liveBook · Manning

Functional API

As explained in Deep Learning with PyTorch, Second Edition

PyTorch's stateless functional namespace providing operations that do not store learnable module state.

PyTorch’s functional API provides stateless operations: each output depends entirely on the input arguments. Unlike an nn.Module, a functional call does not store learnable parameters internally. section 8.3.3

Use modules for parameterized layers such as nn.Conv2d and nn.Linear, so their Parameters are registered and managed during training. Use functional operations for parameter-free pooling and activations. section 8.3.3

The functional and module-based styles differ in how they handle state:

  • Call a module as layer(input); call a functional operation as F.function(input, other_args). section 8.3.3
  • Store module parameters as attributes; pass functional parameters explicitly, such as F.linear(input, weight, bias=None). section 8.3.3
  • Let modules register and track parameters automatically; manage functional parameters manually. section 8.3.3

Mixing modules and functional operations

import torch.nn.functional as F class Net(nn.Module): def __init__(self): super().__init__() self.conv1 = nn.Conv2d(3, 16, 3, padding=1) self.fc = nn.Linear(8 * 8 * 8, 2) def forward(self, x): out = F.max_pool2d(torch.tanh(self.conv1(x)), 2) out = out.view(-1, 8 * 8 * 8) return self.fc(out) model = Net() model(img.unsqueeze(0)) section 8.3.3

Table 8.1 Comparing module-based and functional APIs in PyTorch
Module-based ( nn.Module )
Functional ( torch.nn.functional )
Example: nn.Linear(in_features, out_features)
Example: F.linear(input, weight, bias=None)
Has internal state (parameters stored in the module)
No internal state (stateless functions)
Parameters are stored as attributes.
Parameters are passed as arguments.
Parameters are registered and tracked automatically.
Parameters must be managed manually.
Common for layers like nn.Conv2d, nn.Linear
Common for activations like F.relu, F.tanh
Called via layer(input)
Called via F.function(input, other_args)

Compare module-based and functional calls, including state, parameter storage, registration, and typical uses. section 8.3.3

How Deep Learning with PyTorch, Second Edition covers it

How deep it goes
Explains stateless calls, explicit parameters, and module-selection rationale, but stops short of implementation internals.
Examples
Demonstrates F.max_pool2d and torch.tanh inside a convolutional Net while retaining nn.Conv2d and nn.Linear modules.
What it assumes
Assumes basic familiarity with nn.Module, forward methods, convolutional layers, and learnable parameters.
What you can do after
Applies torch.nn.functional to write concise stateless operations while preserving automatic parameter management where needed.
How it connects
Connects torch.nn.functional with nn.Module, Parameters, submodules, pooling, activations, and sequential network construction.
Cover of Deep Learning with PyTorch, Second Edition
liveBookmanning.com
Read Deep Learning with PyTorch, Second Edition on liveBook Get the book at manning.com

Related concepts

This explainer was generated by Manning's knowledge graph from the text of Deep Learning with PyTorch, Second Edition. The section links point to the parts of the book it draws on.