#!/usr/bin/env python3
"""
META-LEARNING - CAREFUL IMPROVEMENTS
Starting from 87% baseline, making minimal changes
"""
import torch
import torch.nn as nn
import torch.nn.functional as F
import numpy as np
from copy import deepcopy

class MetaNet100(nn.Module):
    """Simple but effective architecture"""
    def __init__(self):
        super().__init__()
        # Keep it simple - what worked before + small boost
        self.net = nn.Sequential(
            nn.Linear(10, 64),
            nn.ReLU(),
            nn.Linear(64, 64),
            nn.ReLU(),
            nn.Linear(64, 5)
        )
    
    def forward(self, x):
        return self.net(x)

class MetaLearner100:
    def __init__(self):
        self.device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
        self.model = MetaNet100().to(self.device)
        print(f"🔥 Device: {self.device}")
    
    def create_task(self):
        """Keep tasks simple and consistent"""
        # Stick with linear tasks - they work!
        w = np.random.randn(10, 5) * 1.5
        b = np.random.randn(5)
        X = np.random.randn(120, 10)
        Y = (X @ w + b).argmax(axis=1)
        
        X = torch.FloatTensor(X).to(self.device)
        Y = torch.LongTensor(Y).to(self.device)
        return X, Y
    
    def meta_train(self, n_epochs=600):
        """Train carefully with proven approach"""
        print(f"META-TRAINING: {n_epochs} epochs")
        
        meta_opt = torch.optim.Adam(self.model.parameters(), lr=0.001)
        scheduler = torch.optim.lr_scheduler.StepLR(meta_opt, step_size=200, gamma=0.5)
        
        for epoch in range(n_epochs):
            epoch_loss = 0
            
            for _ in range(10):
                data, labels = self.create_task()
                support_x, support_y = data[:80], labels[:80]  # Slightly more support
                query_x, query_y = data[80:], labels[80:]
                
                adapted = deepcopy(self.model)
                inner_opt = torch.optim.SGD(adapted.parameters(), lr=0.01, momentum=0.9)
                
                # Inner loop - moderate adaptation
                for _ in range(10):
                    pred = adapted(support_x)
                    loss = F.cross_entropy(pred, support_y)
                    inner_opt.zero_grad()
                    loss.backward()
                    inner_opt.step()
                
                pred_query = adapted(query_x)
                meta_loss = F.cross_entropy(pred_query, query_y)
                epoch_loss += meta_loss
            
            meta_opt.zero_grad()
            epoch_loss.backward()
            torch.nn.utils.clip_grad_norm_(self.model.parameters(), 1.0)
            meta_opt.step()
            scheduler.step()
            
            if (epoch + 1) % 100 == 0:
                print(f"Epoch {epoch+1}: Loss: {epoch_loss.item():.4f}")
        
        print("✅ Training complete")
    
    def test_adaptation(self, n_tests=20):
        """Test with good adaptation"""
        print(f"\nTesting on {n_tests} tasks...")
        accs = []
        
        for i in range(n_tests):
            data, labels = self.create_task()
            support_x, support_y = data[:80], labels[:80]
            query_x, query_y = data[80:], labels[80:]
            
            adapted = MetaNet100().to(self.device)
            adapted.load_state_dict(self.model.state_dict())
            
            # Adapt carefully
            opt = torch.optim.SGD(adapted.parameters(), lr=0.01, momentum=0.9)
            for _ in range(30):  # More steps for better adaptation
                pred = adapted(support_x)
                loss = F.cross_entropy(pred, support_y)
                opt.zero_grad()
                loss.backward()
                opt.step()
            
            with torch.no_grad():
                pred = adapted(query_x)
                acc = (pred.argmax(dim=1) == query_y).float().mean().item()
                accs.append(acc)
                
                status = "✅" if acc >= 0.90 else "⚠️" if acc >= 0.80 else "❌"
                print(f"  {status} Task {i+1}: {acc*100:.1f}%")
        
        return accs

def main():
    print("\n" + "="*70)
    print("🎯 META-LEARNING: CAREFUL PUSH TO 95%+")
    print("="*70)
    
    ml = MetaLearner100()
    ml.meta_train(n_epochs=600)
    
    accs = ml.test_adaptation(n_tests=20)
    
    avg = np.mean(accs)
    std = np.std(accs)
    
    print("\n" + "="*70)
    print(f"Average:  {avg*100:.1f}% (±{std*100:.1f}%)")
    print(f"Min:      {min(accs)*100:.1f}%")
    print(f"Max:      {max(accs)*100:.1f}%")
    print(f"Tasks ≥90%: {sum(1 for a in accs if a >= 0.90)}/20")
    
    if avg >= 0.95:
        print("🏆 PERFECT!")
    elif avg >= 0.90:
        print("✅ EXCELLENT!")
    elif avg >= 0.87:
        print("✅ IMPROVED!")
    print("="*70)

if __name__ == "__main__":
    main()