def __init__(self, in_channel, x):
super().__init__()
#self._save_graphs(save_graph_flag=True, save_graph_path=".")
self.biasadd = P.BiasAdd()
self.equal = P.Equal()
self.addn = P.AddN()
self.conv = Conv2d(in_channels=in_channel,
out_channels=in_channel,
kernel_size=1,
stride=1,
has_bias=False,
weight_init='ones',
pad_mode='same')
self.bn = BatchNorm2d(num_features=in_channel)
self.assignadd = P.AssignAdd()
self.assign = P.Assign()
self.relu = ReLU()
self.mean = P.ReduceMean(keep_dims=False)
self.bias = Parameter(Tensor(
np.random.randint(2, size=(3, )).astype((np.float32))),
name="bias")
self.bias2 = Parameter(Tensor(np.ones([3]).astype(np.float32)),
name="bias2")
self.parameterupdate = ParameterUpdate(self.bias)
self.value = Tensor(np.random.randn(*(3, )), ms.float32)
self.x = x
def __init__(self):
super().__init__()
self.parameter1 = Parameter(Tensor([199.0], ms.float32), name="parameter1")
self.assign = P.Assign()
self.assignadd = P.AssignAdd()
self.addn = P.AddN()
self.depend = P.Depend()
def __init__(self,
params,
learning_rate,
beta1=0.9,
beta2=0.999,
eps=1e-6,
weight_decay=0.0):
super(Lamb, self).__init__(learning_rate, params, weight_decay)
_check_param_value(beta1, beta2, eps, self.cls_name)
# turn them to scalar when me support scalar/tensor mix operations
self.beta1 = Tensor(np.array([beta1]).astype(np.float32))
self.beta2 = Tensor(np.array([beta2]).astype(np.float32))
self.eps = Tensor(np.array([eps]).astype(np.float32))
self.params = self.parameters
self.moments1 = self.params.clone(prefix="lamb_m", init='zeros')
self.moments2 = self.params.clone(prefix="lamb_v", init='zeros')
if not self.dynamic_lr:
self.global_step = Parameter(initializer(0, [1]),
name='global_step')
self.assignadd = P.AssignAdd()
self.hyper_map = C.HyperMap()
self.enable_graph_kernel = context.get_context("enable_graph_kernel") and \
context.get_context("device_target") == "Ascend"
def __init__(self, params, learning_rate, momentum, weight_decay=0.0, loss_scale=1.0,
decay_filter=lambda x: 'beta' not in x.name and 'gamma' not in x.name):
super(Momentum, self).__init__(learning_rate, params)
if isinstance(momentum, float) and momentum < 0.0:
raise ValueError("momentum should be at least 0.0, but got momentum {}".format(momentum))
if isinstance(learning_rate, Iterable) or \
(isinstance(learning_rate, Tensor) and learning_rate.dim() == 1):
self.dynamic_lr = True
self.gather = P.GatherV2()
self.assignadd = P.AssignAdd()
self.global_step = Parameter(initializer(0, [1], mstype.int32), name="global_step")
self.axis = 0
else:
self.dynamic_lr = False
self.gather = None
self.assignadd = None
self.global_step = None
self.axis = None
self.momentum = Parameter(momentum, name="momentum")
self.params = self.parameters
self.moments = self.params.clone(prefix="moments", init='zeros')
self.decay_tf = tuple(decay_filter(x) for x in self.parameters)
self.hyper_map = C.HyperMap()
self.opt = P.ApplyMomentum()
self.weight_decay = weight_decay * loss_scale
self.reciprocal_scale = 1.0 / loss_scale
self.one = Tensor(1, mstype.int32)
def __init__(self,
optimizer,
epsilon=1e-05,
hyperpara=0.001,
weight_decay=0.0,
use_clip=False,
decay_filter=lambda x: 'LayerNorm' not in x.name and 'bias'
not in x.name,
lars_filter=lambda x: 'LayerNorm' not in x.name and 'bias'
not in x.name,
loss_scale=1.0):
super(LARS, self).__init__(0.0,
[Parameter(Tensor(0.0), name="trivial")])
self.opt = optimizer
self.parameters = optimizer.parameters
self.learning_rate = optimizer.learning_rate
self.lars = P.LARSUpdate(epsilon, hyperpara, use_clip)
self.reciprocal_scale = 1.0 / loss_scale
self.weight_decay = weight_decay * loss_scale
self.cast = P.Cast()
self.decay_flag = tuple(decay_filter(x) for x in self.parameters)
self.lars_flag = tuple(lars_filter(x) for x in self.parameters)
self.hyper_map = C.HyperMap()
self.dynamic_lr = False
self.gather = None
self.global_step = None
self.axis = None
if isinstance(self.learning_rate.default_input, Iterable) or \
(isinstance(self.learning_rate.default_input, Tensor) and self.learning_rate.default_input.dim() == 1):
self.dynamic_lr = True
self.assignadd = P.AssignAdd()
self.gather = P.GatherV2()
self.global_step = Parameter(initializer(0, [1], mstype.int32),
name="lars_global_step")
self.axis = 0
def __init__(self):
super().__init__()
self.assign_sub = P.AssignAdd()
self.mul = P.Mul()
self.mul_weight = Parameter(Tensor(
np.full([128, 32], 0.5, dtype=np.float32)),
name="mul_weight")
self.assignsub_weight = Parameter(Tensor(
np.full([128, 32], 1.1, dtype=np.float32)),
name="assignsub_weight")
def __init__(self,
learning_rate,
parameters,
weight_decay=0.0,
loss_scale=1.0,
decay_filter=lambda x: 'beta' not in x.name and 'gamma' not in
x.name):
super(Optimizer, self).__init__()
if isinstance(learning_rate, float):
self.dynamic_lr = False
self.gather = None
self.assignadd = None
self.global_step = None
validator.check_number_range("learning rate", learning_rate, 0.0,
float("inf"), Rel.INC_LEFT)
else:
self.dynamic_lr = True
self.gather = P.GatherV2()
self.assignadd = P.AssignAdd()
self.global_step = Parameter(initializer(0, [1], mindspore.int32),
name='global_step')
if isinstance(learning_rate, Iterable):
learning_rate = Tensor(
np.array(list(learning_rate)).astype(np.float32))
elif isinstance(learning_rate, Tensor):
if learning_rate.dim() > 1:
raise ValueError(
"Learning rate should be a 0 or 1 dim `Tensor`,"
f"but got {learning_rate.dim()}.")
if learning_rate.dim() == 1 and learning_rate.size() < 2:
logger.warning(
"If want to use the dynamic learning rate, please make sure that the number "
"of elements in the list, tuple or tensor passed is greater than 1."
)
else:
raise TypeError(
"Learning rate should be float, Tensor or Iterable.")
if loss_scale <= 0.0:
raise ValueError(
"Loss scale should be greater than 0, but got {}".format(
loss_scale))
if weight_decay < 0.0:
raise ValueError(
"Weight decay should be equal or greater than 0, but got {}".
format(weight_decay))
self.learning_rate = Parameter(learning_rate, name="learning_rate")
self.parameters = ParameterTuple(parameters)
self.reciprocal_scale = 1.0 / loss_scale
self.weight_decay = weight_decay * loss_scale
self.decay_flags = tuple(decay_filter(x) for x in self.parameters)
if not self.parameters:
raise ValueError("optimizer got an empty parameter list.")
def __init__(self,
params,
learning_rate=0.1,
decay=0.9,
momentum=0.0,
epsilon=1e-10,
use_locking=False,
centered=False,
loss_scale=1.0,
weight_decay=0.0,
decay_filter=lambda x: 'beta' not in x.name and 'gamma' not in
x.name):
super(RMSProp, self).__init__(learning_rate, params)
if isinstance(momentum, float) and momentum < 0.0:
raise ValueError(
"momentum should be at least 0.0, but got momentum {}".format(
momentum))
if decay < 0.0:
raise ValueError(
"decay should be at least 0.0, but got dampening {}".format(
decay))
self.decay = decay
self.epsilon = epsilon
validator.check_type("use_locking", use_locking, [bool])
validator.check_type("centered", centered, [bool])
self.centered = centered
if centered:
self.opt = P.ApplyCenteredRMSProp(use_locking)
self.mg = self.parameters.clone(prefix="mean_grad", init='zeros')
else:
self.opt = P.ApplyRMSProp(use_locking)
self.dynamic_lr = False
if not isinstance(learning_rate, float):
self.dynamic_lr = True
self.gather = P.GatherV2()
self.assignadd = P.AssignAdd()
self.global_step = Parameter(initializer(0, [1], mstype.int32),
name="global_step")
self.axis = 0
self.one = Tensor(1, mstype.int32)
self.momentum = momentum
self.ms = self.parameters.clone(prefix="mean_square", init='zeros')
self.moment = self.parameters.clone(prefix="moment", init='zeros')
self.hyper_map = C.HyperMap()
self.decay = decay
self.decay_tf = tuple(decay_filter(x) for x in self.parameters)
self.reciprocal_scale = 1.0 / loss_scale
self.weight_decay = weight_decay * loss_scale
def __init__(self,
params,
learning_rate=0.1,
momentum=0.0,
dampening=0.0,
weight_decay=0.0,
nesterov=False,
loss_scale=1.0):
super(SGD, self).__init__(learning_rate, params)
if isinstance(momentum, float) and momentum < 0.0:
raise ValueError(
"momentum should be at least 0.0, but got momentum {}".format(
momentum))
if dampening < 0.0:
raise ValueError(
"dampening should be at least 0.0, but got dampening {}".
format(dampening))
self.dampening = dampening
if weight_decay < 0.0:
raise ValueError(
"weight_decay should be at least 0.0, but got weight_decay {}".
format(weight_decay))
self.weight_decay = weight_decay
validator.check_type("nesterov", nesterov, [bool])
self.nesterov = nesterov
self.opt = P.SGD(dampening, weight_decay, nesterov)
self.dynamic_lr = False
self.gather = None
self.global_step = None
self.axis = None
if not isinstance(learning_rate, float):
self.dynamic_lr = True
self.gather = P.GatherV2()
self.assignadd = P.AssignAdd()
self.global_step = Parameter(initializer(0, [1], mstype.int32),
name="global_step")
self.axis = 0
self.momentum = Parameter(momentum, name="momentum")
self.params = self.parameters
self.accum = self.params.clone(prefix="accum", init='zeros')
self.stat = self.params.clone(prefix="stat", init='ones')
self.hyper_map = C.HyperMap()
self.weight_decay = weight_decay * loss_scale
self.reciprocal_scale = 1.0 / loss_scale
def __init__(self,
params,
learning_rate=1e-3,
beta1=0.9,
beta2=0.999,
eps=1e-8,
use_locking=False,
use_nesterov=False,
weight_decay=0.0,
loss_scale=1.0,
decay_filter=lambda x: 'beta' not in x.name and 'gamma' not in
x.name):
super(Adam, self).__init__(learning_rate, params)
_check_param_value(beta1, beta2, eps, weight_decay)
validator.check_type("use_locking", use_locking, [bool])
validator.check_type("use_nesterov", use_nesterov, [bool])
validator.check_type("loss_scale", loss_scale, [float])
validator.check_number_range("loss_scale", loss_scale, 1.0,
float("inf"), Rel.INC_LEFT)
self.dynamic_lr = False
if isinstance(learning_rate, Iterable) or \
(isinstance(learning_rate, Tensor) and learning_rate.dim() == 1):
self.dynamic_lr = True
self.gather = P.GatherV2()
self.assignadd = P.AssignAdd()
self.global_step = Parameter(initializer(0, [1], mstype.int32),
name="global_step")
self.axis = 0
self.beta1 = Tensor(beta1, mstype.float32)
self.beta2 = Tensor(beta2, mstype.float32)
self.beta1_power = Parameter(initializer(1, [1], mstype.float32),
name="beta1_power")
self.beta2_power = Parameter(initializer(1, [1], mstype.float32),
name="beta2_power")
self.eps = eps
self.moment1 = self.parameters.clone(prefix="moment1", init='zeros')
self.moment2 = self.parameters.clone(prefix="moment2", init='zeros')
self.decay_tf = tuple(decay_filter(x) for x in self.parameters)
self.hyper_map = C.HyperMap()
self.opt = P.Adam(use_locking, use_nesterov)
self.weight_decay = weight_decay * loss_scale
self.reciprocal_scale = 1.0 / loss_scale
self.pow = P.Pow()
self.sqrt = P.Sqrt()
self.one = Tensor(np.array([1.0]).astype(np.float32))
self.realdiv = P.RealDiv()
def _init_group_params(self, parameters, learning_rate, weight_decay):
"""Init learning rate or weight decay in group params."""
origin_dynamic_lr = self.dynamic_lr
self._parse_group_params(parameters, learning_rate)
if self.dynamic_lr and not origin_dynamic_lr:
self.gather = P.GatherV2()
self.assignadd = P.AssignAdd()
self.global_step = Parameter(initializer(0, [1], mindspore.int32), name='global_step')
params_store = []
for group_param in parameters:
if 'order_params' in group_param.keys():
ordered_parameters = group_param['order_params']
continue
self.group_params += group_param['params']
if 'lr' in group_param.keys():
params_dynamic_lr = isinstance(group_param['lr'], (Iterable, Tensor))
if self.dynamic_lr and not params_dynamic_lr:
lr = Tensor(np.array([group_param['lr']] * self.dynamic_lr_length).astype(np.float32))
else:
lr = self._get_single_lr(group_param['lr'])
else:
if self.dynamic_lr and not origin_dynamic_lr:
lr = Tensor(np.array([self.scalar_lr] * self.dynamic_lr_length).astype(np.float32))
else:
lr = learning_rate
if 'weight_decay' in group_param.keys():
validator.check_float_legal_value('weight_decay', group_param['weight_decay'], None)
validator.check_number_range('weight_decay', group_param['weight_decay'], 0.0, float("inf"),
Rel.INC_LEFT, self.cls_name)
weight_decay_ = group_param['weight_decay'] * self.loss_scale
else:
weight_decay_ = weight_decay * self.loss_scale
for key in group_param.keys():
if key not in ('params', 'lr', 'weight_decay'):
logger.warning(f"The optimizer cannot parse '{key}' when setting parameter groups.")
for param in group_param['params']:
validator.check_value_type("parameter", param, [Parameter], self.cls_name)
if param.name in params_store:
raise RuntimeError(f"The {param.name} parameter has appeared in parameter groups.")
params_store.append(param.name)
self.group_lr.append(Parameter(lr, name="lr_" + param.name))
self.group_weight_decay.append(weight_decay_)
if self.is_group_params_ordered:
self._order_and_adjust_group_params(ordered_parameters, learning_rate, weight_decay)
def __init__(self):
super(AssignAdd, self).__init__()
self.add = P.AssignAdd()
def __init__(self):
super(Assign_RAW, self).__init__()
self.assign_add = P.AssignAdd()
self.greater = P.Greater()
self.add = P.Add()
self.para = Parameter(Tensor(1, dtype=ms.int32), name='para')
def _cumulative_gard(grad_sum, grad):
"""Apply gard sum to cumulative gradient."""
add = P.AssignAdd()
return add(grad_sum, grad)
def __init__(self,
in_channels,
out_channels,
weight_init='normal',
bias_init='zeros',
damping=0.03,
loss_scale=1,
frequency=278,
has_bias=True,
activation=None):
super(Dense_Thor, self).__init__()
self.in_channels = Validator.check_positive_int(in_channels)
self.out_channels = Validator.check_positive_int(out_channels)
self.has_bias = Validator.check_bool(has_bias)
self.thor = True
if isinstance(weight_init, Tensor):
if weight_init.ndim != 2 or weight_init.shape[0] != out_channels or \
weight_init.shape[1] != in_channels:
raise ValueError("weight_init shape error")
self.weight = Parameter(initializer(weight_init, [out_channels, in_channels]), name="weight")
if self.has_bias:
if isinstance(bias_init, Tensor):
if bias_init.ndim != 1 or bias_init.shape[0] != out_channels:
raise ValueError("bias_init shape error")
self.bias = Parameter(initializer(bias_init, [out_channels]), name="bias")
self.matmul = P.MatMul(transpose_b=True)
self.bias_add = P.BiasAdd()
self.activation = get_activation(activation)
self.activation_flag = self.activation is not None
self.matrix_A_inv = Parameter(Tensor(np.zeros([128, 128, 16, 16]).astype(np.float16)), name='matrix_A_inv',
requires_grad=False)
self.matrix_G_inv = Parameter(Tensor(np.zeros([63, 63, 16, 16]).astype(np.float16)), name="matrix_G_inv",
requires_grad=False)
self.fake_G = Tensor(np.zeros([63, 63, 16, 16]).astype(np.float16))
self.matmul = P.MatMul(transpose_b=True)
self.cube_matmul = P.CusMatMulCube(transpose_a=True)
self.matrix_combine = P.CusMatrixCombine()
self.cholesky = P.CusCholeskyTrsm()
self.shape = P.Shape()
self.reshape = P.Reshape()
self.transpose = P.Transpose()
self.cov_step = Parameter(initializer(0, [1], mstype.int32), name="cov_step", requires_grad=False)
self.mul = P.Mul()
self.cast = P.Cast()
self.damping = Tensor(damping)
self.loss_scale = Tensor(1 / loss_scale, mstype.float16)
self.vector_matmul = P.CusBatchMatMul()
self.pad = P.Pad(((0, 24), (0, 24)))
self.pad1 = P.Pad(((0, 8), (0, 8)))
self.slice = P.Slice()
self.gather = P.GatherV2()
self.assignadd = P.AssignAdd()
self.freq = Tensor(frequency, mstype.int32)
self.axis = 0
self.A_inv_max = Parameter(initializer(0, [1], mstype.float32), name="A_inv_max", requires_grad=False)
self.G_inv_max = Parameter(initializer(0, [1], mstype.float32), name="G_inv_max", requires_grad=False)
self.fused_abs_max1 = P.CusFusedAbsMax1([1000, 1000])
self.fused_abs_max2 = P.CusFusedAbsMax1()
self.log = P.Log()
self.exp = P.Exp()
self.dampingA = Tensor(np.identity(2048), mstype.float32)
self.dampingG = Tensor(np.identity(1024), mstype.float32)
self.add = P.TensorAdd()
self.sqrt = P.Sqrt()
self.getG = P.InsertGradientOf(self.save_gradient)
def __init__(self,
in_channels,
out_channels,
weight_init='normal',
bias_init='zeros',
damping=0.03,
loss_scale=1,
frequency=100,
has_bias=False,
activation=None,
batch_size=12):
super(Dense_Thor, self).__init__()
self.in_channels = Validator.check_positive_int(in_channels)
self.out_channels = Validator.check_positive_int(out_channels)
self.has_bias = Validator.check_bool(has_bias)
self.thor = True
if isinstance(weight_init, Tensor):
if weight_init.dim() != 2 or weight_init.shape()[0] != out_channels or \
weight_init.shape()[1] != in_channels:
raise ValueError("weight_init shape error")
self.weight = Parameter(initializer(weight_init,
[out_channels, in_channels]),
name="weight")
if self.has_bias:
if isinstance(bias_init, Tensor):
if bias_init.dim() != 1 or bias_init.shape(
)[0] != out_channels:
raise ValueError("bias_init shape error")
self.bias = Parameter(initializer(bias_init, [out_channels]),
name="bias")
self.matmul = P.MatMul(transpose_b=True)
self.bias_add = P.BiasAdd()
self.activation = get_activation(activation)
self.activation_flag = self.activation is not None
self.matrix_A_inv = Parameter(Tensor(
np.zeros([in_channels, in_channels]).astype(np.float16)),
name='matrix_A_inv',
requires_grad=False)
self.matrix_G_inv = Parameter(Tensor(
np.zeros([out_channels, out_channels]).astype(np.float16)),
name="matrix_G_inv",
requires_grad=False)
self.fake_G = Tensor(
np.zeros([out_channels, out_channels]).astype(np.float16))
self.matmul = P.MatMul(transpose_b=True)
self.cube_matmul = P.CusMatMulCube(transpose_a=True)
self.matrix_combine = P.CusMatrixCombine()
self.cholesky = P.CusCholeskyTrsm()
self.shape = P.Shape()
self.reshape = P.Reshape()
self.transpose = P.Transpose()
self.cov_step = Parameter(initializer(0, [1], mstype.int32),
name="cov_step",
requires_grad=False)
self.mul = P.Mul()
self.cast = P.Cast()
self.damping = damping
self.loss_scale = Tensor(1 / loss_scale, mstype.float16)
self.vector_matmul = P.CusBatchMatMul()
self.gather = P.GatherV2()
self.assignadd = P.AssignAdd()
self.freq = Tensor(frequency, mstype.int32)
self.axis = 0
self.abs = P.Abs()
self.reduce_max = P.ReduceMax(keep_dims=False)
self.log = P.Log()
self.exp = P.Exp()
self.dampingA = Tensor(np.identity(in_channels), mstype.float32)
self.dampingG = Tensor(np.identity(out_channels), mstype.float32)
self.sqrt = P.Sqrt()
self.getG = P.InsertGradientOf(self.save_gradient)
self.batch_size = batch_size
def __init__(self, value):
super(AssignAdd, self).__init__()
self.var = Parameter(value, name="var")
self.add = P.AssignAdd()
def __init__(self):
super(Net, self).__init__()
self.AssignAdd = P.AssignAdd()
self.inputdata = Parameter(initializer(1, [1], ms.int64),
name="global_step")
print("inputdata: ", self.inputdata)
def __init__(self,
in_channels,
out_channels,
kernel_size,
stride=1,
pad_mode='same',
padding=0,
dilation=1,
group=1,
eps=1e-5,
momentum=0.997,
weight_init=None,
beta_init=None,
gamma_init=None,
mean_init=None,
var_init=None,
quant_delay=0,
freeze_bn=100000,
fake=True,
num_bits=8,
per_channel=False,
symmetric=False,
narrow_range=False):
super(Conv2dBatchNormQuant, self).__init__()
self.in_channels = in_channels
self.out_channels = out_channels
self.pad_mode = pad_mode
self.padding = padding
self.dilation = twice(dilation)
self.stride = twice(stride)
self.group = group
self.fake = fake
self.freeze_bn = freeze_bn
self.momentum = momentum
self.quant_delay = quant_delay
if isinstance(kernel_size, int):
self.kernel_size = (kernel_size, kernel_size)
else:
self.kernel_size = kernel_size
if weight_init is None:
weight_init = initializer(
'normal',
[out_channels, in_channels // group, *self.kernel_size])
self.weight = Parameter(weight_init, name='weight')
if gamma_init is None:
gamma_init = initializer('ones', [out_channels])
self.gamma = Parameter(gamma_init, name='gamma')
if beta_init is None:
beta_init = initializer('zeros', [out_channels])
self.beta = Parameter(beta_init, name='beta')
if mean_init is None:
mean_init = initializer('zeros', [out_channels])
self.moving_mean = Parameter(mean_init,
name='moving_mean',
requires_grad=False)
if var_init is None:
var_init = initializer('ones', [out_channels])
self.moving_variance = Parameter(var_init,
name='moving_variance',
requires_grad=False)
self.step = Parameter(initializer('normal', [1], dtype=mstype.int32),
name='step',
requires_grad=False)
self.conv = P.Conv2D(out_channel=self.out_channels,
kernel_size=self.kernel_size,
mode=1,
pad_mode=self.pad_mode,
pad=self.padding,
stride=self.stride,
dilation=self.dilation,
group=self.group)
self.fake_quant_weight = FakeQuantWithMinMax(min_init=-6,
max_init=6,
ema=False,
num_bits=num_bits,
quant_delay=quant_delay,
per_channel=per_channel,
out_channels=out_channels,
symmetric=symmetric,
narrow_range=narrow_range)
self.batchnorm_fold_train = P.BatchNormFold(epsilon=eps,
momentum=momentum,
is_training=True,
freeze_bn=freeze_bn)
self.batchnorm_fold_infer = P.BatchNormFold(epsilon=eps,
momentum=momentum,
is_training=False,
freeze_bn=freeze_bn)
self.correct_mul = P.CorrectionMul()
self.relu = P.ReLU()
self.batchnorm_fold2 = P.BatchNormFold2(freeze_bn=freeze_bn)
self.batchnorm_fold2_infer = P.BatchNormFold2(freeze_bn=0)
self.one = Tensor(1, mstype.int32)
self.assignadd = P.AssignAdd()
def __init__(self, para):
super(AssignAddNet, self).__init__()
self.para = Parameter(para, name="para")
self.assign_add = P.AssignAdd()
def __init__(self,):
super(AssignAddNet, self).__init__()
self.op = P.AssignAdd()
self.inputdata = Parameter(Tensor(np.zeros([1]).astype(np.bool_), mstype.bool_), name="assign_add1")
def __init__(self,
learning_rate,
parameters,
weight_decay=0.0,
loss_scale=1.0):
super(Optimizer, self).__init__(auto_prefix=False)
if parameters and not isinstance(parameters, list):
parameters = list(parameters)
if not parameters:
raise ValueError("Optimizer got an empty parameter list.")
if not isinstance(parameters[0], (dict, Parameter)):
raise TypeError(
"Only a list of Parameter or dict can be supported.")
if isinstance(loss_scale, int):
loss_scale = float(loss_scale)
validator.check_value_type("loss_scale", loss_scale, [float], None)
validator.check_number_range("loss_scale", loss_scale, 0.0,
float("inf"), Rel.INC_NEITHER, None)
if isinstance(weight_decay, int):
weight_decay = float(weight_decay)
validator.check_value_type("weight_decay", weight_decay, [float], None)
validator.check_number_range("weight_decay", weight_decay, 0.0,
float("inf"), Rel.INC_LEFT, None)
self.is_group = False
self.is_group_lr = False
self.loss_scale = loss_scale
if isinstance(learning_rate, float):
self.dynamic_lr = False
self.gather = None
self.assignadd = None
self.global_step = None
self.scalar_lr = learning_rate
else:
self.dynamic_lr = True
self.gather = P.GatherV2()
self.assignadd = P.AssignAdd()
self.global_step = Parameter(initializer(0, [1], mindspore.int32),
name='global_step')
self.scalar_lr = None
learning_rate = self._get_single_lr(learning_rate)
if isinstance(parameters[0], dict):
self.is_group = True
self.params = []
self.group_lr = []
self.group_weight_decay = []
self._init_group_params(parameters, learning_rate, weight_decay)
if self.is_group_lr:
self.learning_rate = ParameterTuple(self.group_lr)
else:
self.learning_rate = Parameter(learning_rate, name="learning_rate")
if self.is_group:
self.parameters = ParameterTuple(self.params)
self.weight_decay = tuple(self.group_weight_decay)
decay_filter = lambda x: x > 0
self.decay_flags = tuple(
decay_filter(x) for x in self.weight_decay)
else:
self.parameters = ParameterTuple(parameters)
self.weight_decay = weight_decay * loss_scale
decay_filter = lambda x: 'beta' not in x.name and 'gamma' not in x.name
self.decay_flags = tuple(decay_filter(x) for x in self.parameters)
self.reciprocal_scale = 1.0 / loss_scale
self.exec_weight_decay = any(self.decay_flags)
self.param_length = len(self.parameters)
('StridedSlice_2_Error', {
'block': (lambda x: P.StridedSlice(end_mask="1"), {
'exception': TypeError
}),
'desc_inputs': [0]
}),
('StridedSlice_3_Error', {
'block': (lambda x: P.StridedSlice(ellipsis_mask=1.1), {
'exception': TypeError
}),
'desc_inputs': [0]
}),
('StridedSlice_4_Error', {
'block': (lambda x: P.StridedSlice(new_axis_mask="1.1"), {
'exception': TypeError
}),
'desc_inputs': [0]
}),
('AssignAdd_Error', {
'block': (P.AssignAdd(), {
'exception': IndexError
}),
'desc_inputs': [[1]]
}),
]
@mindspore_test(pipeline_for_verify_exception_for_case_by_case_config)
def test_check_exception():
return raise_set
def __init__(self,
learning_rate,
parameters,
weight_decay=0.0,
loss_scale=1.0):
super(Optimizer, self).__init__(auto_prefix=False)
if parameters is not None and not isinstance(parameters, list):
parameters = list(parameters)
if not parameters:
raise ValueError("Optimizer got an empty parameter list.")
if not isinstance(parameters[0], (dict, Parameter)):
raise TypeError(
"Only a list of Parameter or dict can be supported.")
if isinstance(loss_scale, int):
loss_scale = float(loss_scale)
validator.check_value_type("loss_scale", loss_scale, [float],
self.cls_name)
validator.check_positive_float(loss_scale, "loss_scale", self.cls_name)
self.loss_scale = loss_scale
weight_decay = self._preprocess_weight_decay(weight_decay)
self.grad_centralization = False
self._unique = True
self._target = context.get_context("device_target")
self.dynamic_lr = False
self.assignadd = None
self.global_step = None
self.is_group = False
self.is_group_lr = False
self.is_group_params_ordered = False
learning_rate = self._preprocess_single_lr(learning_rate)
if isinstance(parameters[0], dict):
self.is_group = True
self.group_params = []
self.group_lr = []
self.group_weight_decay = []
self.group_grad_centralization = []
self._init_group_params(parameters, learning_rate, weight_decay,
self.grad_centralization)
# The final value of dynamic_lr can be determined after the process of parse_single_lr and init_group_params
if self.dynamic_lr:
self.assignadd = P.AssignAdd()
self.global_step = Parameter(initializer(0, [1], mindspore.int32),
name='global_step')
if self.is_group_lr:
self.learning_rate = CellList(
self.group_lr) if self.dynamic_lr else ParameterTuple(
self.group_lr)
else:
self.learning_rate = self._build_single_lr(learning_rate,
'learning_rate')
if self.is_group:
self.parameters = ParameterTuple(self.group_params)
self.weight_decay = tuple(self.group_weight_decay)
self.weight_decay_tensor_tuple = tuple(
Tensor(x, mstype.float32) for x in self.group_weight_decay)
decay_filter = lambda x: x > 0
self.decay_flags = tuple(
decay_filter(x) for x in self.weight_decay)
self.exec_weight_decay = any(self.decay_flags)
self.grad_centralization_flags = tuple(
self.group_grad_centralization)
else:
self.parameters = ParameterTuple(parameters)
self.weight_decay = weight_decay * loss_scale
self.weight_decay_tensor = Tensor(self.weight_decay,
mstype.float32)
decay_filter = lambda x: 'beta' not in x.name and 'gamma' not in x.name
self.decay_flags = tuple(decay_filter(x) for x in self.parameters)
self.exec_weight_decay = self.weight_decay > 0
# when a parameter has been unique, there is no need do another unique in optimizer.
for param in self.parameters:
if param.unique:
self._unique = False
break
ps_filter = lambda x: x.is_param_ps
self.ps_parameters = tuple(ps_filter(x) for x in self.parameters)
cache_filter = lambda x: x.cache_enable
self.cache_enable = tuple(cache_filter(x) for x in self.parameters)
self.reciprocal_scale = Tensor(1.0 / loss_scale, mstype.float32)
self.need_scale = loss_scale != 1.0
self.global_step_increase_tensor = Tensor(1, mstype.int32)
self.param_length = len(self.parameters)
self.map_ = C.Map()
self._use_parallel_optimizer()
def __init__(self):
super().__init__()
self.op = P.AssignAdd()
self.inputdata = Parameter(initializer(1, [1], ms.float32),
name="global_step")
@non_graph_engine
@mindspore_test(pipeline_for_compile_forward_ge_graph_for_case_by_case_config)
def test_exec():
context.set_context(mode=context.GRAPH_MODE)
return test_exec_case
raise_set = [
('StridedSlice_1_Error', {
'block': (lambda x: P.StridedSlice(begin_mask="1"), {'exception': TypeError}),
'desc_inputs': [0]}),
('StridedSlice_2_Error', {
'block': (lambda x: P.StridedSlice(end_mask="1"), {'exception': TypeError}),
'desc_inputs': [0]}),
('StridedSlice_3_Error', {
'block': (lambda x: P.StridedSlice(ellipsis_mask=1.1), {'exception': TypeError}),
'desc_inputs': [0]}),
('StridedSlice_4_Error', {
'block': (lambda x: P.StridedSlice(new_axis_mask="1.1"), {'exception': TypeError}),
'desc_inputs': [0]}),
('AssignAdd_Error', {
'block': (P.AssignAdd(), {'exception': ValueError}),
'desc_inputs': [[1]]}),
]
@mindspore_test(pipeline_for_verify_exception_for_case_by_case_config)
def test_check_exception():
return raise_set
def __init__(self):
super(AssignAddNet, self).__init__()
self.AssignAdd = P.AssignAdd()
self.inputdata = Parameter(initializer(1, [1], ms.float16),
name="KIND_AUTOCAST_SCALAR_TO_TENSOR")
self.one = 1
def __init__(self, learning_rate, parameters, weight_decay=0.0, loss_scale=1.0):
super(Optimizer, self).__init__(auto_prefix=False)
if parameters is not None and not isinstance(parameters, list):
parameters = list(parameters)
if not parameters:
raise ValueError("Optimizer got an empty parameter list.")
if not isinstance(parameters[0], (dict, Parameter)):
raise TypeError("Only a list of Parameter or dict can be supported.")
if isinstance(loss_scale, int):
loss_scale = float(loss_scale)
validator.check_value_type("loss_scale", loss_scale, [float], self.cls_name)
validator.check_number_range("loss_scale", loss_scale, 0.0, float("inf"), Rel.INC_NEITHER, self.cls_name)
self.loss_scale = loss_scale
weight_decay = self._preprocess_weight_decay(weight_decay)
self.dynamic_lr = False
self.assignadd = None
self.global_step = None
self.is_group = False
self.is_group_lr = False
self.is_group_params_ordered = False
learning_rate = self._preprocess_single_lr(learning_rate)
if isinstance(parameters[0], dict):
self.is_group = True
self.group_params = []
self.group_lr = []
self.group_weight_decay = []
self._init_group_params(parameters, learning_rate, weight_decay)
# The final value of dynamic_lr can be determined after the process of parse_single_lr and init_group_params
if self.dynamic_lr:
self.assignadd = P.AssignAdd()
self.global_step = Parameter(initializer(0, [1], mindspore.int32), name='global_step')
if self.is_group_lr:
if self.dynamic_lr:
self.learning_rate = CellList(self.group_lr)
else:
self.learning_rate = ParameterTuple(self.group_lr)
else:
self.learning_rate = self._build_single_lr(learning_rate, 'learning_rate')
if self.is_group:
self.parameters = ParameterTuple(self.group_params)
self.weight_decay = tuple(self.group_weight_decay)
decay_filter = lambda x: x > 0
self.decay_flags = tuple(decay_filter(x) for x in self.weight_decay)
self.exec_weight_decay = any(self.decay_flags)
else:
self.parameters = ParameterTuple(parameters)
self.weight_decay = weight_decay * loss_scale
decay_filter = lambda x: 'beta' not in x.name and 'gamma' not in x.name
self.decay_flags = tuple(decay_filter(x) for x in self.parameters)
self.exec_weight_decay = self.weight_decay > 0
ps_filter = lambda x: x.is_param_ps
self.ps_parameters = tuple(ps_filter(x) for x in self.parameters)
self.reciprocal_scale = 1.0 / loss_scale
self.param_length = len(self.parameters)
self.map_ = C.Map()
use_parallel = context.get_auto_parallel_context("enable_parallel_optimizer")
self.use_parallel = use_parallel
if use_parallel:
if self.cls_name not in ["Lamb", "AdamWeightDecay"]:
raise RuntimeError("Optimizer segmentation does not support optimizer {}".format(self.cls_name))
if _get_parallel_mode() != ParallelMode.DATA_PARALLEL:
raise RuntimeError("Optimizer segmentation does not support parallel mode {}".format
(_get_parallel_mode()))
self.dev_num = _get_device_num()
if self.dev_num > self.param_length:
raise RuntimeError("Optimizer segmentation can not be applied when the number of parameters {} is"
" less than the number of devices {}".format(self.param_length, self.dev_num))
self.param_rank = self._get_parameter_group_id()
self.optim_filter = tuple(map(lambda x: x == _get_global_rank(), self.param_rank))
self.param_names = []
for param in self.parameters:
self.param_names.append(param.name)
else:
self.optim_filter = (True,) * self.param_length
def __init__(self,
learning_rate,
parameters,
weight_decay=0.0,
loss_scale=1.0):
super(Optimizer, self).__init__(auto_prefix=False)
if parameters is not None and not isinstance(parameters, list):
parameters = list(parameters)
if not parameters:
raise ValueError("Optimizer got an empty parameter list.")
if not isinstance(parameters[0], (dict, Parameter)):
raise TypeError(
"Only a list of Parameter or dict can be supported.")
if isinstance(loss_scale, int):
loss_scale = float(loss_scale)
validator.check_value_type("loss_scale", loss_scale, [float],
self.cls_name)
validator.check_positive_float(loss_scale, "loss_scale", self.cls_name)
self.loss_scale = loss_scale
weight_decay = self._preprocess_weight_decay(weight_decay)
self._unique = True
self._target = context.get_context("device_target")
self.dynamic_lr = False
self.assignadd = None
self.global_step = None
self.is_group = False
self.is_group_lr = False
self.is_group_params_ordered = False
learning_rate = self._preprocess_single_lr(learning_rate)
if isinstance(parameters[0], dict):
self.is_group = True
self.group_params = []
self.group_lr = []
self.group_weight_decay = []
self._init_group_params(parameters, learning_rate, weight_decay)
# The final value of dynamic_lr can be determined after the process of parse_single_lr and init_group_params
if self.dynamic_lr:
self.assignadd = P.AssignAdd()
self.global_step = Parameter(initializer(0, [1], mindspore.int32),
name='global_step')
if self.is_group_lr:
if self.dynamic_lr:
self.learning_rate = CellList(self.group_lr)
else:
self.learning_rate = ParameterTuple(self.group_lr)
else:
self.learning_rate = self._build_single_lr(learning_rate,
'learning_rate')
if self.is_group:
self.parameters = ParameterTuple(self.group_params)
self.weight_decay = tuple(self.group_weight_decay)
self.weight_decay_tensor_tuple = tuple(
Tensor(x, mstype.float32) for x in self.group_weight_decay)
decay_filter = lambda x: x > 0
self.decay_flags = tuple(
decay_filter(x) for x in self.weight_decay)
self.exec_weight_decay = any(self.decay_flags)
else:
self.parameters = ParameterTuple(parameters)
self.weight_decay = weight_decay * loss_scale
self.weight_decay_tensor = Tensor(self.weight_decay,
mstype.float32)
decay_filter = lambda x: 'beta' not in x.name and 'gamma' not in x.name
self.decay_flags = tuple(decay_filter(x) for x in self.parameters)
self.exec_weight_decay = self.weight_decay > 0
# when a parameter has been unique, there is no need do another unique in optimizer.
for param in self.parameters:
if param.unique:
self._unique = False
break
ps_filter = lambda x: x.is_param_ps
self.ps_parameters = tuple(ps_filter(x) for x in self.parameters)
ps_cache_filter = lambda x: x.cache_enable
self.cache_enable = tuple(ps_cache_filter(x) for x in self.parameters)
self.reciprocal_scale = Tensor(1.0 / loss_scale, mstype.float32)
self.need_scale = loss_scale != 1.0
self.global_step_increase_tensor = Tensor(1, mstype.int32)
self.param_length = len(self.parameters)
self.map_ = C.Map()
if context.get_auto_parallel_context("enable_parallel_optimizer"):
if _get_parallel_mode(
) == ParallelMode.DATA_PARALLEL and context.get_context(
"device_target") == "Ascend":
self.use_parallel = True
elif _get_parallel_mode() == ParallelMode.DATA_PARALLEL \
and context.get_context("device_target") != "Ascend":
raise RuntimeError(
"Parallel optimizer only supports Ascend in data parallel mode."
)
elif _get_parallel_mode() in (ParallelMode.STAND_ALONE,
ParallelMode.HYBRID_PARALLEL):
raise RuntimeError(
"Parallel optimizer is not supported in {}.".format(
_get_parallel_mode()))
else:
self.use_parallel = False
else:
self.use_parallel = False
if self.use_parallel:
if self.cls_name not in ["Lamb", "AdamWeightDecay"]:
raise RuntimeError(
"Parallel optimizer does not support optimizer {}".format(
self.cls_name))
self.dev_num = _get_device_num()
if self.dev_num > self.param_length:
raise RuntimeError(
"Parallel optimizer can not be applied when the number of parameters {} is"
" less than the number of devices {}".format(
self.param_length, self.dev_num))
self.param_rank = self._get_parameter_group_id()
self.optim_filter = tuple(
map(lambda x: x == _get_global_rank(), self.param_rank))
self.param_names = []
for param in self.parameters:
self.param_names.append(param.name)
else:
self.optim_filter = (True, ) * self.param_length
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ============================================================================
from mindspore.ops import Primitive
from mindspore.ops import operations as P
from mindspore.ops import _constants as Constants
depend = P.Depend()
all_reduce = P.AllReduce()
broadcast = P.Broadcast(1)
tensor_move = Primitive('TensorMove')
make_tuple = Primitive('MakeTuple')
tuple_getitem = Primitive(Constants.kTupleGetItem)
assign_add = P.AssignAdd()
apply_momentun = P.ApplyMomentum()
relu = P.ReLU()
class FnDict:
def __init__(self):
self.fnDict = {}
def __call__(self, fn):
self.fnDict[fn.__name__] = fn
def __getitem__(self, name):
return self.fnDict[name]
