sm.py

import pickle as pkl
import os
import pandas as pd
import numpy as np
import json
import copy
from math import isnan
import matplotlib as mlp
from script import *

mlp.use("agg")
import matplotlib.pyplot as plt

font = {'size': 24}
mlp.rc('font', **font)

NAN_LABEL = -1


def get_metrics(duration_list, queue_length_list, min_duration, min_duration_id, min_queue_length, min_queue_length_id,
                traffic_name, total_summary, mode_name, save_path, num_rounds, min_duration2=None):
    validation_duration_length = 10
    minimum_round = 50 if num_rounds > 50 else 0
    duration_list = np.array(duration_list)
    queue_length_list = np.array(queue_length_list)

    # min_duration, min_duration_id = np.min(duration_list), np.argmin(duration_list)
    # min_queue_length, min_queue_length_id = np.min(queue_length_list), np.argmin(queue_length_list)

    nan_count = len(np.where(duration_list == NAN_LABEL)[0])
    validation_duration = duration_list[-validation_duration_length:]
    final_duration = np.round(np.mean(validation_duration[validation_duration > 0]), decimals=2)
    final_duration_std = np.round(np.std(validation_duration[validation_duration > 0]), decimals=2)

    if nan_count == 0:
        convergence = {1.2: len(duration_list) - 1, 1.1: len(duration_list) - 1}
        for j in range(minimum_round, len(duration_list)):
            for level in [1.2, 1.1]:
                if max(duration_list[j:]) <= level * final_duration:
                    if convergence[level] > j:
                        convergence[level] = j
        conv_12 = convergence[1.2]
        conv_11 = convergence[1.1]
    else:
        conv_12, conv_11 = 0, 0

    # simple plot for each training instance
    f, ax = plt.subplots(1, 1, figsize=(12, 9), dpi=100)
    ax.plot(duration_list, linewidth=2, color='k')
    ax.plot([0, len(duration_list)], [final_duration, final_duration], linewidth=2, color="g")
    ax.plot([conv_12, conv_12], [duration_list[conv_12], duration_list[conv_12] * 3], linewidth=2, color="b")
    ax.plot([conv_11, conv_11], [duration_list[conv_11], duration_list[conv_11] * 3], linewidth=2, color="b")
    ax.plot([0, len(duration_list)], [min_duration, min_duration], linewidth=2, color="r")
    ax.plot([min_duration_id, min_duration_id], [min_duration, min_duration * 3], linewidth=2, color="r")
    ax.set_title(traffic_name + "-" + str(final_duration))
    plt.savefig(save_path + "/" + traffic_name + "-" + mode_name + ".png")
    plt.close()

    total_summary["traffic_file"].append(traffic_name)
    total_summary["traffic"].append(traffic_name.split(".xml")[0])
    total_summary["min_queue_length"].append(min_queue_length)
    total_summary["min_queue_length_round"].append(min_queue_length_id)
    total_summary["min_duration"].append(min_duration)
    total_summary["min_duration_round"].append(min_duration_id)
    total_summary["final_duration"].append(final_duration)
    total_summary["final_duration_std"].append(final_duration_std)
    total_summary["convergence_1.2"].append(conv_12)
    total_summary["convergence_1.1"].append(conv_11)
    total_summary["nan_count"].append(nan_count)
    total_summary["min_duration2"].append(min_duration2)

    return total_summary


def summary_plot(traffic_performance, figure_dir, mode_name, num_rounds):
    minimum_round = 50 if num_rounds > 50 else 0
    validation_duration_length = 10
    anomaly_threshold = 1.3

    for traffic_name in traffic_performance:
        f, ax = plt.subplots(2, 1, figsize=(12, 9), dpi=100)
        performance_tmp = []
        check_list = []
        for ti in range(len(traffic_performance[traffic_name])):
            ax[0].plot(traffic_performance[traffic_name][ti][0], linewidth=2)
            validation_duration = traffic_performance[traffic_name][ti][0][-validation_duration_length:]
            final_duration = np.round(np.mean(validation_duration), decimals=2)
            if len(np.where(traffic_performance[traffic_name][ti][0] == NAN_LABEL)[0]) == 0:
                # and len(traffic_performance[traffic_name][ti][0]) == num_rounds:
                tmp = traffic_performance[traffic_name][ti][0]
                if len(tmp) < num_rounds:
                    tmp.extend([float("nan")] * (num_rounds - len(traffic_performance[traffic_name][ti][0])))
                performance_tmp.append(tmp)
                check_list.append(final_duration)
            else:
                print("the length of traffic {} is shorter than {}".format(traffic_name, num_rounds))
        check_list = np.array(check_list)
        for ci in np.where(check_list > anomaly_threshold * np.mean(check_list))[0][::-1]:
            del performance_tmp[ci]
            print("anomaly traffic_name:{} id:{} err:{}".format(traffic_name, ci, check_list[ci] - np.mean(check_list)))
        if len(performance_tmp) == 0:
            print("The result of {} is not enough for analysis.".format(traffic_name))
            continue
        try:
            performance_summary = np.array(performance_tmp)
            print(traffic_name, performance_summary.shape)
            ax[1].errorbar(x=range(len(traffic_performance[traffic_name][0][0])),
                           y=np.mean(performance_summary, axis=0),
                           yerr=np.std(performance_summary, axis=0))

            psm = np.mean(performance_summary, axis=0)
            validation_duration = psm[-validation_duration_length:]
            final_duration = np.round(np.mean(validation_duration), decimals=2)

            convergence = {1.2: len(psm) - 1, 1.1: len(psm) - 1}
            for j in range(minimum_round, len(psm)):
                for level in [1.2, 1.1]:
                    if max(psm[j:]) <= level * final_duration:
                        if convergence[level] > j:
                            convergence[level] = j
            ax[1].plot([0, len(psm)], [final_duration, final_duration], linewidth=2, color="g")
            ax[1].text(len(psm), final_duration * 2, "final-" + str(final_duration))
            ax[1].plot([convergence[1.2], convergence[1.2]], [psm[convergence[1.2]], psm[convergence[1.2]] * 3],
                       linewidth=2, color="b")
            ax[1].text(convergence[1.2], psm[convergence[1.2]] * 2, "conv 1.2-" + str(convergence[1.2]))
            ax[1].plot([convergence[1.1], convergence[1.1]], [psm[convergence[1.1]], psm[convergence[1.1]] * 3],
                       linewidth=2, color="b")
            ax[1].text(convergence[1.1], psm[convergence[1.1]] * 2, "conv 1.1-" + str(convergence[1.1]))
            ax[1].set_title(traffic_name)
            plt.subplots_adjust(left=None, bottom=None, right=None, top=None, wspace=None, hspace=0.3)
            plt.savefig(figure_dir + "/" + traffic_name + "-" + mode_name + ".png")
            plt.close()
        except:
            print("plot error")


def plot_segment_duration(round_summary, path, mode_name):
    save_path = os.path.join(path, "segments")
    if not os.path.exists(save_path):
        os.makedirs(save_path)
    for key in round_summary.keys():
        if "duration" in key:
            f, ax = plt.subplots(1, 1, figsize=(12, 9), dpi=100)
            ax.plot(round_summary[key], linewidth=2, color='k')
            ax.set_title(key)
            plt.savefig(save_path + "/" + key + "-" + mode_name + ".png")
            plt.close()


def padding_duration(performance_duration):
    for traffic_name in performance_duration.keys():
        max_duration_length = max([len(x[0]) for x in performance_duration[traffic_name]])
        for i, ti in enumerate(performance_duration[traffic_name]):
            performance_duration[traffic_name][i][0].extend((max_duration_length - len(ti[0]))*[ti[0][-1]])

    return performance_duration


def performance_at_min_duration_round_plot(performance_at_min_duration_round, figure_dir, mode_name):
    for traffic_name in performance_at_min_duration_round:
        f, ax = plt.subplots(1, 1, figsize=(12, 9), dpi=100)
        for ti in range(len(performance_at_min_duration_round[traffic_name])):
            ax.plot(performance_at_min_duration_round[traffic_name][ti][0], linewidth=2)
        plt.savefig(figure_dir + "/" + "min_duration_round" + "-" + traffic_name + "_" + mode_name + ".png")
        plt.close()


def summary_detail_train(memo, total_summary):
    # each_round_train_duration

    performance_duration = {}
    performance_at_min_duration_round = {}
    records_dir = os.path.join("records", memo)
    for traffic_file in os.listdir(records_dir):
        if ".xml" not in traffic_file and "json" not in traffic_file:
            continue
        print(traffic_file)

        min_queue_length = min_duration = float('inf')
        min_queue_length_id = min_duration_ind = 0

        # get run_counts to calculate the queue_length each second
        exp_conf = open(os.path.join(records_dir, traffic_file, "exp.conf"), 'r')
        dic_exp_conf = json.load(exp_conf)
        run_counts = dic_exp_conf["RUN_COUNTS"]
        num_rounds = dic_exp_conf["NUM_ROUNDS"]
        num_seg = run_counts // 3600

        traffic_vol = get_total_traffic_volume(dic_exp_conf["TRAFFIC_FILE"][0])
        nan_thres = 120

        duration_each_round_list = []
        queue_length_each_round_list = []

        train_round_dir = os.path.join(records_dir, traffic_file, "train_round")
        round_files = os.listdir(train_round_dir)
        round_files = [f for f in round_files if "round" in f]
        round_files.sort(key=lambda x: int(x[6:]))
        round_summary = {"round": list(range(num_rounds))}
        for round in round_files:
            try:
                round_dir = os.path.join(train_round_dir, round)

                duration_gens = 0
                queue_length_gens = 0
                cnt_gen = 0

                list_duration_seg = [float('inf')] * num_seg
                list_queue_length_seg = [float('inf')] * num_seg
                list_queue_length_id_seg = [0] * num_seg
                list_duration_id_seg = [0] * num_seg
                for gen in os.listdir(round_dir):
                    if "generator" not in gen:
                        continue

                    # summary items (queue_length) from pickle
                    gen_dir = os.path.join(records_dir, traffic_file, "train_round", round, gen)
                    f = open(os.path.join(gen_dir, "inter_0.pkl"), "rb")
                    samples = pkl.load(f)

                    for sample in samples:
                        queue_length_gens += sum(sample['state']['lane_queue_length'])
                    sample_num = len(samples)
                    f.close()

                    # summary items (duration) from csv
                    df_vehicle_inter_0 = pd.read_csv(os.path.join(round_dir, gen, "vehicle_inter_0.csv"),
                                                     sep=',', header=0, dtype={0: str, 1: float, 2: float},
                                                     names=["vehicle_id", "enter_time", "leave_time"])

                    duration = df_vehicle_inter_0["leave_time"].values - df_vehicle_inter_0["enter_time"].values
                    ave_duration = np.mean([time for time in duration if not isnan(time)])
                    real_traffic_vol = 0
                    nan_num = 0
                    for time in duration:
                        if not isnan(time):
                            real_traffic_vol += 1
                        else:
                            nan_num += 1
                    # print(ave_duration)

                    cnt_gen += 1
                    duration_gens += ave_duration

                    for i, interval in enumerate(range(0, run_counts, 3600)):
                        did = np.bitwise_and(df_vehicle_inter_0["enter_time"].values < interval + 3600,
                                             df_vehicle_inter_0["enter_time"].values > interval)
                        # vehicle_in_seg = sum([int(x) for x in (df_vehicle_inter_0["enter_time"][did].values > 0)])
                        # vehicle_out_seg = sum([int(x) for x in (df_vehicle_inter_0["leave_time"][did].values > 0)])
                        duration_seg = df_vehicle_inter_0["leave_time"][did].values - df_vehicle_inter_0["enter_time"][
                            did].values
                        ave_duration_seg = np.mean([time for time in duration_seg if not isnan(time)])
                        # print(traffic_file, round, i, ave_duration)
                        real_traffic_vol_seg = 0
                        nan_num_seg = 0
                        for time in duration_seg:
                            if not isnan(time):
                                real_traffic_vol_seg += 1
                            else:
                                nan_num_seg += 1

                        # print(real_traffic_vol, traffic_vol, traffic_vol - real_traffic_vol, nan_num)

                        if nan_num_seg < nan_thres:
                            # if min_duration[i] > ave_duration and ave_duration > 24:
                            list_duration_seg[i] = ave_duration_seg
                            list_duration_id_seg[i] = int(round[6:])


                list_duration_seg = np.array(list_duration_seg)/cnt_gen
                for j in range(num_seg):
                    key = "min_duration-" + str(j)
                    if key not in round_summary.keys():
                        round_summary[key] = [list_duration_seg[j]]
                    else:
                        round_summary[key].append(list_duration_seg[j])


                duration_each_round_list.append(duration_gens / cnt_gen)
                queue_length_each_round_list.append(queue_length_gens / cnt_gen / sample_num)

                # print(real_traffic_vol, traffic_vol, traffic_vol - real_traffic_vol, nan_num)
                if min_queue_length > queue_length_gens / cnt_gen / sample_num:
                    min_queue_length = queue_length_gens / cnt_gen / sample_num
                    min_queue_length_id = int(round[6:])

                valid_flag = json.load(open(os.path.join(gen_dir, "valid_flag.json")))
                if valid_flag['0']: # temporary for one intersection
                    if min_duration > duration_gens / cnt_gen:
                        min_duration = duration_gens / cnt_gen
                        min_duration_ind = int(round[6:])
                #print(nan_num, nan_thres)

            except:
                # change anomaly label from nan to -1000 for the convenience of following computation
                duration_each_round_list.append(NAN_LABEL)
                queue_length_each_round_list.append(NAN_LABEL)

        result_dir = os.path.join("summary", memo, traffic_file)
        if not os.path.exists(result_dir):
            os.makedirs(result_dir)

        _res = {
            "duration": duration_each_round_list,
            "queue_length": queue_length_each_round_list
        }
        result = pd.DataFrame(_res)
        result.to_csv(os.path.join(result_dir, "train_results.csv"))
        if num_seg > 1:
            round_result = pd.DataFrame(round_summary)
            round_result.to_csv(os.path.join(result_dir, "train_seg_results.csv"), index=False)
            # plot duration segment
            plot_segment_duration(round_summary, result_dir, mode_name="train")
            duration_each_segment_list = round_result.iloc[min_duration_ind][1:].values
            if ".xml" in traffic_file:
                traffic_name, traffic_time = traffic_file.split(".xml")
            elif ".json" in traffic_file:
                traffic_name, traffic_time = traffic_file.split(".json")
            if traffic_name not in performance_at_min_duration_round:
                performance_at_min_duration_round[traffic_name] = [(duration_each_segment_list, traffic_time)]
            else:
                performance_at_min_duration_round[traffic_name].append((duration_each_segment_list, traffic_time))



        # total_summary
        total_summary = get_metrics(duration_each_round_list, queue_length_each_round_list,
                                    min_duration, min_duration_ind, min_queue_length, min_queue_length_id,
                                    traffic_file, total_summary,
                                    mode_name="train", save_path=result_dir, num_rounds=num_rounds)

        if ".xml" in traffic_file:
            traffic_name, traffic_time = traffic_file.split(".xml")
        elif ".json" in traffic_file:
            traffic_name, traffic_time = traffic_file.split(".json")
        if traffic_name not in performance_duration:
            performance_duration[traffic_name] = [(duration_each_round_list, traffic_time)]
        else:
            performance_duration[traffic_name].append((duration_each_round_list, traffic_time))

    figure_dir = os.path.join("summary", memo, "figures")
    if not os.path.exists(figure_dir):
        os.makedirs(figure_dir)
    if dic_exp_conf["EARLY_STOP"]:
        performance_duration = padding_duration(performance_duration)
    summary_plot(performance_duration, figure_dir, mode_name="train", num_rounds=num_rounds)
    total_result = pd.DataFrame(total_summary)
    total_result.to_csv(os.path.join("summary", memo, "total_train_results.csv"))
    performance_at_min_duration_round_plot(performance_at_min_duration_round, figure_dir, mode_name="train")


def summary_detail_test(memo, total_summary):
    # each_round_train_duration

    performance_duration = {}
    performance_at_min_duration_round = {}
    records_dir = os.path.join("records", memo)
    for traffic_file in os.listdir(records_dir):
        if ".xml" not in traffic_file and ".json" not in traffic_file:
            continue

        if "48034" in traffic_file:
           continue
        print(traffic_file)

        min_queue_length = min_duration = min_duration2 = float('inf')
        min_queue_length_id = min_duration_ind = 0

        # get run_counts to calculate the queue_length each second
        # exp_conf = open(os.path.join(records_dir, traffic_file, "exp.conf"), 'r')
        # dic_exp_conf = json.load(exp_conf)
        # run_counts = dic_exp_conf["RUN_COUNTS"]
        # num_rounds = dic_exp_conf["NUM_ROUNDS"]
        run_counts = 3600
        num_rounds = 500
        num_seg = run_counts//3600

        # traffic_vol = get_total_traffic_volume(dic_exp_conf["TRAFFIC_FILE"][0])
        nan_thres = 120

        duration_each_round_list = []
        duration_each_round_list2 = []
        queue_length_each_round_list = []
        num_of_vehicle_in = []
        num_of_vehicle_out = []

        train_round_dir = os.path.join(records_dir, traffic_file, "train_round")
        try:
            round_files = os.listdir(train_round_dir)
        except:
            print("no test round in {}".format(traffic_file))
            continue
        round_files = [f for f in round_files if "round" in f]
        # round_files.sort(key=lambda x: int(x[6:]))
        round_summary = {"round": list(range(num_rounds))}
        for round in range(500):

            try:
                # round_dir = os.path.join(train_round_dir, round)

                list_duration_seg = [float('inf')] * num_seg
                list_queue_length_seg = [float('inf')] * num_seg
                list_queue_length_id_seg = [0] * num_seg
                list_duration_id_seg = [0] * num_seg

                # summary items (queue_length) from pickle
                # f = open(os.path.join(round_dir, "inter_0.pkl"), "rb")
                # samples = pkl.load(f)
                # queue_length_each_round = 0
                # for sample in samples:
                #     queue_length_each_round += sum(sample['state']['lane_queue_length'])
                # sample_num = len(samples)
                # f.close()

                # summary items (duration) from csv
                df_vehicle_inter_0 = pd.read_csv(os.path.join(train_round_dir, "vehicle_inter_0_round_{}.csv".format(round)),
                                                 sep=',', header=0, dtype={0: str, 1: float, 2: float},
                                                 names=["vehicle_id", "enter_time", "leave_time"])

                vehicle_in = sum([int(x) for x in (df_vehicle_inter_0["enter_time"].values > 0)])
                vehicle_out = sum([int(x) for x in (df_vehicle_inter_0["leave_time"].values > 0)])
                total_vol = get_total_traffic_volume(traffic_file)
                duration = df_vehicle_inter_0["leave_time"].values - df_vehicle_inter_0["enter_time"].values
                ave_duration = np.mean([time for time in duration if not isnan(time)])
                # print(ave_duration)

                if "peak" in traffic_file:
                    did1 = df_vehicle_inter_0["enter_time"].values <= run_counts / 2
                    duration = df_vehicle_inter_0["leave_time"][did1].values - df_vehicle_inter_0["enter_time"][
                        did1].values
                    ave_duration = np.mean([time for time in duration if not isnan(time)])

                    did2 = df_vehicle_inter_0["enter_time"].values > run_counts / 2
                    duration2 = df_vehicle_inter_0["leave_time"][did2].values - df_vehicle_inter_0["enter_time"][
                        did2].values
                    ave_duration2 = np.mean([time for time in duration2 if not isnan(time)])
                    duration_each_round_list2.append(ave_duration2)

                    real_traffic_vol2 = 0
                    nan_num2 = 0
                    for time in duration2:
                        if not isnan(time):
                            real_traffic_vol2 += 1
                        else:
                            nan_num2 += 1

                    if nan_num2 < nan_thres:
                        if min_duration2 > ave_duration2 and ave_duration2 > 24:
                            min_duration2 = ave_duration2
                            min_duration_ind2 = int(round[6:])

                real_traffic_vol = 0
                nan_num = 0
                for time in duration:
                    if not isnan(time):
                        real_traffic_vol += 1
                    else:
                        nan_num += 1

                duration_each_round_list.append(ave_duration)
                # queue_length_each_round_list.append(queue_length_each_round / sample_num)
                num_of_vehicle_in.append(vehicle_in)
                num_of_vehicle_out.append(vehicle_out)

                # print(real_traffic_vol, traffic_vol, traffic_vol - real_traffic_vol, nan_num)
                # if min_queue_length > queue_length_each_round / sample_num:
                #     min_queue_length = queue_length_each_round / sample_num
                #     min_queue_length_id = int(round[6:])

                # valid_flag = json.load(open(os.path.join(round_dir, "valid_flag.json")))
                #if valid_flag['0']:  # temporary for one intersection
                if vehicle_out > total_vol * 0.9:
                    if min_duration > ave_duration and ave_duration > 24:
                        print(">", traffic_file)
                        print(">>>", ave_duration, vehicle_out, total_vol)
                        min_duration = ave_duration
                        min_duration_ind = round

                if num_seg > 1:
                    for i, interval in enumerate(range(0, run_counts, 3600)):
                        did = np.bitwise_and(df_vehicle_inter_0["enter_time"].values < interval + 3600,
                                             df_vehicle_inter_0["enter_time"].values > interval)
                        #vehicle_in_seg = sum([int(x) for x in (df_vehicle_inter_0["enter_time"][did].values > 0)])
                        #vehicle_out_seg = sum([int(x) for x in (df_vehicle_inter_0["leave_time"][did].values > 0)])
                        duration_seg = df_vehicle_inter_0["leave_time"][did].values - df_vehicle_inter_0["enter_time"][
                            did].values
                        ave_duration_seg = np.mean([time for time in duration_seg if not isnan(time)])
                        # print(traffic_file, round, i, ave_duration)
                        real_traffic_vol_seg = 0
                        nan_num_seg = 0
                        for time in duration_seg:
                            if not isnan(time):
                                real_traffic_vol_seg += 1
                            else:
                                nan_num_seg += 1

                        # print(real_traffic_vol, traffic_vol, traffic_vol - real_traffic_vol, nan_num)

                        if nan_num_seg < nan_thres:
                            # if min_duration[i] > ave_duration and ave_duration > 24:
                            list_duration_seg[i] = ave_duration_seg
                            list_duration_id_seg[i] = round

                        #round_summary = {}
                    for j in range(num_seg):
                        key = "min_duration-" + str(j)
                        if key not in round_summary.keys():
                            round_summary[key] = [list_duration_seg[j]]
                        else:
                            round_summary[key].append(list_duration_seg[j])
                    #round_result_dir = os.path.join("summary", memo, traffic_file)
                    #if not os.path.exists(round_result_dir):
                    #    os.makedirs(round_result_dir)

            except:
                duration_each_round_list.append(NAN_LABEL)
                # queue_length_each_round_list.append(NAN_LABEL)
                num_of_vehicle_in.append(NAN_LABEL)
                num_of_vehicle_out.append(NAN_LABEL)
                if "peak" in traffic_file:
                    duration_each_round_list2.append(NAN_LABEL)

        # result_dir = os.path.join(records_dir, traffic_file)
        result_dir = os.path.join("summary", memo, traffic_file)
        if not os.path.exists(result_dir):
            os.makedirs(result_dir)
        _res = {
            "duration": duration_each_round_list,
            # "queue_length": queue_length_each_round_list,
            "vehicle_in": num_of_vehicle_in,
            "vehicle_out": num_of_vehicle_out
        }
        result = pd.DataFrame(_res)
        result.to_csv(os.path.join(result_dir, "test_results.csv"))
        if num_seg > 1:
            round_result = pd.DataFrame(round_summary)
            round_result.to_csv(os.path.join(result_dir, "test_seg_results.csv"), index=False)
            plot_segment_duration(round_summary, result_dir, mode_name="test")
            duration_each_segment_list = round_result.iloc[min_duration_ind][1:].values

            traffic_name, traffic_time = traffic_file.split(".json")
            if traffic_name not in performance_at_min_duration_round:
                performance_at_min_duration_round[traffic_name] = [(duration_each_segment_list, traffic_time)]
            else:
                performance_at_min_duration_round[traffic_name].append((duration_each_segment_list, traffic_time))


        # print(os.path.join(result_dir, "test_results.csv"))

        # total_summary
        total_summary = get_metrics(duration_each_round_list, queue_length_each_round_list,
                                    min_duration, min_duration_ind, min_queue_length, min_queue_length_id,
                                    traffic_file, total_summary,
                                    mode_name="test", save_path=result_dir, num_rounds=num_rounds,
                                    min_duration2=None if "peak" not in traffic_file else min_duration2)

        if ".xml" in traffic_file:
            traffic_name, traffic_time = traffic_file.split(".xml")
        elif ".json" in traffic_file:
            traffic_name, traffic_time = traffic_file.split(".json")
        if traffic_name not in performance_duration:
            performance_duration[traffic_name] = [(duration_each_round_list, traffic_time)]
        else:
            performance_duration[traffic_name].append((duration_each_round_list, traffic_time))

    total_result = pd.DataFrame(total_summary)
    total_result.to_csv(os.path.join("summary", memo, "total_test_results.csv"))
    figure_dir = os.path.join("summary", memo, "figures")
    if not os.path.exists(figure_dir):
        os.makedirs(figure_dir)
    # if dic_exp_conf["EARLY_STOP"]:
    #     performance_duration = padding_duration(performance_duration)
    summary_plot(performance_duration, figure_dir, mode_name="test", num_rounds=num_rounds)
    performance_at_min_duration_round_plot(performance_at_min_duration_round, figure_dir, mode_name="test")


def summary_detail_baseline(memo):
    # each_round_train_duration
    total_summary = {
        "traffic": [],
        "min_queue_length": [],
        "min_queue_length_round": [],
        "min_duration": [],
        "min_duration_round": []
    }

    records_dir = os.path.join("records", memo)
    for traffic_file in os.listdir(records_dir):
        if ".xml" not in traffic_file and ".json" not in traffic_file:
            continue
        print(traffic_file)

        # if "650" not in traffic_file:
        #    continue

        # get run_counts to calculate the queue_length each second
        exp_conf = open(os.path.join(records_dir, traffic_file, "exp.conf"), 'r')
        dic_exp_conf = json.load(exp_conf)
        run_counts = dic_exp_conf["RUN_COUNTS"]

        duration_each_round_list = []
        queue_length_each_round_list = []

        train_dir = os.path.join(records_dir, traffic_file)

        # summary items (queue_length) from pickle
        f = open(os.path.join(train_dir, "inter_0.pkl"), "rb")
        try:
            samples = pkl.load(f)
        except:
            continue
        for sample in samples:
            queue_length_each_round = sum(sample['state']['lane_queue_length'])
        f.close()

        # summary items (duration) from csv
        df_vehicle_inter_0 = pd.read_csv(os.path.join(train_dir, "vehicle_inter_0.csv"),
                                         sep=',', header=0, dtype={0: str, 1: float, 2: float},
                                         names=["vehicle_id", "enter_time", "leave_time"])

        duration = df_vehicle_inter_0["leave_time"].values - df_vehicle_inter_0["enter_time"].values
        ave_duration = np.mean([time for time in duration if not isnan(time)])
        # print(ave_duration)

        duration_each_round_list.append(ave_duration)
        ql = queue_length_each_round / len(samples)
        queue_length_each_round_list.append(ql)

        # result_dir = os.path.join(records_dir, traffic_file)
        result_dir = os.path.join("summary", memo, traffic_file)
        if not os.path.exists(result_dir):
            os.makedirs(result_dir)
        _res = {
            "duration": duration_each_round_list,
            "queue_length": queue_length_each_round_list
        }
        result = pd.DataFrame(_res)
        result.to_csv(os.path.join(result_dir, "test_results.csv"))
        # print(os.path.join(result_dir, "test_results.csv"))

        total_summary["traffic"].append(traffic_file)
        total_summary["min_queue_length"].append(ql)
        total_summary["min_queue_length_round"].append(0)
        total_summary["min_duration"].append(ave_duration)
        total_summary["min_duration_round"].append(0)

    total_result = pd.DataFrame(total_summary)
    total_result.to_csv(os.path.join("summary", memo, "total_baseline_test_results.csv"))


def main(memo=None):
    total_summary = {
        "traffic": [],
        "traffic_file": [],
        "min_queue_length": [],
        "min_queue_length_round": [],
        "min_duration": [],
        "min_duration_round": [],
        "final_duration": [],
        "final_duration_std": [],
        "convergence_1.2": [],
        "convergence_1.1": [],
        "nan_count": [],
        "min_duration2": []
    }
    if not memo:
        memo = "pipeline_500"
    #summary_detail_train(memo, copy.deepcopy(total_summary))
    summary_detail_test(memo, copy.deepcopy(total_summary))
    # summary_detail_test_segments(memo, copy.deepcopy(total_summary))


if __name__ == "__main__":
    total_summary = {
        "traffic": [],
        "traffic_file": [],
        "min_queue_length": [],
        "min_queue_length_round": [],
        "min_duration": [],
        "min_duration_round": [],
        "final_duration": [],
        "final_duration_std": [],
        "convergence_1.2": [],
        "convergence_1.1": [],
        "nan_count": [],
        "min_duration2": []
    }

    memo = "syn"
    # summary_detail_train(memo, copy.deepcopy(total_summary))
    summary_detail_test(memo, copy.deepcopy(total_summary))
    # summary_detail_baseline(memo)
    #summary_detail_test_segments(memo, copy.deepcopy(total_summary))