Deep Learning with PyTorch
Train neural networks with PyTorch — tensors, autograd, nn.Module, DataLoader, GPU training, transfer learning, and ONNX export.
Part 1: What You Will Learn
- Create PyTorch tensors and move them to the available CPU or GPU device.
- Define a neural network with
nn.Module. - Train the network with a loss function, backpropagation, and an optimiser.
- Use the trained model to make a prediction.
Part 2: Key Concepts
A neural network contains layers of learnable parameters. During training, PyTorch calculates a loss, uses automatic differentiation to compute gradients, and lets an optimiser update the parameters.
- Tensor: PyTorch's multidimensional numerical object.
- Forward pass: calculate predictions.
- Loss: measure prediction error.
- Backward pass: calculate gradients with autograd.
- Optimiser: update weights using those gradients.
Part 3: Topic-Specific Code Example
import torch
from torch import nn
device = torch.device(
"cuda" if torch.cuda.is_available() else "cpu"
)
print("Using:", device)
X = torch.tensor(
[[1.0], [2.0], [3.0], [4.0], [5.0], [6.0]],
device=device,
)
y = torch.tensor(
[[3.0], [5.0], [7.0], [9.0], [11.0], [13.0]],
device=device,
)
class SimpleRegressor(nn.Module):
def __init__(self) -> None:
super().__init__()
self.network = nn.Sequential(
nn.Linear(1, 8),
nn.ReLU(),
nn.Linear(8, 1),
)
def forward(self, x: torch.Tensor) -> torch.Tensor:
return self.network(x)
model = SimpleRegressor().to(device)
loss_function = nn.MSELoss()
optimizer = torch.optim.Adam(model.parameters(), lr=0.03)
for epoch in range(500):
prediction = model(X)
loss = loss_function(prediction, y)
optimizer.zero_grad()
loss.backward()
optimizer.step()
model.eval()
with torch.no_grad():
new_x = torch.tensor([[7.0]], device=device)
predicted_y = model(new_x)
print("Prediction for x=7:", round(predicted_y.item(), 2))pip install torch
Part 4: How the Example Works
The sample relationship is approximately y = 2x + 1. The network is not given that formula; it learns an approximation from examples. loss.backward() calculates gradients and optimizer.step() changes the model parameters. torch.no_grad() is used during prediction because gradients are unnecessary.
Part 5: Hands-On Practice
Mini project — Exam Score Regressor. Train a network using study hours and attendance as two inputs and exam score as the output. Add a DataLoader when the dataset becomes larger, save the model with torch.save(), and reload it for prediction.
Part 6: Next Steps
Modify the network size and learning rate, then continue to Lesson 33 to connect Python applications to large language models.