checkpoints

sneat/evolve.py

@@ -23,6 +23,19 @@ def evaluate_population(pop, ff):
         for g, fitness in zip(pop.genomes, fitness_scores):
             g.fitness = fitness
 
+def save_checkpoint(pop):
+    with open('checkpoint.pkl', 'wb') as f:
+        pkl.dump(pop, f)
+
+def load_checkpoint():
+    try:
+        with open('checkpoint.pkl', 'rb') as f:
+            pop = pkl.load(f)
+            print(f'[i] Restoring from checkpoint (gen. {pop.generation})...')
+            return pop
+    except FileNotFoundError:
+        return None
+
 def print_stats(pop):
     print(f'\n\n[i] Gen. {pop.generation}:')
     headers = ['Species', 'Members', 'Best Fitness', 'Average Fitness']
@@ -35,7 +48,8 @@ def print_stats(pop):
 
 def evolve(fitness_function):
     config = get_config()
-    pop = Population()
+
+    pop = load_checkpoint() or Population()
 
     max_generations = config.getint('Evolution', 'max_generations') or np.inf
     max_fitness = config.getfloat('Evolution', 'max_fitness') or np.inf
@@ -52,6 +66,10 @@ def evolve(fitness_function):
             # reproduce
             pop.reproduce()
 
+            # save checkpoint
+            if pop.generation % 10 == 0:
+                save_checkpoint(pop)
+
             best_fitness = max(g.fitness for g in pop.genomes)
             if best_fitness >= max_fitness:
                 winner = max(pop.genomes, key=lambda x: x.fitness)