!8695 add AdamOffload optimizer

From: @wangnan39
Reviewed-by: @stsuteng,@stsuteng,@kingxian
Signed-off-by: @kingxian

mindspore/ccsrc/backend/kernel_compiler/cpu/adam_delta_cpu_kernel.h

@@ -36,7 +36,7 @@ class AdamDeltaCPUKernel : public CPUKernel {
   size_t elem_num_{0};
 };
 
-MS_REG_CPU_KERNEL(AdamDelta,
+MS_REG_CPU_KERNEL(AdamNoUpdateParam,
                   KernelAttr()
                     .AddInputAttr(kNumberTypeFloat32)
                     .AddInputAttr(kNumberTypeFloat32)
@@ -47,7 +47,6 @@ MS_REG_CPU_KERNEL(AdamDelta,
                     .AddInputAttr(kNumberTypeFloat32)
                     .AddInputAttr(kNumberTypeFloat32)
                     .AddInputAttr(kNumberTypeFloat32)
-                    .AddInputAttr(kNumberTypeFloat32)
                     .AddOutputAttr(kNumberTypeFloat32),
                   AdamDeltaCPUKernel);
 }  // namespace kernel

mindspore/nn/optim/__init__.py

@@ -20,7 +20,7 @@ The optimizer is used to calculate and update the gradients.
 """
 from .optimizer import Optimizer
 from .momentum import Momentum
-from .adam import Adam, AdamWeightDecay
+from .adam import Adam, AdamWeightDecay, AdamOffload
 from .lamb import Lamb
 from .sgd import SGD
 from .lars import LARS
@@ -29,5 +29,5 @@ from .rmsprop import RMSProp
 from .proximal_ada_grad import ProximalAdagrad
 from .lazyadam import LazyAdam
 
-__all__ = ['Optimizer', 'Momentum', 'LARS', 'Adam', 'AdamWeightDecay', 'LazyAdam',
+__all__ = ['Optimizer', 'Momentum', 'LARS', 'Adam', 'AdamWeightDecay', 'LazyAdam', 'AdamOffload',
            'Lamb', 'SGD', 'FTRL', 'RMSProp', 'ProximalAdagrad']

mindspore/nn/optim/adam.py

@@ -90,22 +90,22 @@ def _update_run_op(beta1, beta2, eps, lr, weight_decay, param, m, v, gradient, d
 @_adam_opt.register("Function", "Function", "Function", "Function", "Bool", "Bool", "Bool", "Tensor", "Tensor",
                     "Tensor", "Tensor", "Tensor", "Tensor", "RowTensor", "Tensor", "Tensor", "Tensor", "Bool")
 def _run_opt_with_sparse(opt, sparse_opt, push, pull, use_locking, use_nesterov, target, beta1_power,
-                         beta2_power, beta1, beta2, eps, lr, gradient, params, m, v, ps_parameter):
+                         beta2_power, beta1, beta2, eps, lr, gradient, param, m, v, ps_parameter):
     """Apply sparse adam optimizer to the weight parameter when the gradient is sparse."""
     success = True
     indices = gradient.indices
     values = gradient.values
     if ps_parameter:
         op_shape = P.Shape()
-        shapes = (op_shape(params), op_shape(m), op_shape(v),
+        shapes = (op_shape(param), op_shape(m), op_shape(v),
                   op_shape(beta1_power), op_shape(beta2_power), op_shape(lr), op_shape(beta1),
                   op_shape(beta2), op_shape(eps), op_shape(values), op_shape(indices))
         success = F.depend(success, pull(push((beta1_power, beta2_power, lr, beta1, beta2,
-                                               eps, values, indices), shapes), params))
+                                               eps, values, indices), shapes), param))
         return success
 
     if not target:
-        success = F.depend(success, sparse_opt(params, m, v, beta1_power, beta2_power, lr, beta1, beta2,
+        success = F.depend(success, sparse_opt(param, m, v, beta1_power, beta2_power, lr, beta1, beta2,
                                                eps, values, indices))
     else:
         op_mul = P.Mul()
@@ -145,12 +145,12 @@ def _run_opt_with_sparse(opt, sparse_opt, push, pull, use_locking, use_nesterov,
 
         lr_t = lr * op_sqrt(1 - beta2_power) / (1 - beta1_power)
 
-        next_param = params - lr_t * param_update
+        next_param = param - lr_t * param_update
 
         F.control_depend(assign_m, next_m)
         F.control_depend(assign_v, next_v)
 
-        success = F.depend(success, F.assign(params, next_param))
+        success = F.depend(success, F.assign(param, next_param))
         success = F.depend(success, F.assign(m, next_m))
         success = F.depend(success, F.assign(v, next_v))
 
@@ -160,18 +160,29 @@ def _run_opt_with_sparse(opt, sparse_opt, push, pull, use_locking, use_nesterov,
 @_adam_opt.register("Function", "Function", "Function", "Function", "Bool", "Bool", "Bool", "Tensor", "Tensor",
                     "Tensor", "Tensor", "Tensor", "Tensor", "Tensor", "Tensor", "Tensor", "Tensor", "Bool")
 def _run_opt_with_one_number(opt, sparse_opt, push, pull, use_locking, use_nesterov, target, beta1_power,
-                             beta2_power, beta1, beta2, eps, lr, gradient, params, moment1, moment2, ps_parameter):
+                             beta2_power, beta1, beta2, eps, lr, gradient, param, moment1, moment2, ps_parameter):
     """Apply adam optimizer to the weight parameter using Tensor."""
     success = True
     if ps_parameter:
         op_shape = P.Shape()
         success = F.depend(success, pull(push((beta1_power, beta2_power, lr, beta1, beta2, eps, gradient),
-                                              (op_shape(params), op_shape(moment1), op_shape(moment2))), params))
+                                              (op_shape(param), op_shape(moment1), op_shape(moment2))), param))
     else:
-        success = F.depend(success, opt(params, moment1, moment2, beta1_power, beta2_power, lr, beta1, beta2,
+        success = F.depend(success, opt(param, moment1, moment2, beta1_power, beta2_power, lr, beta1, beta2,
                                         eps, gradient))
     return success
 
+
+@_adam_opt.register("Function", "Tensor", "Tensor", "Tensor", "Tensor", "Tensor", "Tensor", "Tensor", "Tensor",
+                    "Tensor", "Tensor")
+def _run_off_load_opt(opt, beta1_power, beta2_power, beta1, beta2, eps, lr, gradient, param, moment1, moment2):
+    """Apply AdamOffload optimizer to the weight parameter using Tensor."""
+    success = True
+    delat_param = opt(moment1, moment2, beta1_power, beta2_power, lr, beta1, beta2, eps, gradient)
+    success = F.depend(success, F.assign_add(param, delat_param))
+    return success
+
+
 def _check_param_value(beta1, beta2, eps, prim_name):
     """Check the type of inputs."""
     validator.check_value_type("beta1", beta1, [float], prim_name)
@@ -443,3 +454,146 @@ class AdamWeightDecay(Optimizer):
         if self.use_parallel:
             self.broadcast_params(optim_result)
         return optim_result
+
+
+class AdamOffload(Optimizer):
+    r"""
+    Updates gradients by the Adaptive Moment Estimation (Adam) algorithm. This optimizer will offload Adam optimizer to
+    host CPU and keep parameters being updated on the device, to minimize the memory cost. Although that would bring
+    about an increase of performance overhead, the optimizer could be used to run a larger model.
+
+    The Adam algorithm is proposed in `Adam: A Method for Stochastic Optimization <https://arxiv.org/abs/1412.6980>`_.
+
+    The updating formulas are as follows,
+
+    .. math::
+        \begin{array}{ll} \\
+            m = \beta_1 * m + (1 - \beta_1) * g \\
+            v = \beta_2 * v + (1 - \beta_2) * g * g \\
+            l = \alpha * \frac{\sqrt{1-\beta_2^t}}{1-\beta_1^t} \\
+            w = w - l * \frac{m}{\sqrt{v} + \epsilon}
+        \end{array}
+
+    :math:`m` represents the 1st moment vector `moment1`, :math:`v` represents the 2nd moment vector `moment2`,
+    :math:`g` represents `gradients`, :math:`l` represents scaling factor `lr`, :math:`\beta_1, \beta_2` represent
+    `beta1` and `beta2`, :math:`t` represents updating step while :math:`beta_1^t` and :math:`beta_2^t` represent
+    `beta1_power` and `beta2_power`, :math:`\alpha` represents `learning_rate`, :math:`w` represents `params`,
+    :math:`\epsilon` represents `eps`.
+
+    Note:
+        This optimizer only supports `GRAPH_MODE` currently.
+
+        When separating parameter groups, the weight decay in each group will be applied on the parameters if the
+        weight decay is positive. When not separating parameter groups, the `weight_decay` in the API will be applied
+        on the parameters without 'beta' or 'gamma' in their names if `weight_decay` is positive.
+
+        To improve parameter groups performance, the customized order of parameters is supported.
+
+    Args:
+        params (Union[list[Parameter], list[dict]]): When the `params` is a list of `Parameter` which will be updated,
+            the element in `params` must be class `Parameter`. When the `params` is a list of `dict`, the "params",
+            "lr", "weight_decay" and "order_params" are the keys can be parsed.
+
+            - params: Required. The value must be a list of `Parameter`.
+
+            - lr: Optional. If "lr" is in the keys, the value of the corresponding learning rate will be used.
+              If not, the `learning_rate` in the API will be used.
+
+            - weight_decay: Optional. If "weight_decay" is in the keys, the value of the corresponding weight decay
+              will be used. If not, the `weight_decay` in the API will be used.
+
+            - order_params: Optional. If "order_params" is in the keys, the value must be the order of parameters and
+              the order will be followed in the optimizer. There are no other keys in the `dict` and the parameters
+              which in the 'order_params' must be in one of group parameters.
+
+        learning_rate (Union[float, Tensor, Iterable, LearningRateSchedule]): A value or a graph for the learning rate.
+            When the learning_rate is an Iterable or a Tensor in a 1D dimension, use the dynamic learning rate, then
+            the i-th step will take the i-th value as the learning rate. When the learning_rate is LearningRateSchedule,
+            use dynamic learning rate, the i-th learning rate will be calculated during the process of training
+            according to the formula of LearningRateSchedule. When the learning_rate is a float or a Tensor in a zero
+            dimension, use fixed learning rate. Other cases are not supported. The float learning rate must be
+            equal to or greater than 0. If the type of `learning_rate` is int, it will be converted to float.
+            Default: 1e-3.
+        beta1 (float): The exponential decay rate for the 1st moment estimations. Should be in range (0.0, 1.0).
+                       Default: 0.9.
+        beta2 (float): The exponential decay rate for the 2nd moment estimations. Should be in range (0.0, 1.0).
+                       Default: 0.999.
+        eps (float): Term added to the denominator to improve numerical stability. Should be greater than 0. Default:
+                     1e-8.
+        use_locking (bool): Whether to enable a lock to protect variable tensors from being updated.
+            If true, updates of the var, m, and v tensors will be protected by a lock.
+            If false, the result is unpredictable. Default: False.
+        use_nesterov (bool): Whether to use Nesterov Accelerated Gradient (NAG) algorithm to update the gradients.
+            If true, update the gradients using NAG.
+            If false, update the gradients without using NAG. Default: False.
+        weight_decay (float): Weight decay (L2 penalty). It must be equal to or greater than 0. Default: 0.0.
+        loss_scale (float): A floating point value for the loss scale. Should be greater than 0. Default: 1.0.
+
+    Inputs:
+        - **gradients** (tuple[Tensor]) - The gradients of `params`, the shape is the same as `params`.
+
+    Outputs:
+        Tensor[bool], the value is True.
+
+    Examples:
+        >>> net = Net()
+        >>> #1) All parameters use the same learning rate and weight decay
+        >>> optim = nn.AdamOffload(params=net.trainable_params())
+        >>>
+        >>> #2) Use parameter groups and set different values
+        >>> conv_params = list(filter(lambda x: 'conv' in x.name, net.trainable_params()))
+        >>> no_conv_params = list(filter(lambda x: 'conv' not in x.name, net.trainable_params()))
+        >>> group_params = [{'params': conv_params, 'weight_decay': 0.01},
+        >>>                 {'params': no_conv_params, 'lr': 0.01},
+        >>>                 {'order_params': net.trainable_params()}]
+        >>> optim = nn.AdamOffload(group_params, learning_rate=0.1, weight_decay=0.0)
+        >>> # The conv_params's parameters will use default learning rate of 0.1 and weight decay of 0.01.
+        >>> # The no_conv_params's parameters will use learning rate of 0.01 and defaule weight decay of 0.0.
+        >>> # The final parameters order in which the optimizer will be followed is the value of 'order_params'.
+        >>>
+        >>> loss = nn.SoftmaxCrossEntropyWithLogits()
+        >>> model = Model(net, loss_fn=loss, optimizer=optim)
+    """
+
+    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):
+        super(AdamOffload, self).__init__(learning_rate, params, weight_decay, loss_scale)
+        _check_param_value(beta1, beta2, eps, self.cls_name)
+        validator.check_value_type("use_locking", use_locking, [bool], self.cls_name)
+        validator.check_value_type("use_nesterov", use_nesterov, [bool], self.cls_name)
+
+        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 = Tensor(eps, mstype.float32)
+        self.use_nesterov = use_nesterov
+        self.use_locking = use_locking
+        self.moment1 = self.parameters.clone(prefix="moment1", init='zeros')
+        self.moment2 = self.parameters.clone(prefix="moment2", init='zeros')
+
+        self.hyper_map = C.HyperMap()
+        self.opt = P.AdamNoUpdateParam(use_locking, use_nesterov)
+        self.opt.add_prim_attr("primitive_target", "CPU")
+
+    def construct(self, gradients):
+        params = self.parameters
+        moment1 = self.moment1
+        moment2 = self.moment2
+        gradients = self.decay_weight(gradients)
+        gradients = self.scale_grad(gradients)
+        lr = self.get_lr()
+
+        beta1_power = self.beta1_power * self.beta1
+        self.beta1_power = beta1_power
+        beta2_power = self.beta2_power * self.beta2
+        self.beta2_power = beta2_power
+        if self.is_group_lr:
+            success = self.map_(F.partial(_adam_opt, self.opt,
+                                          beta1_power, beta2_power, self.beta1, self.beta2, self.eps),
+                                lr, gradients, params, moment1, moment2)
+        else:
+            success = self.map_(F.partial(_adam_opt, self.opt,
+                                          beta1_power, beta2_power, self.beta1, self.beta2, self.eps, lr),
+                                gradients, params, moment1, moment2)
+        return success

mindspore/ops/operations/__init__.py

@@ -58,7 +58,7 @@ from .math_ops import (Abs, ACos, Asin, Asinh, AddN, AccumulateNV2, AssignAdd, A
 
 from .random_ops import (RandomChoiceWithMask, StandardNormal, Gamma, Poisson, UniformInt, UniformReal,
                          RandomCategorical, StandardLaplace, Multinomial)
-from .nn_ops import (LSTM, SGD, Adam, FusedSparseAdam, FusedSparseLazyAdam, ApplyMomentum, BatchNorm,
+from .nn_ops import (LSTM, SGD, Adam, FusedSparseAdam, FusedSparseLazyAdam, AdamNoUpdateParam, ApplyMomentum, BatchNorm,
                      BiasAdd, Conv2D,
                      DepthwiseConv2dNative,
                      DropoutDoMask, Dropout,
@@ -133,6 +133,7 @@ __all__ = [
     'Adam',
     'FusedSparseAdam',
     'FusedSparseLazyAdam',
+    'AdamNoUpdateParam',
     'Softplus',
     'Softmax',
     'Softsign',

mindspore/ops/operations/nn_ops.py

@@ -3315,6 +3315,107 @@ class Adam(PrimitiveWithInfer):
         return var_dtype, m_dtype, v_dtype
 
 
+class AdamNoUpdateParam(PrimitiveWithInfer):
+    r"""
+    Updates gradients by Adaptive Moment Estimation (Adam) algorithm. This operator do not update the parameter, but
+    calculate the value that should be added to the parameter instead.
+
+    The Adam algorithm is proposed in `Adam: A Method for Stochastic Optimization <https://arxiv.org/abs/1412.6980>`_.
+
+    The updating formulas are as follows,
+
+    .. math::
+        \begin{array}{ll} \\
+            m = \beta_1 * m + (1 - \beta_1) * g \\
+            v = \beta_2 * v + (1 - \beta_2) * g * g \\
+            l = \alpha * \frac{\sqrt{1-\beta_2^t}}{1-\beta_1^t} \\
+            \Delta{w} = - l * \frac{m}{\sqrt{v} + \epsilon}
+        \end{array}
+
+    :math:`m` represents the 1st moment vector, :math:`v` represents the 2nd moment vector, :math:`g` represents
+    `gradient`, :math:`l` represents scaling factor `lr`, :math:`\beta_1, \beta_2` represent `beta1` and `beta2`,
+    :math:`t` represents updating step while :math:`beta_1^t` and :math:`beta_2^t` represent `beta1_power` and
+    `beta2_power`, :math:`\alpha` represents `learning_rate`, :math:`w` represents the parameter to be updated,
+    :math:`\epsilon`represents `epsilon`.
+
+    Args:
+        use_locking (bool): Whether to enable a lock to protect variable tensors from being updated.
+            If true, updates of the var, m, and v tensors will be protected by a lock.
+            If false, the result is unpredictable. Default: False.
+        use_nesterov (bool): Whether to use Nesterov Accelerated Gradient (NAG) algorithm to update the gradients.
+            If true, update the gradients using NAG.
+            If false, update the gradients without using NAG. Default: False.
+
+    Inputs:
+        - **m** (Tensor) - The 1st moment vector in the updating formula. The data type must be float32.
+        - **v** (Tensor) - the 2nd moment vector in the updating formula. The shape must be the same as `m`.
+          The data type must be float32.
+        - **beta1_power** (Tensor) - :math:`beta_1^t` in the updating formula. The data type must be float32.
+        - **beta2_power** (Tensor) - :math:`beta_2^t` in the updating formula. The data type must be float32.
+        - **lr** (Tensor) - :math:`l` in the updating formula. The data type must be float32.
+        - **beta1** (Tensor) - The exponential decay rate for the 1st moment estimations. The data type must be float32.
+        - **beta2** (Tensor) - The exponential decay rate for the 2nd moment estimations. The data type must be float32.
+        - **epsilon** (Tensor) - Term added to the denominator to improve numerical stability. The data type must be
+          float32.
+        - **gradient** (Tensor) - Gradient, the shape must be the same as `m`, the data type must be float32.
+
+    Outputs:
+        Tensor, whose shape and data type are the same with `gradient`, is a value that should be added to the
+        parameter to be updated.
+
+    Examples:
+        >>> import numpy as np
+        >>> import mindspore as ms
+        >>> import mindspore.nn as nn
+        >>> from mindspore import Tensor, Parameter
+        >>> from mindspore.ops import operations as P
+        >>>
+        >>> class Net(nn.Cell):
+        >>>     def __init__(self):
+        >>>         super(Net, self).__init__()
+        >>>         self.adam = P.AdamNoUpdateParam()
+        >>>         self.m = Parameter(Tensor(np.array([[0.1, 0.1, 0.1], [0.2, 0.2, 0.2]]).astype(np.float32)),
+        >>>                            name="m")
+        >>>         self.v = Parameter(Tensor(np.array([[0.1, 0.1, 0.1], [0.2, 0.2, 0.2]]).astype(np.float32)),
+        >>>                            name="v")
+        >>>     def construct(self, beta1_power, beta2_power, lr, beta1, beta2, epsilon, grad):
+        >>>         out = self.adam(self.m, self.v, beta1_power, beta2_power, lr, beta1, beta2, epsilon, grad)
+        >>>         return out
+        >>> net = Net()
+        >>> beta1_power = Tensor(0.9, ms.float32)
+        >>> beta2_power = Tensor(0.999, ms.float32)
+        >>> lr = Tensor(0.001, ms.float32)
+        >>> beta1 = Tensor(0.9, ms.float32)
+        >>> beta2 = Tensor(0.999, ms.float32)
+        >>> epsilon = Tensor(1e-8, ms.float32)
+        >>> gradient = Tensor(np.array([[0.1, 0.1, 0.1], [0.1, 0.1, 0.1]]).astype(np.float32))
+        >>>
+        >>> result = net(beta1_power, beta2_power, lr, beta1, beta2, epsilon, gradient)
+        >>> print(result)
+        [[-0.00010004 -0.00010004 -0.00010004]
+        [-0.00013441 -0.00013441 -0.00013441]]
+
+    """
+    @prim_attr_register
+    def __init__(self, use_locking=False, use_nesterov=False):
+        validator.check_value_type("use_locking", use_locking, [bool], self.name)
+        validator.check_value_type("use_nesterov", use_nesterov, [bool], self.name)
+
+    def infer_shape(self, m_shape, v_shape, beta1_power_shape, beta2_power_shape, lr_shape,
+                    beta1_shape, beta2_shape, epsilon_shape, grad_shape):
+        validator.check("grad_shape", grad_shape, "m_shape", m_shape, Rel.EQ, self.name)
+        validator.check("grad_shape", grad_shape, "v_shape", v_shape, Rel.EQ, self.name)
+        return grad_shape
+
+    def infer_dtype(self, m_dtype, v_dtype, beta1_power_dtype, beta2_power_dtype, lr_dtype,
+                    beta1_dtype, beta2_dtype, epsilon_dtype, grad_dtype):
+        args = {"m": m_dtype, "v": v_dtype, "grad": grad_dtype,
+                "beta1_power": beta1_power_dtype, "beta2_power": beta2_power_dtype, 'lr': lr_dtype,
+                "beta1": beta1_dtype, "beta2": beta2_dtype, "epsilon": epsilon_dtype}
+        validator.check_tensors_dtypes_same_and_valid(args, [mstype.float32], self.name)
+        return grad_dtype
+
+
 class FusedSparseAdam(PrimitiveWithInfer):
     r"""
     Merges the duplicate value of the gradient and then updates parameters by Adaptive Moment Estimation (Adam)

tests/ut/python/nn/optim/test_adam.py

@@ -183,6 +183,24 @@ def test_adamweightdecay_group():
     _executor.compile(train_network, inputs, label)
 
 
+def test_adamoffload_group():
+    """ test_adam_group_lr_and_weight_decay """
+    inputs = Tensor(np.ones([1, 64]).astype(np.float32))
+    label = Tensor(np.zeros([1, 10]).astype(np.float32))
+    net = Net()
+    net.set_train()
+    loss = nn.SoftmaxCrossEntropyWithLogits()
+    net_with_loss = WithLossCell(net, loss)
+    all_params = net.trainable_params()
+
+    schedule_lr = lr_schedules.PolynomialDecayLR(0.01, 0.0001, 3, power=1.0)
+    group_params = [{'params': [all_params[0]], 'lr': 0.02, 'weight_decay': 0.9},
+                    {'params': [all_params[1]]}]
+    optimizer = nn.AdamOffload(group_params, learning_rate=schedule_lr)
+    train_network = TrainOneStepCell(net_with_loss, optimizer)
+    _executor.compile(train_network, inputs, label)
+
+
 def test_AdamWeightDecay_beta1():
     net = Net()
     print("**********", net.get_parameters())