optimization.py

# Copyright 2023 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.

"""Customized optimizer to match paper results."""

import dataclasses
from typing import List, Optional

import tensorflow as tf

from official.modeling import optimization
from official.nlp import optimization as nlp_optimization


@dataclasses.dataclass
class ViTAdamWConfig(optimization.AdamWeightDecayConfig):
  layer_decay: Optional[float] = 1.0
  vars_substr: Optional[List[str]] = None
  layers_idx: Optional[List[int]] = None


@dataclasses.dataclass
class OptimizerConfig(optimization.OptimizerConfig):
  vit_adamw: ViTAdamWConfig = ViTAdamWConfig()


@dataclasses.dataclass
class OptimizationConfig(optimization.OptimizationConfig):
  """Configuration for optimizer and learning rate schedule.

  Attributes:
    optimizer: optimizer oneof config.
    ema: optional exponential moving average optimizer config, if specified, ema
      optimizer will be used.
    learning_rate: learning rate oneof config.
    warmup: warmup oneof config.
  """
  optimizer: OptimizerConfig = OptimizerConfig()


# TODO(frederickliu): figure out how to make this configuable.
# TODO(frederickliu): Study if this is needed.
class _ViTAdamW(nlp_optimization.AdamWeightDecay):
  """Custom AdamW to support different lr scaling for backbone.

  The code is copied from AdamWeightDecay and Adam with learning scaling.
  """

  def __init__(self,
               learning_rate=0.001,
               beta_1=0.9,
               beta_2=0.999,
               epsilon=1e-7,
               amsgrad=False,
               weight_decay_rate=0.0,
               include_in_weight_decay=None,
               exclude_from_weight_decay=None,
               gradient_clip_norm=1.0,
               layer_decay=1.0,
               vars_substr=None,
               layers_idx=None,
               name='ViTAdamWeightDecay',
               **kwargs):
    super(_ViTAdamW,
          self).__init__(learning_rate, beta_1, beta_2, epsilon, amsgrad,
                         weight_decay_rate, include_in_weight_decay,
                         exclude_from_weight_decay, gradient_clip_norm, name,
                         **kwargs)
    self._layer_decay = layer_decay
    self._vars_substr = vars_substr
    self._layers_idx = layers_idx
    self._max_idx = max(layers_idx) if layers_idx is not None else 0

  def _resource_apply_dense(self, grad, var, apply_state=None):
    lr_t, kwargs = self._get_lr(var.device, var.dtype.base_dtype, apply_state)
    apply_state = kwargs['apply_state']
    if self._layer_decay != 1.0 and self._vars_substr is not None and self._layers_idx is not None:
      for var_substr, idx in zip(self._vars_substr, self._layers_idx):
        if var_substr in var.name:
          lr_t = lr_t * (self._layer_decay ** (self._max_idx - idx))
          break
    decay = self._decay_weights_op(var, lr_t, apply_state)
    with tf.control_dependencies([decay]):
      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 = coefficients['lr_t']
      if self._layer_decay != 1.0 and self._vars_substr is not None and self._layers_idx is not None:
        for var_substr, idx in zip(self._vars_substr, self._layers_idx):
          if var_substr in var.name:
            lr = lr * (self._layer_decay ** (self._max_idx - idx))
            break

      if not self.amsgrad:
        return tf.raw_ops.ResourceApplyAdam(
            var=var.handle,
            m=m.handle,
            v=v.handle,
            beta1_power=coefficients['beta_1_power'],
            beta2_power=coefficients['beta_2_power'],
            lr=lr,
            beta1=coefficients['beta_1_t'],
            beta2=coefficients['beta_2_t'],
            epsilon=coefficients['epsilon'],
            grad=grad,
            use_locking=self._use_locking)
      else:
        vhat = self.get_slot(var, 'vhat')
        return tf.raw_ops.ResourceApplyAdamWithAmsgrad(
            var=var.handle,
            m=m.handle,
            v=v.handle,
            vhat=vhat.handle,
            beta1_power=coefficients['beta_1_power'],
            beta2_power=coefficients['beta_2_power'],
            lr=lr,
            beta1=coefficients['beta_1_t'],
            beta2=coefficients['beta_2_t'],
            epsilon=coefficients['epsilon'],
            grad=grad,
            use_locking=self._use_locking)

  def _resource_apply_sparse(self, grad, var, indices, apply_state=None):
    lr_t, kwargs = self._get_lr(var.device, var.dtype.base_dtype, apply_state)
    apply_state = kwargs['apply_state']
    if self._layer_decay != 1.0 and self._vars_substr is not None and self._layers_idx is not None:
      for var_substr, idx in zip(self._vars_substr, self._layers_idx):
        if var_substr in var.name:
          lr_t = lr_t * (self._layer_decay ** (self._max_idx - idx))
          break
    decay = self._decay_weights_op(var, lr_t, apply_state)
    with tf.control_dependencies([decay]):
      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 = tf.compat.v1.assign(m, m * coefficients['beta_1_t'],
                                use_locking=self._use_locking)
      with tf.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 = tf.compat.v1.assign(v, v * coefficients['beta_2_t'],
                                use_locking=self._use_locking)
      with tf.control_dependencies([v_t]):
        v_t = self._resource_scatter_add(v, indices, v_scaled_g_values)
      lr = coefficients['lr_t']
      if self._layer_decay != 1.0 and self._vars_substr is not None and self._layers_idx is not None:
        for var_substr, idx in zip(self._vars_substr, self._layers_idx):
          if var_substr in var.name:
            lr = lr * (self._layer_decay ** (self._max_idx - idx))
            break
      if not self.amsgrad:
        v_sqrt = tf.sqrt(v_t)
        var_update = tf.compat.v1.assign_sub(
            var, lr * m_t / (v_sqrt + coefficients['epsilon']),
            use_locking=self._use_locking)
        return tf.group(*[var_update, m_t, v_t])
      else:
        v_hat = self.get_slot(var, 'vhat')
        v_hat_t = tf.maximum(v_hat, v_t)
        with tf.control_dependencies([v_hat_t]):
          v_hat_t = tf.compat.v1.assign(
              v_hat, v_hat_t, use_locking=self._use_locking)
        v_hat_sqrt = tf.sqrt(v_hat_t)
        var_update = tf.compat.v1.assign_sub(
            var,
            lr* m_t / (v_hat_sqrt + coefficients['epsilon']),
            use_locking=self._use_locking)
        return tf.group(*[var_update, m_t, v_t, v_hat_t])

optimization.register_optimizer_cls('vit_adamw', _ViTAdamW)