Quick Reference

The Quick Reference collects the PyTorch patterns, tensor conventions, formulas, and workflow rules that recur throughout the course. Use this section when you know roughly what you need to do but do not remember the exact syntax, tensor shape, formula, or workflow pattern. For conceptual explanations and complete examples, return to the lessons or the tutorials.

Important

The Quick Reference is not a tutorial. It is a concise summary of key concepts for experienced users. If you are new to deep learning with PyTorch, start with the lessons or the tutorials.

Find What You Need

I need to…	Go to
inspect, reshape, or move a tensor	Tensors & Devices
create a dataset or mini-batches	Data Management
define a model or inspect its parameters	Models
calculate a loss or update parameters	Training
evaluate a model or restore the best checkpoint	Evaluation
diagnose poor learning or design a controlled comparison	Diagnostics
calculate CNN shapes or convolution parameters	Convolutional Networks
use batch normalization, residual connections, or global pooling	Residual Networks
use or fine-tune a pretrained model	Transfer Learning

---

Common Tensor Shapes

Object	Typical shape
Tabular example	(D,)
Tabular batch	(N, D)
Grayscale image	(1, H, W)
RGB image	(3, H, W)
Image batch	(N, C, H, W)
Multiclass logits	(N, K)
Multiclass targets	(N,)
Binary logits	(N,) or (N, 1)
Binary targets	(N,)
Regression predictions	(N,) or (N, T)

Notation:

• N = batch size;

• D = feature dimension;

• K = number of classes;

• T = number of regression targets;

• C = number of channels;

• H = height;

• W = width;

Training, Validation, Test

Data is typically split into three partitions, each with a different role in the development workflow.

Partition	May influence
Training	Model parameters and fitted preprocessing
Validation	Model selection and development decisions
Test	Final reported evaluation only

Training and Evaluation

Training updates the model parameters. Evaluation measures a fixed model without parameter updates.

Training	Evaluation
model.train()	model.eval()
gradients recorded for backpropagation	gradient recording disabled with torch.inference_mode()
backward pass computes gradients	no backward pass
optimizer updates parameters	parameters remain fixed

PyTorch Imports

import torch
from torch import nn
import torch.nn.functional as F
from torch.utils.data import DataLoader, Dataset, Subset, random_split

from torchvision import datasets
import torchvision.transforms.v2 as v2