# print("Angular RMSE after training:", final_eval["out_p_filt_angular_rmse"]) print("Generating visualization set") vis_set_fname = 'saved_params/policy_viz_set.npz' if not os.path.exists(vis_set_fname): vis_input, vis_output, batch_size = create_policy_vis_set( data=data, args=args, n_mazes=args.n_mazes, encoding_func=encoding_func, # maze_indices=[2, 4, 5, 6], maze_indices=[1, 0, 5, 7], goal_indices=[ 0, ], # TEMP: for debugging # maze_indices=[0, ], # goal_indices=[2, 3, 4, 5, 6, 7, 8], # x_offset=0.25, # y_offset=0.25, ) np.savez(vis_set_fname, vis_input=vis_input, vis_output=vis_output, batch_size=batch_size) else: data = np.load(vis_set_fname) vis_input = data['vis_input']
) final_eval = sim.evaluate(test_input, {out_p_filt: test_output}, verbose=0) print("Loss after training:", final_eval["loss"]) print("Angular RMSE after training:", final_eval["out_p_filt_angular_rmse"]) if args.plot_vis_set: print("Generating visualization set") vis_input, vis_output, batch_size = create_policy_vis_set( data=data, args=args, n_mazes=args.n_mazes, encoding_func=encoding_func, maze_indices=[0, 1, 2, 3], goal_indices=[0, 1], ) vis_input = np.tile(vis_input[:, None, :], (1, n_steps, 1)) # vis_output = np.tile(vis_output[:, None, :], (1, n_steps, 1)) print("Running visualization") # viz_eval = sim.evaluate(test_input, {out_p_filt: test_output}, verbose=0) n_batches = 4 * 2 for bi in range(n_batches): viz_eval = sim.predict(vis_input[bi * batch_size:(bi + 1) * batch_size])