train.py

import torch
import torch.nn as nn
import torch.optim as optim
from torch.utils.data import DataLoader
from torchvision import datasets, transforms
from transformers import ViTConfig
from sae.sae import ViTSAE

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print(device)
# Assuming the ViTSAE model and EarlyExitLayer classes are defined as previously provided

# Model Configuration
config = ViTConfig(
    image_size=224,
    num_hidden_layers=12,
    hidden_size=768,
    num_attention_heads=12,
    num_labels=10,
)
model = ViTSAE(config, num_classes=10, early_exit_thresholds=[0.9, 0.95, 0.99])
model.to(device)

# Dataset and DataLoader
transform = transforms.Compose(
    [
        transforms.Resize((224, 224)),
        transforms.ToTensor(),
    ]
)
train_dataset = datasets.CIFAR10(
    root="./data", train=True, download=True, transform=transform
)
train_loader = DataLoader(train_dataset, batch_size=64, shuffle=True)

# Loss Function and Optimizer
criterion = nn.CrossEntropyLoss()
optimizer = optim.Adam(model.parameters(), lr=0.001)

# Training Loop
num_epochs = 10
for epoch in range(num_epochs):
    model.train()
    running_loss = 0.0
    for images, labels in train_loader:
        images, labels = images.to(device), labels.to(device)

        optimizer.zero_grad()

        outputs, exit_point = model(images)
        loss = criterion(outputs, labels)
        loss.backward()
        optimizer.step()

        running_loss += loss.item()
        print(f"Exit point: {exit_point}")
        print(f"Loss: {loss.item()}")
    print(f"Epoch {epoch+1}/{num_epochs}, Loss: {running_loss/len(train_loader)}")

print("Finished Training")