#!/usr/bin/env python3
"""
ULTIMATE CONTINUAL LEARNING
Train on huge datasets (5000 samples) to match test distribution
"""
import torch
import torch.nn as nn
import torch.nn.functional as F
import numpy as np

torch.manual_seed(777)
np.random.seed(777)

device = torch.device('cuda')
print(f"Device: {device}\n")

class UltimateNet(nn.Module):
    def __init__(self, n_tasks=5):
        super().__init__()
        self.backbone = nn.Sequential(
            nn.Linear(20, 512),
            nn.ReLU(),
            nn.Dropout(0.3),
            nn.Linear(512, 512),
            nn.ReLU(),
            nn.Dropout(0.3),
            nn.Linear(512, 256),
            nn.ReLU()
        )
        
        self.heads = nn.ModuleList([
            nn.Sequential(
                nn.Linear(256, 128),
                nn.ReLU(),
                nn.Dropout(0.2),
                nn.Linear(128, 5)
            ) for _ in range(n_tasks)
        ])
    
    def forward(self, x, task_id):
        return self.heads[task_id](self.backbone(x))

def create_task(task_id, n_samples=5000):  # HUGE dataset
    X, Y = [], []
    for _ in range(n_samples):
        x = np.random.randn(20)
        if task_id == 0:
            y = 0 if x[:10].sum() > 0 else 1
        elif task_id == 1:
            y = 2 if x[10:].sum() > 0 else 3
        elif task_id == 2:
            y = 4 if x.std() > 1.2 else 0
        elif task_id == 3:
            y = 1 if x[::2].sum() > 0 else 2
        else:
            y = 3 if x[1::2].sum() > 0 else 4
        X.append(x)
        Y.append(y)
    return torch.FloatTensor(X).to(device), torch.LongTensor(Y).to(device)

print("="*70)
print("ULTIMATE CONTINUAL LEARNING - 5K SAMPLES PER TASK")
print("="*70)

model = UltimateNet().to(device)
opt = torch.optim.Adam(model.parameters(), lr=0.001)

all_tasks = []

for task_id in range(5):
    print(f"\n{'='*70}")
    print(f"TASK {task_id} - Training on 5000 samples")
    print(f"{'='*70}")
    
    X, Y = create_task(task_id, n_samples=5000)
    all_tasks.append((X, Y, task_id))
    
    # Train extensively
    for epoch in range(200):
        # Mini-batch training
        perm = torch.randperm(len(X))
        for i in range(0, len(X), 128):
            batch_idx = perm[i:i+128]
            batch_X = X[batch_idx]
            batch_Y = Y[batch_idx]
            
            # Train current task
            pred = model(batch_X, task_id)
            loss = F.cross_entropy(pred, batch_Y)
            opt.zero_grad()
            loss.backward()
            opt.step()
        
        # Also replay previous tasks
        for X_t, Y_t, tid in all_tasks[:-1]:
            perm_t = torch.randperm(len(X_t))[:500]  # Sample 500
            pred = model(X_t[perm_t], tid)
            loss = F.cross_entropy(pred, Y_t[perm_t])
            opt.zero_grad()
            loss.backward()
            opt.step()
        
        if epoch % 40 == 0:
            accs = []
            for tid in range(task_id + 1):
                X_t, Y_t, _ = all_tasks[tid]
                with torch.no_grad():
                    # Test on first 1000 samples
                    pred = model(X_t[:1000], tid)
                    acc = (pred.argmax(1) == Y_t[:1000]).float().mean().item()
                    accs.append(acc)
            print(f"  Epoch {epoch}: Avg={np.mean(accs)*100:.1f}%, Min={min(accs)*100:.1f}%")

# Final mega-test
print("\n" + "="*70)
print("FINAL TEST (5000 samples)")
print("="*70)

final_accs = []
for tid in range(5):
    X_test, Y_test = create_task(tid, n_samples=5000)
    with torch.no_grad():
        pred = model(X_test, tid)
        acc = (pred.argmax(1) == Y_test).float().mean().item()
        final_accs.append(acc)
        status = "🎉" if acc >= 0.98 else "✅" if acc >= 0.95 else "⚠️"
        print(f"  {status} Task {tid}: {acc*100:.2f}%")

avg = np.mean(final_accs)
min_acc = np.min(final_accs)

print(f"\n{'='*70}")
print(f"Average: {avg*100:.2f}%")
print(f"Minimum: {min_acc*100:.2f}%")

if avg >= 0.97:
    print("🎉 NEAR-PERFECT!")
elif avg >= 0.95:
    print("✅ EXCELLENT!")
else:
    print("✅ Very strong!")

torch.save(model.state_dict(), 'continual_ultimate.pth')
print("💾 Saved!")