""" MAML Meta-Training - Learn the optimal initialization """ import torch import torch.nn as nn import numpy as np from maml import MAML, MAMLModel, create_classification_task import matplotlib.pyplot as plt class MAMLTrainer: """Train MAML on a distribution of tasks""" def __init__( self, input_size=10, hidden_sizes=[40, 40], output_size=2, inner_lr=0.01, outer_lr=0.001, inner_steps=5 ): self.model = MAMLModel(input_size, hidden_sizes, output_size) self.maml = MAML( self.model, inner_lr=inner_lr, outer_lr=outer_lr, inner_steps=inner_steps ) self.train_history = [] def generate_task_batch(self, n_tasks, task_offset=0): """Generate batch of tasks for meta-training""" task_batch = [] for i in range(n_tasks): task_id = task_offset + i support_x, support_y, query_x, query_y = create_classification_task( task_id, n_samples=40 ) task_batch.append((support_x, support_y, query_x, query_y)) return task_batch def meta_train(self, n_iterations=1000, tasks_per_batch=4, eval_every=100): """ Meta-train MAML on distribution of tasks """ print("\n" + "="*70) print("META-TRAINING MAML") print(f"Iterations: {n_iterations} | Tasks per batch: {tasks_per_batch}") print("="*70) for iteration in range(n_iterations): # Generate batch of tasks task_offset = iteration * tasks_per_batch task_batch = self.generate_task_batch(tasks_per_batch, task_offset) # Meta-training step loss, acc = self.maml.meta_train_step(task_batch) self.train_history.append({ 'iteration': iteration, 'loss': loss, 'accuracy': acc }) # Periodic evaluation if (iteration + 1) % eval_every == 0: print(f"Iteration {iteration+1}/{n_iterations} - " f"Loss: {loss:.4f} - Acc: {acc*100:.2f}%") # Evaluate on new tasks eval_acc = self.evaluate_on_new_tasks(n_tasks=20) print(f" Eval on new tasks: {eval_acc*100:.2f}%") print("\nāœ… Meta-training complete!") def evaluate_on_new_tasks(self, n_tasks=20): """Evaluate on completely new tasks (not seen during training)""" total_acc = 0 for i in range(n_tasks): # Use high task_id to ensure these are new tasks task_id = 10000 + i support_x, support_y, query_x, query_y = create_classification_task( task_id, n_samples=40 ) _, acc_after = self.maml.evaluate_task( support_x, support_y, query_x, query_y ) total_acc += acc_after return total_acc / n_tasks def compare_with_baseline(self, n_test_tasks=50): """ Compare MAML vs baseline (no meta-learning) """ print("\n" + "="*70) print("COMPARING MAML VS BASELINE") print("="*70) maml_before = [] maml_after = [] baseline_after = [] for i in range(n_test_tasks): task_id = 20000 + i support_x, support_y, query_x, query_y = create_classification_task( task_id, n_samples=40 ) # MAML performance acc_before, acc_after = self.maml.evaluate_task( support_x, support_y, query_x, query_y ) maml_before.append(acc_before) maml_after.append(acc_after) # Baseline: random init + fine-tune baseline_model = MAMLModel(10, [40, 40], 2) baseline_maml = MAML(baseline_model, inner_lr=0.01, inner_steps=5) _, baseline_acc = baseline_maml.evaluate_task( support_x, support_y, query_x, query_y ) baseline_after.append(baseline_acc) # Compute statistics maml_before_avg = np.mean(maml_before) * 100 maml_after_avg = np.mean(maml_after) * 100 baseline_after_avg = np.mean(baseline_after) * 100 improvement = maml_after_avg - maml_before_avg vs_baseline = maml_after_avg - baseline_after_avg print(f"\nResults on {n_test_tasks} new tasks:") print(f"{'='*70}") print(f"MAML (before adaptation): {maml_before_avg:.2f}%") print(f"MAML (after 5 steps): {maml_after_avg:.2f}% (+{improvement:.2f}%)") print(f"Baseline (after 5 steps): {baseline_after_avg:.2f}%") print(f"{'='*70}") print(f"MAML vs Baseline: +{vs_baseline:.2f}%") # Check if goal achieved if maml_after_avg >= 85 and improvement >= 30: print(f"\nšŸŽÆ GOAL ACHIEVED!") print(f" āœ… >85% accuracy after 5 steps") print(f" āœ… >30% improvement from adaptation") else: print(f"\nāš ļø Goal not yet met") print(f" Need: >85% accuracy, >30% improvement") return { 'maml_before': maml_before_avg, 'maml_after': maml_after_avg, 'baseline': baseline_after_avg, 'improvement': improvement } def visualize_training(self): """Plot training progress""" iterations = [h['iteration'] for h in self.train_history] accuracies = [h['accuracy'] * 100 for h in self.train_history] losses = [h['loss'] for h in self.train_history] fig, (ax1, ax2) = plt.subplots(1, 2, figsize=(14, 5)) # Accuracy plot ax1.plot(iterations, accuracies, color='blue', linewidth=2) ax1.set_xlabel('Meta-Training Iteration') ax1.set_ylabel('Query Set Accuracy (%)') ax1.set_title('MAML: Meta-Training Progress (Accuracy)') ax1.grid(True, alpha=0.3) ax1.axhline(y=85, color='green', linestyle='--', alpha=0.5, label='Goal: 85%') ax1.legend() # Loss plot ax2.plot(iterations, losses, color='red', linewidth=2) ax2.set_xlabel('Meta-Training Iteration') ax2.set_ylabel('Query Set Loss') ax2.set_title('MAML: Meta-Training Progress (Loss)') ax2.grid(True, alpha=0.3) plt.tight_layout() plt.savefig('real_capabilities/learning/meta/maml_training.png', dpi=150) print(f"\nšŸ“Š Training plot saved to: real_capabilities/learning/meta/maml_training.png") if __name__ == "__main__": print("\n" + "="*70) print("MAML META-LEARNING: FULL TRAINING & EVALUATION") print("="*70) # Create trainer trainer = MAMLTrainer( input_size=10, hidden_sizes=[40, 40], output_size=2, inner_lr=0.01, outer_lr=0.001, inner_steps=5 ) # Meta-train trainer.meta_train(n_iterations=2000, tasks_per_batch=4, eval_every=200) # Visualize training trainer.visualize_training() # Compare with baseline results = trainer.compare_with_baseline(n_test_tasks=50) print("\n" + "="*70) print("WEEK 3 COMPLETE!") print("="*70) print("\nMeta-Learning Capability: āœ…") print(f"Fast adaptation in 5 steps: {results['maml_after']:.1f}%") print(f"Improvement: {results['improvement']:.1f}%") print("\nTotal Capabilities: 3/30 (10%)")