Adam OptimizerV2 implementation, weights frozen -- reference issue: https://github.com/tensorflow/tensorflow/issues/33227

adam_v2_frozen.py

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, control_flow_ops, math_ops, state_ops
from tensorflow.python.util.tf_export import keras_export
import keras.backend as K


@keras_export('keras.optimizers.Adam')
class Adam(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).
    """
    r"""Construct a new Adam optimizer.

    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".
        name: Optional name for the operations created when applying gradients.
            Defaults to "Adam".  @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
        **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.
    """

    def __init__(self, learning_rate=0.001, beta_1=0.9, beta_2=0.999,
                 epsilon=None, decay=0., amsgrad=False,
                 name="Adam", **kwargs):
        super(Adam, 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


    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 _resource_apply_dense(self, grad, var):
        var_dtype = var.dtype.base_dtype
        lr_t = self._decayed_lr(var_dtype)
        m = self.get_slot(var, 'm')
        v = self.get_slot(var, 'v')
        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))
        epsilon_t = ops.convert_to_tensor(self.epsilon, var_dtype)
        
        m_t = state_ops.assign(m,
        		   beta_1_t * m + (1.0 - beta_1_t) * grad,
        		   use_locking=self._use_locking)
        v_t = state_ops.assign(v,
        		   beta_2_t * v + (1.0 - beta_2_t) * math_ops.square(grad),
        		   use_locking=self._use_locking)
        if self.amsgrad:
            vhat = self.get_slot(var, 'vhat')
            vhat_t = state_ops.assign(vhat, math_ops.maximum(vhat, v_t),
                                      use_locking=self._use_locking)
            var_delta = m_t / (K.sqrt(vhat_t) + epsilon_t)
        else:
            var_delta = m_t / (K.sqrt(v_t) + epsilon_t)
        var_t = math_ops.sub(var, lr_t * var_delta)
    
        if 'dense_1/kernel' in var.name:
            print(var)
            print(K.eval(var_t))
        var_update = state_ops.assign_sub(var, lr_t * var_delta,
                                          use_locking=self._use_locking)
        if 'dense_1/kernel' in var.name:
            print(var)
        	
        updates = [var_update, m_t, v_t]
        if self.amsgrad:
            updates.append(vhat_t)
        return control_flow_ops.group(*updates)

            
    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)

        if not self.amsgrad:
            v_sqrt = math_ops.sqrt(v_t)
            var_update = state_ops.assign_sub(
                var, coefficients['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,
                coefficients['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 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(Adam, self).set_weights(weights)
        
    def get_config(self):
        config = super(Adam, 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,
        })
        return config