DEV Community

Super Kai (Kazuya Ito)
Super Kai (Kazuya Ito)

Posted on • Edited on

PReLU and ELU in PyTorch

Buy Me a Coffee

*Memos:

PReLU() can get the 1D or more D tensor of the zero or more values computed by PReLU function from the 1D or more D tensor of zero or more elements as shown below:

*Memos:

  • The 1st argument for initialization is num_parameters(Optional-Default:1-Type:int): *Memos:
    • It's the number of learnable parameters.
    • It must be 1 <= x.
  • The 2nd argument for initialization is init(Optional-Default:0.25-Type:float). *It's the initial value for learnable parameters.
  • The 3rd argument is device(Optional-Default:None-Type:str, int or device()): *Memos:
  • The 4th argument is dtype(Optional-Default:None-Type:dtype): *Memos:
  • The 1st argument is input(Required-Type:tensor of float).
  • You can also use prelu() with a tensor.

Image description

import torch
from torch import nn

my_tensor = torch.tensor([8., -3., 0., 1., 5., -2., -1., 4.])

prelu = nn.PReLU()
prelu(input=my_tensor)
# tensor([8.0000, -0.7500, 0.0000, 1.0000,
#         5.0000, -0.5000, -0.2500, 4.0000],
#        grad_fn=<PreluKernelBackward0>)

my_tensor.prelu(weight=torch.tensor(0.25))
# tensor([8.0000, -0.7500, 0.0000, 1.0000,
#         5.0000, -0.5000, -0.2500, 4.0000])

prelu
# PReLU(num_parameters=1)

prelu.num_parameters
# 1

prelu.init
# 0.25

prelu.weight
# Parameter containing:
# tensor([0.2500], requires_grad=True)

prelu = nn.PReLU(num_parameters=1, init=0.25)
prelu(input=my_tensor)
# tensor([8.0000, -0.7500, 0.0000, 1.0000,
#         5.0000, -0.5000, -0.2500, 4.0000],
#        grad_fn=<PreluKernelBackward0>)

prelu.weight
# Parameter containing:
# tensor([0.2500], requires_grad=True)

my_tensor = torch.tensor([[8., -3., 0., 1.],
                          [5., -2., -1., 4.]])
prelu = nn.PReLU()
prelu(input=my_tensor)
# tensor([[8.0000, -0.7500, 0.0000, 1.0000],
#         [5.0000, -0.5000, -0.2500, 4.0000]],
#        grad_fn=<PreluKernelBackward0>)

prelu.weight
# Parameter containing:
# tensor([0.2500], requires_grad=True)

prelu = nn.PReLU(num_parameters=4, init=0.25, device=None, dtype=None)
prelu(input=my_tensor)
# tensor([[8.0000, -0.7500, 0.0000, 1.0000],
#         [5.0000, -0.5000, -0.2500, 4.0000]],
#        grad_fn=<PreluKernelBackward0>)

prelu.weight
# Parameter containing:
# tensor([0.2500, 0.2500, 0.2500, 0.2500], requires_grad=True)

my_tensor = torch.tensor([[[8., -3.], [0., 1.]],
                          [[5., -2.], [-1., 4.]]])
prelu = nn.PReLU()
prelu(input=my_tensor)
# tensor([[[8.0000, -0.7500], [0.0000, 1.0000]],
#         [[5.0000, -0.5000], [-0.2500, 4.0000]]],
#        grad_fn=<PreluKernelBackward0>)

prelu.weight
# Parameter containing:
# tensor([0.2500], requires_grad=True)

prelu = nn.PReLU(num_parameters=2, init=0.25)
prelu(input=my_tensor)
# tensor([[[8.0000, -0.7500], [0.0000, 1.0000]],
#         [[5.0000, -0.5000], [-0.2500, 4.0000]]],
#        grad_fn=<PreluKernelBackward0>)

prelu.weight
# Parameter containing:
# tensor([0.2500, 0.2500], requires_grad=True)
Enter fullscreen mode Exit fullscreen mode

ELU() can get the 0D or more D tensor of the zero or more values computed by ELU function from the 0D or more D tensor of zero or more elements as shown below:

*Memos:

  • The 1st argument for initialization is alpha(Optional-Default:1.0-Type:float). *It's applied to negative input values.
  • The 2nd argument for initialization is inplace(Optional-Default:False-Type:bool): *Memos:
    • It does in-place operation.
    • Keep it False because it's problematic with True.
  • The 1st argument is input(Required-Type:tensor of float).

Image description

import torch
from torch import nn

my_tensor = torch.tensor([8., -3., 0., 1., 5., -2., -1., 4.])

elu = nn.ELU()
elu(input=my_tensor)
# tensor([8.0000, -0.9502, 0.0000, 1.0000, 5.0000, -0.8647, -0.6321, 4.0000])

elu
# ELU(alpha=1.0)

elu.alpha
# 1.0

elu.inplace
# False

elu = nn.ELU(alpha=1.0, inplace=True)
elu(input=my_tensor)
# tensor([8.0000, -0.9502, 0.0000, 1.0000, 5.0000, -0.8647, -0.6321, 4.0000])

my_tensor = torch.tensor([[8., -3., 0., 1.],
                          [5., -2., -1., 4.]])
elu = nn.ELU()
elu(input=my_tensor)
# tensor([[8.0000, -0.9502, 0.0000, 1.0000],
#         [5.0000, -0.8647, -0.6321, 4.0000]])

my_tensor = torch.tensor([[[8., -3.], [0., 1.]],
                          [[5., -2.], [-1., 4.]]])
elu = nn.ELU()
elu(input=my_tensor)
# tensor([[[8.0000, -0.9502], [0.0000, 1.0000]],
#         [[5.0000, -0.8647], [-0.6321, 4.0000]]])
Enter fullscreen mode Exit fullscreen mode

Heroku

Build apps, not infrastructure.

Dealing with servers, hardware, and infrastructure can take up your valuable time. Discover the benefits of Heroku, the PaaS of choice for developers since 2007.

Visit Site

Top comments (0)

Billboard image

The Next Generation Developer Platform

Coherence is the first Platform-as-a-Service you can control. Unlike "black-box" platforms that are opinionated about the infra you can deploy, Coherence is powered by CNC, the open-source IaC framework, which offers limitless customization.

Learn more