pytorch_neural_network_fashionmnist_image_classification_quickstart.py

1import torch
2from torch import nn
3from torch.utils.data import DataLoader
4from torchvision import datasets
5from torchvision.transforms import ToTensor
6
7# Download training data from open datasets.
8training_data = datasets.FashionMNIST(
9    root="data",
10    train=True,
11    download=True,
12    transform=ToTensor(),
13)
14
15# Download test data from open datasets.
16test_data = datasets.FashionMNIST(
17    root="data",
18    train=False,
19    download=True,
20    transform=ToTensor(),
21)
22
23batch_size = 64
24
25# Create data loaders.
26train_dataloader = DataLoader(training_data, batch_size=batch_size)
27test_dataloader = DataLoader(test_data, batch_size=batch_size)
28
29for X, y in test_dataloader:
30    print(f"Shape of X [N, C, H, W]: {X.shape}")
31    print(f"Shape of y: {y.shape} {y.dtype}")
32    break
33
34# Get cpu, gpu or mps device for training.
35device = (
36    "cuda"
37    if torch.cuda.is_available()
38    else "mps"
39    if torch.backends.mps.is_available()
40    else "cpu"
41)
42print(f"Using {device} device")
43
44# Define model
45class NeuralNetwork(nn.Module):
46    def __init__(self):
47        super().__init__()
48        self.flatten = nn.Flatten()
49        self.linear_relu_stack = nn.Sequential(
50            nn.Linear(28*28, 512),
51            nn.ReLU(),
52            nn.Linear(512, 512),
53            nn.ReLU(),
54            nn.Linear(512, 10)
55        )
56
57    def forward(self, x):
58        x = self.flatten(x)
59        logits = self.linear_relu_stack(x)
60        return logits
61
62model = NeuralNetwork().to(device)
63print(model)
64
65loss_fn = nn.CrossEntropyLoss()
66optimizer = torch.optim.SGD(model.parameters(), lr=1e-3)
67
68def train(dataloader, model, loss_fn, optimizer):
69    size = len(dataloader.dataset)
70    model.train()
71    for batch, (X, y) in enumerate(dataloader):
72        X, y = X.to(device), y.to(device)
73
74        # Compute prediction error
75        pred = model(X)
76        loss = loss_fn(pred, y)
77
78        # Backpropagation
79        loss.backward()
80        optimizer.step()
81        optimizer.zero_grad()
82
83        if batch % 100 == 0:
84            loss, current = loss.item(), (batch + 1) * len(X)
85            print(f"loss: {loss:>7f}  [{current:>5d}/{size:>5d}]")
86
87def test(dataloader, model, loss_fn):
88    size = len(dataloader.dataset)
89    num_batches = len(dataloader)
90    model.eval()
91    test_loss, correct = 0, 0
92    with torch.no_grad():
93        for X, y in dataloader:
94            X, y = X.to(device), y.to(device)
95            pred = model(X)
96            test_loss += loss_fn(pred, y).item()
97            correct += (pred.argmax(1) == y).type(torch.float).sum().item()
98    test_loss /= num_batches
99    correct /= size
100    print(f"Test Error: \n Accuracy: {(100*correct):>0.1f}%, Avg loss: {test_loss:>8f} \n")
101
102epochs = 5
103for t in range(epochs):
104    print(f"Epoch {t+1}\n-------------------------------")
105    train(train_dataloader, model, loss_fn, optimizer)
106    test(test_dataloader, model, loss_fn)
107print("Done!")
108
109# Save model
110torch.save(model.state_dict(), "model.pth")
111print("Saved PyTorch Model State to model.pth")
112
113# Load model
114model = NeuralNetwork().to(device)
115model.load_state_dict(torch.load("model.pth"))
116
117classes = [
118    "T-shirt/top",
119    "Trouser",
120    "Pullover",
121    "Dress",
122    "Coat",
123    "Sandal",
124    "Shirt",
125    "Sneaker",
126    "Bag",
127    "Ankle boot",
128]
129
130model.eval()
131x, y = test_data[0][0], test_data[0][1]
132with torch.no_grad():
133    x = x.to(device)
134    pred = model(x)
135    predicted, actual = classes[pred[0].argmax(0)], classes[y]
136    print(f'Predicted: "{predicted}", Actual: "{actual}"')