为什么在 Tensorflow 中构建自定义优化器时出现“NotImplementedError()”

Question

我正在研究图像分类，并且正在尝试在 Tensorflow 中实现自定义优化器（基于 ELSEVIER 上发表的论文），

我尝试修改代码如下：我有一些其他功能，但都与预处理和模型架构等有关。我的优化器代码如下；

import os
os.environ['TF_KERAS'] = '1'
from tensorflow import keras
import tensorflow as tf
from sklearn.model_selection import train_test_split
from sklearn.utils import shuffle
import cv2
import imutils
import matplotlib.pyplot as plt
from os import listdir
from sklearn.metrics import confusion_matrix,classification_report
import logging, warnings
import numpy as np
from tensorflow.python.training import optimizer
from tensorflow.python.ops import math_ops, state_ops, control_flow_ops, variable_scope
from tensorflow.python.framework import ops

class BPVAM(optimizer.Optimizer):
     """Back-propagation algorithm with variable adaptive momentum.

    Variables are updated in two steps:
    1) v(t + 1) = alpha * v(t)- lr * g(t)
    2) w(t + 1) = w(t) + v(t + 1)

    where
        - v(t + 1): delta for update at step t + 1
        - w(t + 1): weights at step t + 1 (after update)
        - g(t): gradients at step t.
        - lr: learning rate
        - alpha: momentum parameter

    In the algorithm alpha is not fixed. It is variable and it is parametrized by:
        alpha(t) = lambda / (1 - beta ^ t)
    """

    def __init__(
            self,
            lr: float = 0.001,
            lam: float = 0.02,
            beta: float = 0.998,
            use_locking: bool = False,
            name: str = 'BPVAM'
    ):
        """
            Args:
                lr: learning rate
                lam: momentum parameter
                beta: momentum parameter
                use_locking:
                name:
        """
        super(BPVAM, self).__init__(use_locking, name)

        self._lr = lr
        self._lambda = lam
        self._beta = beta

        self._lr_tensor = None
        self._lambda_tensor = None
        self._beta_tensor = None

    def _create_slots(self, var_list):
        for v in var_list:
            self._zeros_slot(v, 'v', self._name)
            self._get_or_make_slot(v,
                                   ops.convert_to_tensor(self._beta),
                                   'beta',
                                   self._name)

    def _prepare(self):
        self._lr_tensor = ops.convert_to_tensor(self._lr, name='lr')
        self._lambda_tensor = ops.convert_to_tensor(self._lambda, name='lambda')

    def _apply_dense(self, grad, var):
        lr_t = math_ops.cast(self._lr_tensor, var.dtype.base_dtype)
        lambda_t = math_ops.cast(self._lambda_tensor, var.dtype.base_dtype)

        v = self.get_slot(var, 'v')
        betas = self.get_slot(var, 'beta')

        beta_t = state_ops.assign(betas, betas * betas)

        alpha = lambda_t / (1 - beta_t)

        v_t = state_ops.assign(v, alpha * v - lr_t * grad)

        var_update = state_ops.assign_add(var, v_t, use_locking=self._use_locking)

        return control_flow_ops.group(*[beta_t, v_t, var_update])

在我创建优化器并运行之后；

myopt = BPVAM()
model.compile(optimizer= myopt, loss='sparse_categorical_crossentropy', metrics=['accuracy'])

我收到此错误消息；

Traceback (most recent call last):
  File "/Users/classification.py", line 264, in <module>model.fit(x=X_train,y=y_train, batch_size=32, epochs=50, validation_data=(X_val, y_val))

File"/Users/ venv/lib/python3.7/sitepackages/tensorflow/python/keras/engine/training.py", line 780, in fit steps_name='steps_per_epoch')

File"/Users/venv/lib/python3.7/sitepackages/tensorflow/python/keras/engine/training_arrays.py", line 157, in model_iteration f = _make_execution_function(model, mode)

 File"/Users/ venv/lib/python3.7/sitepackages/tensorflow/python/keras/engine/training_arrays.py", line 532, in _make_execution_function return model._make_execution_function(mode)

File"/Users/ venv/lib/python3.7/sitepackages/tensorflow/python/keras/engine/training.py", line 2276, in _make_execution_function self._make_train_function()

File"/Users/ venv/lib/python3.7/sitepackages/tensorflow/python/keras/engine/training.py", line 2219, in _make_train_function params=self._collected_trainable_weights, loss=self.total_loss)

File "/Users/ venv/lib/python3.7/site-packages/tensorflow/python/keras/optimizers.py", line 753, in get_updates grads, global_step=self.iterations)

File "/Users/ venv/lib/python3.7/site-packages/tensorflow/python/training/optimizer.py", line 614, in apply_gradients update_ops.append(processor.update_op(self, grad))

File "/Users/venv/lib/python3.7/site-packages/tensorflow/python/training/optimizer.py", line 171, in update_op update_op = optimizer._resource_apply_dense(g, self._v)

File "/Users/venv/lib/python3.7/site-packages/tensorflow/python/training/optimizer.py", line 954, in _resource_apply_dense

raise NotImplementedError()
NotImplementedError

我不明白问题出在哪里。 我使用的是 Tensorflow 1.14.0 版本和 python 3.7。 我创建了虚拟环境并尝试了其他 tensorflow 和 python 版本，但它仍然不起作用。

Answer 1

为了使用从tensorflow.python.training.optimizer.Optimizer继承的类，您必须至少实现以下方法：

• _apply_dense
• _resource_apply_dense
• _apply_sparse

查看Optimizer的源代码以获取更多信息。

由于您尝试实现自定义动量方法，因此您可能希望直接MomentumOptimizer 。