比较与tf.keras.optimizers.Adagrad的功能差异

查看源文件

tf.keras.optimizers.Adagrad

tf.keras.optimizers.Adagrad(
    learning_rate=0.001,
    initial_accumulator_value=0.1,
    epsilon=1e-07,
    name='Adagrad',
    **kwargs
) -> Tensor

更多内容详见tf.keras.optimizers.Adagrad

mindspore.nn.Adagrad

class mindspore.nn.Adagrad(
    params,
    accum=0.1,
    learning_rate=0.001,
    update_slots=True,
    loss_scale=1.0,
    weight_decay=0.0
)(grads) -> Tensor

更多内容详见mindspore.nn.Adagrad

差异对比

TensorFlow:Adagrad是一个具有特定参数学习率的优化器,用来实现Adagrad算法,它根据训练期间参数更新的频率进行调整。参数接收的更新越多,更新越小。

MindSpore:MindSpore此API实现功能与TensorFlow基本一致,部分参数名不一样,并且比TensorFlow多出update_slots、loss_scale、weight_decay参数。

分类

子类

TensorFlow

MindSpore

差异

参数

参数1

learning_rate

learning_rate

-

参数2

initial_accumulator_value

accum

功能一致,参数名不同

参数3

epsilon

-

TensorFlow用于保持数值稳定性的小浮点值,MindSpore无此参数

参数4

name

-

不涉及

参数5

**kwargs

-

不涉及

参数6

-

params

Parameter组成的列表或字典组成的列表,TensorFlow中无此参数

参数7

-

update_slots

值如果为True,则更新累加器,TensorFlow中无此参数

参数8

-

loss_scale

梯度缩放系数,默认值:1.0,TensorFlow中无此参数

参数9

-

weight_decay

权重衰减(L2 penalty),默认值:0.0,TensorFlow中无此参数

输入

单输入

-

grads

优化器中 params 的梯度,TensorFlow中无此参数

代码示例

两API实现功能基本一致。

# TensorFlow
import tensorflow as tf

opt = tf.keras.optimizers.Adagrad(initial_accumulator_value=0.1, learning_rate=0.1)
var = tf.Variable(1.0)
val0 = var.value()
loss = lambda: (var ** 2)/2.0
step_count = opt.minimize(loss, [var]).numpy()
val1 = var.value()
print([val1.numpy()])
# [0.9046537]
step_count = opt.minimize(loss, [var]).numpy()
val2 = var.value()
print([val2.numpy()])
# [0.8393387]

# MindSpore
import numpy as np
import mindspore.nn as nn
import mindspore as ms
from mindspore.dataset import NumpySlicesDataset
from mindspore.train import Model

class Net(nn.Cell):
    def __init__(self):
        super(Net, self).__init__()
        self.w = ms.Parameter(ms.Tensor(np.array([1.0], np.float32)), name='w')

    def construct(self, x):
        f = self.w * x
        return f

class MyLoss(nn.LossBase):
    def __init__(self, reduction='none'):
        super(MyLoss, self).__init__()

    def construct(self, y, y_pred):
        return (y - y_pred) ** 2 / 2.0
net = Net()
loss = MyLoss()
optim = nn.Adagrad(params=net.trainable_params(), accum=0.1, learning_rate=0.1)
model = Model(net, loss_fn=loss, optimizer=optim)
data_x = np.array([1.0], dtype=np.float32)
data_y = np.array([0.0], dtype=np.float32)
data = NumpySlicesDataset((data_x, data_y), ["x", "y"])
input_x = ms.Tensor(np.array([1.0], np.float32))
y0 = net(input_x)
model.train(1, data)
y1 = net(input_x)
print(y1)
# [0.9046537]
model.train(1, data)
y2 = net(input_x)
print(y2)
# [0.8393387]