L1Decay

class paddle.regularizer. L1Decay ( coeff=0.0 ) [source]

Implement the L1 Weight Decay Regularization, which encourages the weights to be sparse.

It can be set in api_paddle_ParamAttr or optimizer (such as api_paddle_optimizer_Momentum ). When set in ParamAttr , it only takes effect for trainable parameters in this layer. When set in optimizer , it takes effect for all trainable parameters. When set together, ParamAttr has higher priority than optimizer , which means that for a trainable parameter, if regularizer is defined in its ParamAttr, then the regularizer in Optimizer will be ignored. Otherwise the regularizer in Optimizer will be used.

In the implementation, the loss function of L1 Weight Decay Regularization is as follows:

\[loss = coeff * reduce\_sum(abs(x))\]
Parameters

coeff (float, optional) – regularization coeff. Default:0.0.

Examples

# Example1: set Regularizer in optimizer
import paddle
from paddle.regularizer import L1Decay
import numpy as np
linear = paddle.nn.Linear(10, 10)
inp = paddle.rand(shape=[10, 10], dtype="float32")
out = linear(inp)
loss = paddle.mean(out)
beta1 = paddle.to_tensor([0.9], dtype="float32")
beta2 = paddle.to_tensor([0.99], dtype="float32")
momentum = paddle.optimizer.Momentum(
    learning_rate=0.1,
    parameters=linear.parameters(),
    weight_decay=L1Decay(0.0001))
back = out.backward()
momentum.step()
momentum.clear_grad()

# Example2: set Regularizer in parameters
# Set L1 regularization in parameters.
# Global regularizer does not take effect on my_conv2d for this case.
from paddle.nn import Conv2D
from paddle import ParamAttr
from paddle.regularizer import L2Decay

my_conv2d = Conv2D(
        in_channels=10,
        out_channels=10,
        kernel_size=1,
        stride=1,
        padding=0,
        weight_attr=ParamAttr(regularizer=L2Decay(coeff=0.01)),
        bias_attr=False)