dtype=ms.float32)
z = Tensor(np.ones([64, 64]), dtype=ms.float32)
w = Tensor(np.ones([64, 64]), dtype=ms.float32)
net = NetWithLoss(Net())
context.set_auto_parallel_context(parallel_mode="auto_parallel")
net.set_auto_parallel()
reset_op_id()
_executor.compile(net, x, y, z, w, phase='train')
strategies = _executor._get_strategy(net)
expected_strategies = {
'Default/network-Net/MatMul-op1': [[8, 1], [1, 1]],
'Default/network-Net/MatMul-op3': [[8, 1], [1, 1]],
'Default/network-Net/Cast-op2': [[1, 1]],
'Default/network-Net/MatMul-op0': [[8, 1], [1, 1]],
'Default/network-Net/Cast-op4': [[1, 1]]
}
assert strategies == expected_strategies
def test_double_subgraphs():
cost_model_context.set_cost_model_context(multi_subgraphs=True)
context.set_context(save_graphs=True)
context.set_auto_parallel_context(device_num=8, global_rank=0)
net = TrainStepWarp(NetWithLoss(Net()))
context.set_auto_parallel_context(parallel_mode="auto_parallel")
x = Tensor(np.ones([8, 8, 8, 8]), dtype=ms.float32)
reset_op_id()
_executor.compile(net, x, phase='train')
strategies = _executor._get_strategy(net)
expected_strategies = {'Default/network-NetWithLoss/ReduceMean-op0': [[8, 1, 1, 1]],
'Default/network-NetWithLoss/net-Net/ReLU-op1': [[8, 1, 1, 1]],
'Default/network-NetWithLoss/net-Net/Mul-op2': [[8, 1, 1, 1], [8, 1, 1, 1]],
'Default/network-NetWithLoss/net-Net/Mul-op3': [[8, 1, 1, 1], [8, 1, 1, 1]],
'Default/network-NetWithLoss/ReduceSum-op4': [[8, 1, 1, 1]]}
assert strategies == expected_strategies
def test_auto_parallel_assign_sub_with_ref_key():
size = 8
context.set_auto_parallel_context(device_num=size, global_rank=0)
x = Tensor(np.random.rand(4, 4, 32, 64), dtype=ms.float32)
net = NetWithLoss(nn.PReLU(4))
context.set_auto_parallel_context(parallel_mode="auto_parallel")
reset_op_id()
_executor.compile(net, x, phase="train")
strategies = _executor._get_strategy(net)
for (k, v) in strategies.items():
if re.search('PReLU-op', k) is not None:
assert v == [[1, 1, 1, 8], [1]]
elif re.search('ReLU-op', k) is not None:
assert v == [[1]]
def test_auto_parallel_assign_sub_with_ref_key():
size = 8
context.set_auto_parallel_context(device_num=size, global_rank=0)
x = Tensor(np.random.rand(4, 4, 32, 64), dtype=ms.float32)
net = NetWithLoss(nn.PReLU(4))
context.set_auto_parallel_context(parallel_mode="auto_parallel")
reset_op_id()
_executor.compile(net, x, phase="train")
strategies = _executor._get_strategy(net)
expected_strategies = {
'Default/network-PReLU/PReLU-op2': [[1, 1, 1, 8], [1]],
'Default/network-PReLU/ReLU-op3': [[1]]
}
assert strategies == expected_strategies
def all_to_all_common():
learning_rate = 0.1
momentum = 0.9
epoch_size = 2
context.reset_auto_parallel_context()
context.set_auto_parallel_context(parallel_mode=ParallelMode.AUTO_PARALLEL, device_num=1, global_rank=0)
predict = Tensor(np.ones([32, 128]), dtype=ms.float32)
label = Tensor(np.ones([32]), dtype=ms.int32)
dataset = Dataset(predict, label, 2)
net = all_to_all_net()
loss = SoftmaxCrossEntropyWithLogits(is_grad=False, sparse=True)
opt = Momentum(net.trainable_params(), learning_rate, momentum)
model = Model(net, loss, opt)
model.train(epoch_size, dataset, dataset_sink_mode=False)
strategys = _executor._get_strategy(model._train_network)
return strategys
def test_double_star_graph():
class Net(nn.Cell):
def __init__(self):
super().__init__()
self.matmul1 = P.MatMul()
self.matmul2 = P.MatMul()
self.matmul3 = P.MatMul()
self.cast1 = P.Cast()
self.cast2 = P.Cast()
def construct(self, x, y, z, w):
m1_result = self.matmul1(x, y)
m2_result = self.matmul2(z, w)
m3_result = self.matmul3(self.cast1(m2_result, mstype.float16),
self.cast2(m1_result, mstype.float16))
return m3_result
size = 8
context.set_auto_parallel_context(device_num=size, global_rank=0)
x = Tensor(np.ones([32, 8]), dtype=ms.float32)
y = Tensor(np.ones([8, 16]), dtype=ms.float32)
z = Tensor(np.ones([8, 16]), dtype=ms.float32)
w = Tensor(np.ones([16, 32]), dtype=ms.float32)
net = NetWithLoss(Net())
context.set_auto_parallel_context(parallel_mode="auto_parallel")
net.set_auto_parallel()
reset_op_id()
_executor.compile(net, x, y, z, w, phase='train')
strategies = _executor._get_strategy(net)
expected_strategies = {
'Default/network-Net/Cast-op0': [[8, 1]],
'Default/network-Net/Cast-op1': [[1, 8]],
'Default/network-Net/MatMul-op3': [[8, 1], [1, 1]],
'Default/network-Net/MatMul-op4': [[1, 1], [1, 8]],
'Default/network-Net/MatMul-op2': [[1, 8], [8, 1]]
}
assert strategies == expected_strategies
def test_two_bn():
class Net(nn.Cell):
def __init__(self):
super().__init__()
self.block1 = getBlock()
self.block2 = getBlock()
self.relu = P.ReLU()
self.add = P.TensorAdd()
self.bias = Tensor(np.ones([64, 64]), dtype=ms.float32)
def construct(self, x):
out = self.block1(x)
out = self.relu(out)
out = self.add(out, self.bias)
out = self.block2(out)
return out
net = NetWithLoss(Net())
x = Tensor(np.ones([64, 64]), dtype=ms.float32)
context.set_context(save_graphs=True)
context.set_auto_parallel_context(device_num=8, global_rank=0)
context.set_auto_parallel_context(parallel_mode="auto_parallel")
net.set_auto_parallel()
set_algo_parameters(elementwise_op_strategy_follow=True)
reset_op_id()
_executor.compile(net, x, phase='train')
strategies = _executor._get_strategy(net)
assert len(strategies) == 4
for (k, v) in strategies.items():
if re.search('BatchNorm-op', k) is not None:
assert v == [[8, 1], [1], [1], [1], [1]]
elif re.search('TensorAdd-op', k) is not None:
assert v == [[8, 1], [8, 1]]
elif re.search('ReLU-op', k) is not None:
assert v == [[8, 1]]
def test_common_parameter():
class Net(nn.Cell):
def __init__(self):
super().__init__()
self.matmul1 = P.MatMul()
self.matmul2 = P.MatMul()
self.matmul3 = P.MatMul()
self.weight1 = Parameter(Tensor(np.ones([64, 64]).astype(np.float16) * 0.01), "w", requires_grad=True)
self.cast1 = P.Cast()
self.cast2 = P.Cast()
def construct(self, x, y):
m1_result = self.matmul1(x, self.cast1(self.weight1, mstype.float32))
m2_result = self.matmul2(y, self.cast2(self.weight1, mstype.float32))
m3_result = self.matmul3(m2_result, m1_result)
return m3_result
size = 8
context.set_auto_parallel_context(device_num=size, global_rank=0)
set_algo_parameters(elementwise_op_strategy_follow=True)
x = Tensor(np.ones([64, 64]), dtype=ms.float32)
y = Tensor(np.ones([64, 64]), dtype=ms.float32)
net = NetWithLoss(Net())
context.set_auto_parallel_context(parallel_mode="auto_parallel")
net.set_auto_parallel()
reset_op_id()
_executor.compile(net, x, y, phase='train')
strategies = _executor._get_strategy(net)
for (k, v) in strategies.items():
if re.search('MatMul-op', k) is not None:
assert v == [[8, 1], [1, 1]]
elif re.search('Cast-op', k) is not None:
assert v == [[1, 1]]
def test_two_matmul():
class Net(nn.Cell):
def __init__(self):
super().__init__()
self.matmul1 = P.MatMul()
self.matmul2 = P.MatMul()
def construct(self, x, y, b):
out = self.matmul1(x, y)
out = self.matmul2(out, b)
return out
size = 16
context.set_auto_parallel_context(device_num=size, global_rank=0)
cost_model_context.set_cost_model_context(
device_memory_capacity=32.0 * 1024.0 * 1024.0 * 1024.0,
costmodel_alpha=1.0,
costmodel_beta=60.0,
costmodel_gamma=0.1,
costmodel_communi_threshold=1024.0,
costmodel_communi_const=2222.0,
costmodel_communi_bias=1111.0)
dev_mem_cap = cost_model_context.get_cost_model_context(
"device_memory_capacity")
assert dev_mem_cap == 32.0 * 1024.0 * 1024.0 * 1024.0
costmodel_alpha = cost_model_context.get_cost_model_context(
"costmodel_alpha")
assert costmodel_alpha == 1.0
costmodel_beta = cost_model_context.get_cost_model_context(
"costmodel_beta")
assert costmodel_beta == 60.0
costmodel_gamma = cost_model_context.get_cost_model_context(
"costmodel_gamma")
assert costmodel_gamma == 0.1
costmodel_communi_threshold = cost_model_context.get_cost_model_context(
"costmodel_communi_threshold")
assert costmodel_communi_threshold == 1024.0
costmodel_communi_const = cost_model_context.get_cost_model_context(
"costmodel_communi_const")
assert costmodel_communi_const == 2222.0
costmodel_communi_bias = cost_model_context.get_cost_model_context(
"costmodel_communi_bias")
assert costmodel_communi_bias == 1111.0
cost_model_context.reset_cost_model_context()
dev_mem_cap = cost_model_context.get_cost_model_context(
"device_memory_capacity")
assert dev_mem_cap == 16.0 * 1024.0 * 1024.0 * 1024.0
costmodel_alpha = cost_model_context.get_cost_model_context(
"costmodel_alpha")
assert costmodel_alpha == 1.0
costmodel_beta = cost_model_context.get_cost_model_context(
"costmodel_beta")
assert costmodel_beta == 260.0
costmodel_gamma = cost_model_context.get_cost_model_context(
"costmodel_gamma")
assert costmodel_gamma == 0.001
costmodel_communi_threshold = cost_model_context.get_cost_model_context(
"costmodel_communi_threshold")
assert costmodel_communi_threshold == 2048.0
costmodel_communi_const = cost_model_context.get_cost_model_context(
"costmodel_communi_const")
assert costmodel_communi_const == 3072.0
costmodel_communi_bias = cost_model_context.get_cost_model_context(
"costmodel_communi_bias")
assert costmodel_communi_bias == 1024.0
set_algo_parameters(simplify_cal=True,
tensor_slice_align_enable=False,
tensor_slice_align_size=32,
fully_use_devices=False,
elementwise_op_strategy_follow=False)
para_simplify_cal = get_algo_parameters("simplify_cal")
assert para_simplify_cal == True
para_slice_align_enable = get_algo_parameters("tensor_slice_align_enable")
assert para_slice_align_enable == False
para_slice_align_size = get_algo_parameters("tensor_slice_align_size")
assert para_slice_align_size == 32
fully_use_devices = get_algo_parameters("fully_use_devices")
assert fully_use_devices == False
elementwise_op_strategy_follow = get_algo_parameters(
"elementwise_op_strategy_follow")
assert elementwise_op_strategy_follow == False
reset_algo_parameters()
para_simplify_cal = get_algo_parameters("simplify_cal")
assert para_simplify_cal == True
para_slice_align_enable = get_algo_parameters("tensor_slice_align_enable")
assert para_slice_align_enable == False
para_slice_align_size = get_algo_parameters("tensor_slice_align_size")
assert para_slice_align_size == 16
fully_use_devices = get_algo_parameters("fully_use_devices")
assert fully_use_devices == True
elementwise_op_strategy_follow = get_algo_parameters(
"elementwise_op_strategy_follow")
assert elementwise_op_strategy_follow == False
x = Tensor(np.ones([128, 32]), dtype=ms.float32)
y = Tensor(np.ones([32, 64]), dtype=ms.float32)
b = Tensor(np.ones([64, 64]), dtype=ms.float32)
net = NetWithLoss(Net())
context.set_auto_parallel_context(parallel_mode="auto_parallel")
reset_op_id()
_executor.compile(net, x, y, b, phase='train')
strategies = _executor._get_strategy(net)
expected_strategies = {
'Default/network-Net/MatMul-op0': [[16, 1], [1, 1]],
'Default/network-Net/MatMul-op1': [[16, 1], [1, 1]]
}
assert strategies == expected_strategies
def test_train_32k_8p(epoch_size=3, batch_size=32, num_classes=32768): #1048576 #131072 #32768 #8192
dev_num = 8
context.set_auto_parallel_context(parallel_mode=ParallelMode.AUTO_PARALLEL, device_num=dev_num)
cost_model_context.set_cost_model_context(costmodel_gamma=0.001, costmodel_beta=260.0)
cost_model_context.set_cost_model_context(costmodel_allreduce_fusion_algorithm=1)
cost_model_context.set_cost_model_context(costmodel_allreduce_fusion_times=2)
cost_model_context.set_cost_model_context(costmodel_allreduce_fusion_tail_percent=0.5)
set_algo_parameters(elementwise_op_strategy_follow=True)
resset_op_id()
np.random.seed(6)
input_np = np.ones([batch_size, 3, 224, 224]).astype(np.float32)
label_np = np.zeros([batch_size]).astype(np.int32)
for i in range(0, batch_size):
label_np[i] = i % num_classes
dataset = DatasetLenet(Tensor(input_np), Tensor(label_np), 1)
net = resnet50(num_classes)
loss = SoftmaxCrossEntropyExpand(sparse=True)
opt = Momentum(filter(lambda x: x.requires_grad, net.get_parameters()), 0.01, 0.9)
model = Model(net, loss_fn=loss, optimizer=opt)
model.train(5, dataset, dataset_sink_mode=False)
strategies = _executor._get_strategy(model._train_network)
for (k, v) in strategies.items():
if re.search('Conv2D-op', k) is not None:
assert v[0][0] == dev_num
elif re.search('MatMul-op', k) is not None:
assert v == [[dev_num, 1], [1, 1]]
elif re.search('ReduceSum-op', k) is not None:
assert v == [[dev_num, 1]]
allreduce_fusion_dict = _executor._get_allreduce_fusion(model._train_network)
print(allreduce_fusion_dict)
expect_dict = {'end_point.bias': 2,
'end_point.weight': 2,
'layer4.2.bn3.beta': 2,
'layer4.2.bn3.gamma': 2,
'layer4.2.conv3.weight': 2,
'layer4.2.bn2.beta': 2,
'layer4.2.bn2.gamma': 2,
'layer4.2.conv2.weight': 2,
'layer4.2.bn1.beta': 2,
'layer4.2.bn1.gamma': 2,
'layer4.2.conv1.weight': 2,
'layer4.1.bn3.beta': 2,
'layer4.1.bn3.gamma': 2,
'layer4.1.conv3.weight': 2,
'layer4.1.bn2.beta': 2,
'layer4.1.bn2.gamma': 2,
'layer4.1.conv2.weight': 2,
'layer4.1.bn1.beta': 2,
'layer4.1.bn1.gamma': 2,
'layer4.1.conv1.weight': 2,
'layer4.0.bn_down_sample.beta': 2,
'layer4.0.bn_down_sample.gamma': 2,
'layer4.0.conv_down_sample.weight': 2,
'layer4.0.bn3.beta': 2,
'layer4.0.bn3.gamma': 2,
'layer4.0.conv3.weight': 2,
'layer4.0.bn2.beta': 2,
'layer4.0.bn2.gamma': 2,
'layer4.0.conv2.weight': 2,
'layer4.0.bn1.beta': 2,
'layer4.0.bn1.gamma': 2,
'layer4.0.conv1.weight': 2,
'layer3.5.bn3.beta': 2,
'layer3.5.bn3.gamma': 2,
'layer3.5.conv3.weight': 2,
'layer3.5.bn2.beta': 2,
'layer3.5.bn2.gamma': 2,
'layer3.5.conv2.weight': 2,
'layer3.5.bn1.beta': 2,
'layer3.5.bn1.gamma': 2,
'layer3.5.conv1.weight': 2,
'layer3.4.bn3.beta': 2,
'layer3.4.bn3.gamma': 2,
'layer3.4.conv3.weight': 2,
'layer3.4.bn2.beta': 2,
'layer3.4.bn2.gamma': 2,
'layer3.4.conv2.weight': 2,
'layer3.4.bn1.beta': 2,
'layer3.4.bn1.gamma': 2,
'layer3.4.conv1.weight': 2,
'layer3.3.bn3.beta': 2,
'layer3.3.bn3.gamma': 2,
'layer3.3.conv3.weight': 2,
'layer3.3.bn2.beta': 2,
'layer3.3.bn2.gamma': 2,
'layer3.3.conv2.weight': 2,
'layer3.3.bn1.beta': 2,
'layer3.3.bn1.gamma': 2,
'layer3.3.conv1.weight': 2,
'layer3.2.bn3.beta': 2,
'layer3.2.bn3.gamma': 2,
'layer3.2.conv3.weight': 2,
'layer3.2.bn2.beta': 2,
'layer3.2.bn2.gamma': 2,
'layer3.2.conv2.weight': 2,
'layer3.2.bn1.beta': 2,
'layer3.2.bn1.gamma': 2,
'layer3.2.conv1.weight': 2,
'layer3.1.bn3.beta': 2,
'layer3.1.bn3.gamma': 2,
'layer3.1.conv3.weight': 2,
'layer3.1.bn2.beta': 2,
'layer3.1.bn2.gamma': 2,
'layer3.1.conv2.weight': 2,
'layer3.1.bn1.beta': 2,
'layer3.1.bn1.gamma': 2,
'layer3.1.conv1.weight': 2,
'layer3.0.bn_down_sample.beta': 1,
'layer3.0.bn_down_sample.gamma': 1,
'layer3.0.conv_down_sample.weight': 2,
'layer3.0.bn3.beta': 1,
'layer3.0.bn3.gamma': 1,
'layer3.0.conv3.weight': 2,
'layer3.0.bn2.beta': 2,
'layer3.0.bn2.gamma': 2,
'layer3.0.conv2.weight': 2,
'layer3.0.bn1.beta': 2,
'layer3.0.bn1.gamma': 2,
'layer3.0.conv1.weight': 2,
'layer2.3.bn3.beta': 2,
'layer2.3.bn3.gamma': 2,
'layer2.3.conv3.weight': 2,
'layer2.3.bn2.beta': 2,
'layer2.3.bn2.gamma': 2,
'layer2.3.conv2.weight': 2,
'layer2.3.bn1.beta': 2,
'layer2.3.bn1.gamma': 2,
'layer2.3.conv1.weight': 2,
'layer2.2.bn3.beta': 2,
'layer2.2.bn3.gamma': 2,
'layer2.2.conv3.weight': 2,
'layer2.2.bn2.beta': 2,
'layer2.2.bn2.gamma': 2,
'layer2.2.conv2.weight': 2,
'layer2.2.bn1.beta': 2,
'layer2.2.bn1.gamma': 2,
'layer2.2.conv1.weight': 2,
'layer2.1.bn3.beta': 1,
'layer2.1.bn3.gamma': 1,
'layer2.1.conv3.weight': 2,
'layer2.1.bn2.beta': 2,
'layer2.1.bn2.gamma': 2,
'layer2.1.conv2.weight': 2,
'layer2.1.bn1.beta': 2,
'layer2.1.bn1.gamma': 2,
'layer2.1.conv1.weight': 2,
'layer2.0.bn_down_sample.beta': 1,
'layer2.0.bn_down_sample.gamma': 1,
'layer2.0.conv_down_sample.weight': 2,
'layer2.0.bn3.beta': 1,
'layer2.0.bn3.gamma': 1,
'layer2.0.conv3.weight': 2,
'layer2.0.bn2.beta': 2,
'layer2.0.bn2.gamma': 2,
'layer2.0.conv2.weight': 2,
'layer2.0.bn1.beta': 2,
'layer2.0.bn1.gamma': 2,
'layer2.0.conv1.weight': 2,
'layer1.2.bn3.beta': 2,
'layer1.2.bn3.gamma': 2,
'layer1.2.conv3.weight': 2,
'layer1.2.bn2.beta': 2,
'layer1.2.bn2.gamma': 2,
'layer1.2.conv2.weight': 2,
'layer1.2.bn1.beta': 2,
'layer1.2.bn1.gamma': 2,
'layer1.2.conv1.weight': 2,
'layer1.1.bn3.beta': 1,
'layer1.1.bn3.gamma': 1,
'layer1.1.conv3.weight': 2,
'layer1.1.bn2.beta': 2,
'layer1.1.bn2.gamma': 2,
'layer1.1.conv2.weight': 2,
'layer1.1.bn1.beta': 2,
'layer1.1.bn1.gamma': 2,
'layer1.1.conv1.weight': 2,
'layer1.0.bn_down_sample.beta': 1,
'layer1.0.bn_down_sample.gamma': 1,
'layer1.0.conv_down_sample.weight': 2,
'layer1.0.bn3.beta': 1,
'layer1.0.bn3.gamma': 1,
'layer1.0.conv3.weight': 2,
'layer1.0.bn2.beta': 2,
'layer1.0.bn2.gamma': 2,
'layer1.0.conv2.weight': 2,
'layer1.0.bn1.beta': 2,
'layer1.0.bn1.gamma': 2,
'layer1.0.conv1.weight': 2,
'bn1.beta': 1,
'bn1.gamma': 1,
'conv1.weight': 2}
assert (allreduce_fusion_dict == expect_dict)
cost_model_context.reset_cost_model_context()