Comparing the function difference with torch.nn.Module.buffers

torch.nn.Module.buffers

torch.nn.Module.buffers(recurse=True)

For more information, see torch.nn.Module.buffers.

mindspore.nn.Cell.untrainable_params

mindspore.nn.Cell.untrainable_params(recurse=True)

For more information, see mindspore.nn.Cell.untrainable_params.

Differences

In PyTorch, the network has three concepts: parameter, buffer, and state, where state is the collection of parameter and buffer. parameter can use the requires_grad attribute to distinguish whether the parameter in the network needs to be optimized; buffer is mostly defined as an invariant in the network, for example, when defining the network, the running_mean and running_var in BN will be automatically register as buffer; users can also register parameter and buffer through related interfaces.

-torch.nn.Module.buffers: Get the buffer in the network, and return a generator.

-torch.nn.Module.named_buffers: Get the name of buffer and buffer itself in the network, and return a generator.

In MindSpore, there is only the concept of parameter currently. The requires_grad attribute is used to distinguish whether the parameter in the network needs to be optimized. For example, when defining the network, the moving_mean and moving_var in BN will be defined as parameter with attribute requires_grad=False.

-mindspore.nn.Cell.untrainable_params: The function returns a list of all untrainable parameters. Parameter has an attribute name in MindSpore, names of parameters can be obtained after getting parameters by using the untrainable_params method.

Code Example

import numpy as np
from mindspore import nn

class ConvBN(nn.Cell):
  def __init__(self):
    super(ConvBN, self).__init__()
    self.conv = nn.Conv2d(3, 64, 3)
    self.bn = nn.BatchNorm2d(64)
  def construct(self, x):
    x = self.conv(x)
    x = self.bn(x)
    return x

class MyNet(nn.Cell):
  def __init__(self):
    super(MyNet, self).__init__()
    self.build_block = nn.SequentialCell(ConvBN(), nn.ReLU())
  def construct(self, x):
    return self.build_block(x)

# The following implements mindspore.nn.Cell.untrainable_params() with MindSpore.
net = MyNet()
print(type(net.untrainable_params()), "\n")
for params in net.untrainable_params():
  print("Name: ", params.name)
  print("params: ", params)

# Out:
<class 'list'>

Name:  build_block.0.bn.moving_mean
params:  Parameter (name=build_block.0.bn.moving_mean, shape=(64,), dtype=Float32, requires_grad=False)
Name:  build_block.0.bn.moving_variance
params:  Parameter (name=build_block.0.bn.moving_variance, shape=(64,), dtype=Float32, requires_grad=False)

import torch.nn as nn

class ConvBN(nn.Module):
  def __init__(self):
    super(ConvBN, self).__init__()
    self.conv = nn.Conv2d(3, 64, 3)
    self.bn = nn.BatchNorm2d(64)
  def forward(self, x):
    x = self.conv(x)
    x = self.bn(x)
    return x

class MyNet(nn.Module):
  def __init__(self):
    super(MyNet, self).__init__()
    self.build_block = nn.Sequential(ConvBN(), nn.ReLU())
  def construct(self, x):
    return self.build_block(x)

# The following implements torch.nn.Module.buffers() with torch.
net = MyNet()
print(type(net.buffers()), "\n")
for name, params in net.named_buffers():
  print("Name: ", name)
  print("params: ", params.size())

# Out:
<class 'generator'>

Name:  build_block.0.bn.running_mean
params:  torch.Size([64])
Name:  build_block.0.bn.running_var
params:  torch.Size([64])
Name:  build_block.0.bn.num_batches_tracked
params:  torch.Size([])