evo_transformer.py

# evo_transformer.py

import random
import json

class EvoTransformer:
    def __init__(self, config=None):
        self.config = config or {
            "layers": 4,
            "attention_heads": 4,
            "ffn_dim": 1024,
            "dropout": 0.1,
            "memory": False,
        }
        self.history = [self.config.copy()]

    def mutate(self):
        new_config = self.config.copy()
        trait = random.choice(list(new_config.keys()))

        if trait == "layers":
            new_config[trait] = max(1, new_config[trait] + random.choice([-1, 1]))
        elif trait == "attention_heads":
            new_config[trait] = random.choice([2, 4, 6, 8])
        elif trait == "ffn_dim":
            new_config[trait] = random.choice([512, 1024, 2048])
        elif trait == "dropout":
            change = round(random.uniform(-0.05, 0.05), 2)
            new_config[trait] = round(min(max(0.0, new_config[trait] + change), 0.5), 2)
        elif trait == "memory":
            new_config[trait] = not new_config[trait]

        self.config = new_config
        self.history.append(new_config.copy())

    def evolve(self, generations=3):
        for _ in range(generations):
            self.mutate()

    def get_history(self):
        return self.history

    def evaluate(self):
        # Simulated accuracy and parameter count
        accuracy = round(0.8 + random.uniform(0.01, 0.15), 4)
        return {
            "accuracy": accuracy,
            "params": self.estimate_params()
        }

    def estimate_params(self):
        # Lightweight param estimate formula (not real count)
        multiplier = 1.0 + (0.5 if self.config["memory"] else 0.0)
        return int(self.config["layers"] * self.config["ffn_dim"] * multiplier)

    def export_csv(self):
        import pandas as pd
        df = pd.DataFrame(self.history)
        return df.to_csv(index=False)

    def export_json(self):
        return json.dumps(self.history, indent=2)