adamlrm.py

# Copyright 2018 The TensorFlow Authors. All Rights Reserved.
#
# 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.
# ==============================================================================
"""Adam with learning rate multipliers for TensorFlow."""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function

from tensorflow.python.framework import ops
from tensorflow.python.keras import backend_config
from tensorflow.python.keras.optimizer_v2 import optimizer_v2
from tensorflow.python.ops import array_ops
from tensorflow.python.ops import control_flow_ops
from tensorflow.python.ops import math_ops
from tensorflow.python.ops import state_ops
from tensorflow.python.training import training_ops
from tensorflow.python.util.tf_export import keras_export


@keras_export('keras.optimizers.AdamLRM')
class AdamLRM(optimizer_v2.OptimizerV2):
  """Optimizer that implements the Adam algorithm.

  Adam optimization is a stochastic gradient descent method that is based on
  adaptive estimation of first-order and second-order moments.
  According to the paper
  [Adam: A Method for Stochastic Optimization. Kingma et al.,
  2014](http://arxiv.org/abs/1412.6980),
   the method is "*computationally efficient, has little memory
  requirement, invariant to diagonal rescaling of gradients, and is well suited
  for problems that are large in terms of data/parameters*".

  For AMSGrad see [On The Convergence Of Adam And Beyond.
  Reddi et al., 5-8](https://openreview.net/pdf?id=ryQu7f-RZ).
  """

  def __init__(self,
               learning_rate=0.001,
               beta_1=0.9,
               beta_2=0.999,
               epsilon=1e-7,
               amsgrad=False,
               lr_multiplier={},
               name='AdamLRM',
               **kwargs):
    r"""Construct a new Adam optimizer with learning rate multipliers.

    If amsgrad = False:
      Initialization:

      $$m_0 := 0 \text{(Initialize initial 1st moment vector)}$$
      $$v_0 := 0 \text{(Initialize initial 2nd moment vector)}$$
      $$t := 0 \text{(Initialize timestep)}$$

      The update rule for `variable` with gradient `g` uses an optimization
      described at the end of section 2 of the paper:

      $$t := t + 1$$
      $$lr_t := \text{learning\_rate} * \sqrt{1 - beta_2^t} / (1 - beta_1^t)$$

      $$m_t := beta_1 * m_{t-1} + (1 - beta_1) * g$$
      $$v_t := beta_2 * v_{t-1} + (1 - beta_2) * g * g$$
      $$variable := variable - lr_t * m_t / (\sqrt{v_t} + \epsilon)$$

    If amsgrad = True:
      Initialization:

      $$m_0 := 0 \text{(Initialize initial 1st moment vector)}$$
      $$v_0 := 0 \text{(Initialize initial 2nd moment vector)}$$
      $$v_hat_0 := 0 \text{(Initialize initial 2nd moment vector)}$$
      $$t := 0 \text{(Initialize timestep)}$$

      The update rule for `variable` with gradient `g` uses an optimization
      described at the end of section 2 of the paper:

      $$t := t + 1$$
      $$lr_t := \text{learning\_rate} * \sqrt{1 - beta_2^t} / (1 - beta_1^t)$$

      $$m_t := beta_1 * m_{t-1} + (1 - beta_1) * g$$
      $$v_t := beta_2 * v_{t-1} + (1 - beta_2) * g * g$$
      $$v_hat_t := max(v_hat_{t-1}, v_t)$$
      $$variable := variable - lr_t * m_t / (\sqrt{v_hat_t} + \epsilon)$$

    The default value of 1e-7 for epsilon might not be a good default in
    general. For example, when training an Inception network on ImageNet a
    current good choice is 1.0 or 0.1. Note that since AdamOptimizer uses the
    formulation just before Section 2.1 of the Kingma and Ba paper rather than
    the formulation in Algorithm 1, the "epsilon" referred to here is "epsilon
    hat" in the paper.

    The sparse implementation of this algorithm (used when the gradient is an
    IndexedSlices object, typically because of `tf.gather` or an embedding
    lookup in the forward pass) does apply momentum to variable slices even if
    they were not used in the forward pass (meaning they have a gradient equal
    to zero). Momentum decay (beta1) is also applied to the entire momentum
    accumulator. This means that the sparse behavior is equivalent to the dense
    behavior (in contrast to some momentum implementations which ignore momentum
    unless a variable slice was actually used).

    Args:
      learning_rate: A Tensor or a floating point value.  The learning rate.
      beta_1: A float value or a constant float tensor. The exponential decay
        rate for the 1st moment estimates.
      beta_2: A float value or a constant float tensor. The exponential decay
        rate for the 2nd moment estimates.
      epsilon: A small constant for numerical stability. This epsilon is
        "epsilon hat" in the Kingma and Ba paper (in the formula just before
        Section 2.1), not the epsilon in Algorithm 1 of the paper.
      amsgrad: boolean. Whether to apply AMSGrad variant of this algorithm from
        the paper "On the Convergence of Adam and beyond".
      lr_multiplier: A dictionary with variable names as keys and learning rate
        multipliers as values. Learning rates of all variables which names
        start with each key are multiplied by these values.
      name: Optional name for the operations created when applying gradients.
        Defaults to "Adam".
      **kwargs: keyword arguments. Allowed to be {`clipnorm`, `clipvalue`, `lr`,
        `decay`}. `clipnorm` is clip gradients by norm; `clipvalue` is clip
        gradients by value, `decay` is included for backward compatibility to
        allow time inverse decay of learning rate. `lr` is included for backward
        compatibility, recommended to use `learning_rate` instead.

    @compatibility(eager)
    When eager execution is enabled, `learning_rate`, `beta_1`, `beta_2`,
    and `epsilon` can each be a callable that takes no arguments and
    returns the actual value to use. This can be useful for changing these
    values across different invocations of optimizer functions.
    @end_compatibility
    """

    super(AdamLRM, self).__init__(name, **kwargs)
    self._set_hyper('learning_rate', kwargs.get('lr', learning_rate))
    self._set_hyper('decay', self._initial_decay)
    self._set_hyper('beta_1', beta_1)
    self._set_hyper('beta_2', beta_2)
    self.epsilon = epsilon or backend_config.epsilon()
    self.amsgrad = amsgrad
    self._lrm_names = list(lr_multiplier.keys())
    for k,v in lr_multiplier.items():
      self._set_hyper(f'lrm_{k}', v)

  def _create_slots(self, var_list):
    # Create slots for the first and second moments.
    # Separate for-loops to respect the ordering of slot variables from v1.
    for var in var_list:
      self.add_slot(var, 'm')
    for var in var_list:
      self.add_slot(var, 'v')
    if self.amsgrad:
      for var in var_list:
        self.add_slot(var, 'vhat')

  def _prepare_local(self, var_device, var_dtype, apply_state):
    super(AdamLRM, self)._prepare_local(var_device, var_dtype, apply_state)

    local_step = math_ops.cast(self.iterations + 1, var_dtype)
    beta_1_t = array_ops.identity(self._get_hyper('beta_1', var_dtype))
    beta_2_t = array_ops.identity(self._get_hyper('beta_2', var_dtype))
    beta_1_power = math_ops.pow(beta_1_t, local_step)
    beta_2_power = math_ops.pow(beta_2_t, local_step)
    lr = (apply_state[(var_device, var_dtype)]['lr_t'] *
          (math_ops.sqrt(1 - beta_2_power) / (1 - beta_1_power)))
    apply_state[(var_device, var_dtype)].update(dict(
        lr=lr,
        epsilon=ops.convert_to_tensor(self.epsilon, var_dtype),
        beta_1_t=beta_1_t,
        beta_1_power=beta_1_power,
        one_minus_beta_1_t=1 - beta_1_t,
        beta_2_t=beta_2_t,
        beta_2_power=beta_2_power,
        one_minus_beta_2_t=1 - beta_2_t
    ))

  def set_weights(self, weights):
    params = self.weights
    # If the weights are generated by Keras V1 optimizer, it includes vhats
    # even without amsgrad, i.e, V1 optimizer has 3x + 1 variables, while V2
    # optimizer has 2x + 1 variables. Filter vhats out for compatibility.
    num_vars = int((len(params) - 1) / 2)
    if len(weights) == 3 * num_vars + 1:
      weights = weights[:len(params)]
    super(AdamLRM, self).set_weights(weights)

  def _resource_apply_dense(self, grad, var, apply_state=None):
    var_device, var_dtype = var.device, var.dtype.base_dtype
    coefficients = ((apply_state or {}).get((var_device, var_dtype))
                    or self._fallback_apply_state(var_device, var_dtype))

    m = self.get_slot(var, 'm')
    v = self.get_slot(var, 'v')

    lr_t = coefficients['lr_t']
    for k in self._lrm_names:
      if var.name.startswith(k):
        lr_t = coefficients['lr_t'] * self._get_hyper(f'lrm_{k}', var.dtype)

    if not self.amsgrad:
      return training_ops.resource_apply_adam(
          var.handle,
          m.handle,
          v.handle,
          coefficients['beta_1_power'],
          coefficients['beta_2_power'],
          lr_t,
          coefficients['beta_1_t'],
          coefficients['beta_2_t'],
          coefficients['epsilon'],
          grad,
          use_locking=self._use_locking)
    else:
      vhat = self.get_slot(var, 'vhat')
      return training_ops.resource_apply_adam_with_amsgrad(
          var.handle,
          m.handle,
          v.handle,
          vhat.handle,
          coefficients['beta_1_power'],
          coefficients['beta_2_power'],
          lr_t,
          coefficients['beta_1_t'],
          coefficients['beta_2_t'],
          coefficients['epsilon'],
          grad,
          use_locking=self._use_locking)

  def _resource_apply_sparse(self, grad, var, indices, apply_state=None):
    var_device, var_dtype = var.device, var.dtype.base_dtype
    coefficients = ((apply_state or {}).get((var_device, var_dtype))
                    or self._fallback_apply_state(var_device, var_dtype))

    # m_t = beta1 * m + (1 - beta1) * g_t
    m = self.get_slot(var, 'm')
    m_scaled_g_values = grad * coefficients['one_minus_beta_1_t']
    m_t = state_ops.assign(m, m * coefficients['beta_1_t'],
                           use_locking=self._use_locking)
    with ops.control_dependencies([m_t]):
      m_t = self._resource_scatter_add(m, indices, m_scaled_g_values)

    # v_t = beta2 * v + (1 - beta2) * (g_t * g_t)
    v = self.get_slot(var, 'v')
    v_scaled_g_values = (grad * grad) * coefficients['one_minus_beta_2_t']
    v_t = state_ops.assign(v, v * coefficients['beta_2_t'],
                           use_locking=self._use_locking)
    with ops.control_dependencies([v_t]):
      v_t = self._resource_scatter_add(v, indices, v_scaled_g_values)

    lr = coefficients['lr']
    for k in self._lrm_names:
      if var.name.startswith(k):
        lr = coefficients['lr'] * self._get_hyper(f'lrm_{k}', var.dtype)

    if not self.amsgrad:
      v_sqrt = math_ops.sqrt(v_t)
      var_update = state_ops.assign_sub(
          var, lr * m_t / (v_sqrt + coefficients['epsilon']),
          use_locking=self._use_locking)
      return control_flow_ops.group(*[var_update, m_t, v_t])
    else:
      v_hat = self.get_slot(var, 'vhat')
      v_hat_t = math_ops.maximum(v_hat, v_t)
      with ops.control_dependencies([v_hat_t]):
        v_hat_t = state_ops.assign(
            v_hat, v_hat_t, use_locking=self._use_locking)
      v_hat_sqrt = math_ops.sqrt(v_hat_t)
      var_update = state_ops.assign_sub(
          var,
          lr * m_t / (v_hat_sqrt + coefficients['epsilon']),
          use_locking=self._use_locking)
      return control_flow_ops.group(*[var_update, m_t, v_t, v_hat_t])

  def get_config(self):
    config = super(AdamLRM, self).get_config()
    config.update({
        'learning_rate': self._serialize_hyperparameter('learning_rate'),
        'decay': self._serialize_hyperparameter('decay'),
        'beta_1': self._serialize_hyperparameter('beta_1'),
        'beta_2': self._serialize_hyperparameter('beta_2'),
        'epsilon': self.epsilon,
        'amsgrad': self.amsgrad,
    })
    for k in self._lrm_names:
      config[k] = self._serialize_hyperparameter(f'lrm_{k}')
    return config