import torch
import torch.nn as nn
import torch.optim as optim
from torchvision import datasets, transforms
from torch.utils.data import DataLoader

# Represents a single "Rational Term" in the nRAN formula [1]
class MLPRationalTerm(nn.Module):
    def __init__(self, input_dim, hidden_dim, output_dim):
        super(MLPRationalTerm, self).__init__()
        self.net = nn.Sequential(
            nn.Linear(input_dim, hidden_dim),
            nn.ReLU(),
            nn.Linear(hidden_dim, output_dim)
        )

    def forward(self, x):
        return self.net(x)

class nRANClassifier(nn.Module):
    def __init__(self, n_terms, input_dim=784, hidden_dim=100, output_dim=10):
        super(nRANClassifier, self).__init__()
        self.n_terms = n_terms
        
        # The "Base Offset" (c) from source [1]
        self.base_offset = nn.Linear(input_dim, output_dim)
        
        # The list of "Rational Correction Terms" (a_i / b_i) from source [1]
        self.rational_terms = nn.ModuleList([
            MLPRationalTerm(input_dim, hidden_dim, output_dim) for _ in range(n_terms)
        ])
        
        # Scale factors to mimic multi-scale rational expansions (1/10, 1/100...) [2]
        self.scales = [0.1 ** i for i in range(1, n_terms + 1)]

    def forward(self, x):
        x = x.view(x.size(0), -1) # Flatten MNIST images
        
        # Start with the base offset
        out = self.base_offset(x)
        
        # Perform the "Collapse" by summing all functional rational terms [1]
        for i, term in enumerate(self.rational_terms):
            out += self.scales[i] * term(x)
            
        return out

# --- Training and Testing Logic ---

def train_test_mnist():
    # Load Data
    transform = transforms.Compose([transforms.ToTensor(), transforms.Normalize((0.1307,), (0.3081,))])
    train_loader = DataLoader(datasets.MNIST('../data', train=True, download=True, transform=transform), batch_size=64, shuffle=True)
    test_loader = DataLoader(datasets.MNIST('../data', train=False, transform=transform), batch_size=1000)

    # Initialize nRAN Model with n=5 terms [3]
    model = nRANClassifier(n_terms=5)
    optimizer = optim.Adam(model.parameters(), lr=0.001)
    criterion = nn.CrossEntropyLoss()

    # Train
    model.train()
    for _ in range(100):
        for epoch in range(1, 2):
            for batch_idx, (data, target) in enumerate(train_loader):
                for i in range(10):
                    optimizer.zero_grad()
                    output = model(data)
                    loss = criterion(output, target)
                    loss.backward()
                    optimizer.step()
                    if i==0:
                        if batch_idx % 200 == 0:
                            print(f"Epoch {epoch} [{batch_idx*64}/60000] Loss: {loss.item():.6f}")

        # Test
        model.eval()
        test_loss = 0
        correct = 0
        with torch.no_grad():
            for data, target in test_loader:
                output = model(data)
                test_loss += criterion(output, target).item()
                pred = output.argmax(dim=1, keepdim=True)
                correct += pred.eq(target.view_as(pred)).sum().item()

        print(f"\nTest set: Average loss: {test_loss/len(test_loader):.4f}, "
              f"Accuracy: {correct}/{len(test_loader.dataset)} ({100. * correct / len(test_loader.dataset)}%)")

if __name__ == "__main__":
    train_test_mnist()