比较与tf.compat.v1.train.RMSPropOptimizer的功能差异

tf.compat.v1.train.RMSPropOptimizer

tf.compat.v1.train.RMSPropOptimizer(
    learning_rate,
    decay=0.9,
    momentum=0.0,
    epsilon=1e-10,
    use_locking=False,
    centered=False,
    name='RMSProp'
) -> Tensor

更多内容详见tf.compat.v1.train.RMSPropOptimizer。

mindspore.nn.RMSProp

class mindspore.nn.RMSProp(
    params,
    learning_rate=0.1,
    decay=0.9,
    momentum=0.0,
    epsilon=1e-10,
    use_locking=False,
    centered=False,
    loss_scale=1.0,
    weight_decay=0.0
)(gradients) -> Tensor

更多内容详见mindspore.nn.RMSProp。

差异对比

TensorFlow：实现均方根传播算法（RMSProp）的优化器功能。

MindSpore：MindSpore此API实现功能与TensorFlow基本一致。

分类	子类	TensorFlow	MindSpore	差异
参数	参数1	learning_rate	learning_rate	功能一致，TensorFlow无默认值
	参数2	decay	decay	-
	参数3	momentum	momentum	-
	参数4	epsilon	epsilon	-
	参数5	use_locking	use_locking	-
	参数6	centered	centered	-
	参数7	name	-	不涉及
	参数8	-	params	由Parameter类组成的列表或由字典组成的列表，TensorFlow中无此参数
	参数9	-	loss_scale	梯度缩放系数，默认值：1.0，TensorFlow中无此参数
	参数10	-	weight_decay	权重衰减（L2 penalty），默认值：0.0，TensorFlow中无此参数
	参数11	-	gradients	优化器中 params 的梯度，TensorFlow中无此参数

代码示例

两API实现功能一致。

# TensorFlow
import tensorflow as tf
import numpy as np

tf.compat.v1.disable_v2_behavior()
tf.compat.v1.disable_eager_execution()

param_np = np.ones(7).astype(np.float32)
indices = np.array([1, 2, 3, 4, 5, 6]).astype(np.int32)
label = np.zeros((2, 3)).astype(np.float32)
label_shape = (2, 3)
axis = 0
epoch = 3
param_tf = tf.Variable(param_np)
indices_tf = tf.Variable(indices)
label = tf.Variable(label)
net = tf.raw_ops.GatherV2(params=param_tf, indices=indices_tf, axis=axis, batch_dims=0, name=None)
net = tf.reshape(net, label_shape)
criterion = tf.compat.v1.losses.softmax_cross_entropy(onehot_labels=label, logits=net,
                                                      reduction=tf.compat.v1.losses.Reduction.MEAN)
opt = tf.compat.v1.train.RMSPropOptimizer(learning_rate=0.1).minimize(criterion)
init = tf.compat.v1.global_variables_initializer()
with tf.compat.v1.Session() as ss:
    ss.run(init)
    for i in range(epoch):
        loss = criterion.eval()
        ss.run(opt)
    output = net.eval()
    net = ss.run(net)
out_tf = output.astype(np.float32)
print(out_tf)
# [[0.94458014 0.94458014 0.94458014]
#  [0.94458014 0.94458014 0.94458014]]

# MindSpore
import numpy as np
import mindspore.nn as nn
from mindspore import Tensor
from mindspore.ops import operations as op
from mindspore import Parameter
from mindspore.nn import SoftmaxCrossEntropyWithLogits
from mindspore.nn import TrainOneStepCell, WithLossCell
from mindspore.nn import Cell

class NetWithGatherV2(Cell):
    def __init__(self, param_np, label, axis=0):
        super(NetWithGatherV2, self).__init__()
        self.param = Parameter(Tensor(param_np), name="w1")
        self.gatherv2 = op.GatherV2()
        self.reshape = op.Reshape()
        self.axis = axis
        self.label = label

    def construct(self, indices):
        x = self.gatherv2(self.param, indices, self.axis)
        return self.reshape(x, self.label)


param_np = np.ones(7).astype(np.float32)
indices = np.array([1, 2, 3, 4, 5, 6]).astype(np.int32)
label = np.zeros((2, 3)).astype(np.float32)
label_shape = (2, 3)
epoch = 3
inputs = Tensor(indices)
label = Tensor(label)
net = NetWithGatherV2(param_np, label_shape, axis=0)
criterion = SoftmaxCrossEntropyWithLogits(reduction='mean')
optimizer = nn.RMSProp(params=net.trainable_params())
net_with_criterion = WithLossCell(net, criterion)
train_network = TrainOneStepCell(net_with_criterion, optimizer)
train_network.set_train()
for i in range(epoch):
    train_network(inputs, label)
out_ms = net(inputs).asnumpy()
print(out_ms)
# [[0.94458014 0.94458014 0.94458014]
#  [0.94458014 0.94458014 0.94458014]]