mindspore/mindspore/nn/optim/lamb.py

# Copyright 2020 Huawei Technologies Co., Ltd
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# 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.
# ============================================================================
"""lamb"""
import numpy as np
from mindspore import context
from mindspore.common import dtype as mstype
from mindspore.common.initializer import initializer
from mindspore.ops import operations as P
from mindspore.ops import composite as C
from mindspore.ops import functional as F
from mindspore.common.parameter import Parameter
from mindspore.common.tensor import Tensor
from mindspore._checkparam import Validator as validator
from mindspore._checkparam import Rel
from .optimizer import Optimizer
from .. import layer
from .. import _graph_kernels as G

num_one = Tensor(np.ones([1]), mstype.float32)

_lamb_opt = C.MultitypeFuncGraph("lamb_opt")

@_lamb_opt.register("Tensor", "Tensor", "Tensor", "Tensor", "Tensor", "Number", "Tensor", "Tensor", "Tensor",
                    "Tensor", "Bool", "Bool")
def _update_run_op(beta1, beta2, eps, global_step, lr, weight_decay, param, m, v, gradient, decay_flag, optim_filter):
    """
    Update parameters.

    Args:
        beta1 (Tensor): The exponential decay rate for the 1st moment estimations. Should be in range (0.0, 1.0).
        beta2 (Tensor): The exponential decay rate for the 2nd moment estimations. Should be in range (0.0, 1.0).
        eps (Tensor): Term added to the denominator to improve numerical stability. Should be greater than 0.
        lr (Tensor): Learning rate.
        weight_decay (Number): Weight decay. Should be equal to or greater than 0.
        global_step (Tensor): Global step.
        param (Tensor): Parameters.
        m (Tensor): m value of parameters.
        v (Tensor): v value of parameters.
        gradient (Tensor): Gradient of parameters.
        decay_flag (bool): Specifies whether param update with weight decay.
        optim_filter(bool): Applies parameter update or not.

    Returns:
        Tensor, the new value of v after updating.
    """
    if optim_filter:
        op_mul = P.Mul()
        op_sqrt = P.Sqrt()
        op_rsqrt = P.Rsqrt()
        op_square = P.Square()
        op_cast = P.Cast()
        op_reshape = P.Reshape()
        op_shape = P.Shape()
        op_pow = P.Pow()
        op_norm = layer.Norm()
        op_select = P.Select()
        op_greater = P.Greater()
        op_fill = P.Fill()
        op_dtype = P.DType()

        param_fp32 = op_cast(param, mstype.float32)
        m_fp32 = op_cast(m, mstype.float32)
        v_fp32 = op_cast(v, mstype.float32)
        gradient_fp32 = op_cast(gradient, mstype.float32)

        next_m = op_mul(beta1, m_fp32) + op_mul(op_cast(num_one, mstype.float32) - beta1, gradient_fp32)

        next_v = op_mul(beta2, v_fp32) + op_mul(op_cast(num_one, mstype.float32) - beta2, op_square(gradient_fp32))

        next_mm = next_m / (op_cast(num_one, mstype.float32)
                            - op_pow(beta1, op_cast(global_step + num_one, mstype.float32)))
        next_vv = next_v / (op_cast(num_one, mstype.float32) -
                            op_pow(beta2, op_cast(global_step + num_one, mstype.float32)))
        w_norm = op_norm(param_fp32)
        g_norm = op_norm(gradient_fp32)

        g_norm_hat = op_norm(op_mul(next_mm, op_rsqrt(next_vv + eps)) + weight_decay * param_fp32)
        zeros = F.zeros_like(w_norm)
        ones = op_fill(op_dtype(w_norm), op_shape(w_norm), 1.0)
        trust_ratio = op_select(
            op_greater(w_norm, zeros),
            op_select(op_greater(g_norm, zeros), w_norm / g_norm_hat, ones),
            ones)
        tens = op_fill(op_dtype(trust_ratio), op_shape(trust_ratio), 10.0)
        trust_ratio = C.clip_by_value(trust_ratio, zeros, tens)
        update = next_mm / (op_sqrt(next_vv) + eps)

        if decay_flag:
            update = update + op_mul(weight_decay, param_fp32)

        update_with_lr = op_mul(op_mul(trust_ratio, lr), update)

        next_param = param_fp32 - op_reshape(update_with_lr, op_shape(param_fp32))

        next_param = F.depend(next_param, F.assign(param, op_cast(next_param, F.dtype(param))))
        next_param = F.depend(next_param, F.assign(m, op_cast(next_m, F.dtype(m))))
        next_param = F.depend(next_param, F.assign(v, op_cast(next_v, F.dtype(v))))

        return op_cast(next_param, F.dtype(param))
    return gradient


lamb_opt_graph_kernel = C.MultitypeFuncGraph("lamb_opt_graph_kernel")


@lamb_opt_graph_kernel.register("Tensor", "Tensor", "Tensor", "Tensor", "Tensor", "Number",
                                "Tensor", "Tensor", "Tensor", "Tensor", "Bool")
def _update_run_op_graph_kernel(beta1, beta2, eps, global_step, lr, weight_decay, param, m, v, gradient, decay_flag):
    """
    Update parameters.

    Args:
        beta1 (Tensor): The exponential decay rate for the 1st moment estimations. Should be in range (0.0, 1.0).
        beta2 (Tensor): The exponential decay rate for the 2nd moment estimations. Should be in range (0.0, 1.0).
        eps (Tensor): Term added to the denominator to improve numerical stability. Should be greater than 0.
        lr (Tensor): Learning rate.
        weight_decay (Number): Weight decay. Should be equal to or greater than 0.
        global_step (Tensor): Global step.
        param (Tensor): Parameters.
        m (Tensor): m value of parameters.
        v (Tensor): v value of parameters.
        gradient (Tensor): Gradient of parameters.
        decay_flag (bool): Specifies whether param update with weight decay.

    Returns:
        Tensor, the new value of v after updating.
    """
    op_mul = P.Mul()
    op_square = P.Square()
    op_cast = P.Cast()
    op_shape = P.Shape()
    op_pow = P.Pow()
    op_norm = layer.Norm()
    op_fill = P.Fill()
    op_dtype = P.DType()

    param_fp32 = op_cast(param, mstype.float32)
    gradient_fp32 = op_cast(gradient, mstype.float32)

    i6_ex = op_cast(global_step + num_one, mstype.float32)
    i9 = op_cast(num_one, mstype.float32) - beta1
    x1 = op_cast(num_one, mstype.float32) - beta2
    i6 = op_cast(num_one, mstype.float32) - op_pow(beta1, i6_ex)
    i3 = op_cast(num_one, mstype.float32) - op_pow(beta2, i6_ex)
    i1 = op_square(gradient_fp32)
    add3, update = G.LambNextMV()(i1, v, i3, gradient, m, i6, param, beta1, i9, beta2, x1, weight_decay, eps)

    if decay_flag:
        update = update + op_mul(weight_decay, param_fp32)

    w_norm = op_norm(param_fp32)
    g_norm = op_norm(gradient_fp32)
    g_norm_hat = op_norm(add3)

    zeros = F.zeros_like(w_norm)
    ones = op_fill(op_dtype(w_norm), op_shape(w_norm), 1.0)
    tens = op_fill(op_dtype(w_norm), op_shape(w_norm), 10.0)

    next_param = G.LambUpdateWithLR()(g_norm, w_norm, g_norm_hat, lr, update, param, zeros, ones, tens)
    next_v = F.control_depend(add3, next_param)
    return next_v


def _check_param_value(beta1, beta2, eps, prim_name):
    validator.check_value_type("beta1", beta1, [float], prim_name)
    validator.check_value_type("beta2", beta2, [float], prim_name)
    validator.check_value_type("eps", eps, [float], prim_name)
    validator.check_float_range(beta1, 0.0, 1.0, Rel.INC_NEITHER, "beta1", prim_name)
    validator.check_float_range(beta2, 0.0, 1.0, Rel.INC_NEITHER, "beta2", prim_name)
    validator.check_positive_float(eps, "eps", prim_name)


class Lamb(Optimizer):
    """
    Lamb Dynamic Learning Rate.

    LAMB is an optimization algorithm employing a layerwise adaptive large batch
    optimization technique. Refer to the paper `LARGE BATCH OPTIMIZATION FOR DEEP LEARNING: TRAINING BERT IN 76
    MINUTES <https://arxiv.org/abs/1904.00962>`_.

    Note:
        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 can be 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" in the keys, the value of corresponding learning rate will be used.
              If not, the `learning_rate` in the API will be used.

            - weight_decay: Optional. If "weight_decay" in the keys, the value of corresponding weight decay
              will be used. If not, the `weight_decay` in the API will be used.

            - order_params: Optional. If "order_params" in the keys, the value must be the order of parameters and
              the order will be followed in optimizer. There are no other keys in the `dict` and the parameters which
              in the value of '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 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.
        beta1 (float): The exponential decay rate for the 1st moment estimations. Default: 0.9.
            Should be in range (0.0, 1.0).
        beta2 (float): The exponential decay rate for the 2nd moment estimations. Default: 0.999.
            Should be in range (0.0, 1.0).
        eps (float): Term added to the denominator to improve numerical stability. Default: 1e-6.
            Should be greater than 0.
        weight_decay (float): Weight decay (L2 penalty). Default: 0.0. Should be equal to or greater than 0.

    Inputs:
        - **gradients** (tuple[Tensor]) - The gradients of `params`, the shape is the same as `params`.

    Outputs:
        tuple[bool], all elements are True.

    Supported Platforms:
        ``Ascend`` ``GPU``

    Examples:
        >>> net = Net()
        >>> #1) All parameters use the same learning rate and weight decay
        >>> optim = nn.Lamb(params=net.trainable_params(), learning_rate=0.1)
        >>>
        >>> #2) Use parameter groups and set different values
        >>> poly_decay_lr = learning_rate_schedule.PolynomialDecayLR(learning_rate=0.1, end_learning_rate=0.01,
        ...                                                    decay_steps=4, power = 0.5)
        >>> 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': poly_decay_lr},
        ...                 {'order_params': net.trainable_params(0.01)}]
        >>> optim = nn.Lamb(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 dynamic learning rate of poly decay learning rate and default
        >>> # 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, 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("device_target") == "Ascend" and \
            context.get_context("enable_graph_kernel")

    def construct(self, gradients):
        lr = self.get_lr()
        if self.enable_graph_kernel:
            if self.is_group:
                if self.is_group_lr:
                    optim_result = self.hyper_map(F.partial(lamb_opt_graph_kernel, self.beta1, self.beta2, self.eps,
                                                            self.global_step),
                                                  lr, self.weight_decay, self.params, self.moments1, self.moments2,
                                                  gradients, self.decay_flags)
                else:
                    optim_result = self.hyper_map(F.partial(lamb_opt_graph_kernel, self.beta1, self.beta2, self.eps,
                                                            self.global_step, lr),
                                                  self.weight_decay, self.params, self.moments1, self.moments2,
                                                  gradients, self.decay_flags)
            else:
                optim_result = self.hyper_map(F.partial(lamb_opt_graph_kernel, self.beta1, self.beta2, self.eps,
                                                        self.global_step, lr, self.weight_decay),
                                              self.params, self.moments1, self.moments2, gradients, self.decay_flags)
        else:
            if self.is_group:
                if self.is_group_lr:
                    optim_result = self.hyper_map(F.partial(_lamb_opt, self.beta1, self.beta2, self.eps,
                                                            self.global_step),
                                                  lr, self.weight_decay, self.params, self.moments1, self.moments2,
                                                  gradients, self.decay_flags, self.optim_filter)
                else:
                    optim_result = self.hyper_map(F.partial(_lamb_opt, self.beta1, self.beta2, self.eps,
                                                            self.global_step, lr),
                                                  self.weight_decay, self.params, self.moments1, self.moments2,
                                                  gradients, self.decay_flags, self.optim_filter)
            else:
                optim_result = self.hyper_map(F.partial(_lamb_opt, self.beta1, self.beta2, self.eps,
                                                        self.global_step, lr, self.weight_decay),
                                              self.params, self.moments1, self.moments2, gradients,
                                              self.decay_flags, self.optim_filter)

        if self.use_parallel:
            self.broadcast_params(optim_result)

        if not self.dynamic_lr:
            F.control_depend(lr, self.assignadd(self.global_step, 1))

        return optim_result