YuXT/18_Multi_cycle_inventory_prediction_algorithm

1import time2 3import matplotlib.pyplot as plt4import numpy as np5import pandas as pd6from keras.layers.core import Dense, Activation, Dropout7from keras.layers import LSTM8from keras.models import Sequential9from sklearn.metrics import explained_variance_score, mean_absolute_error, mean_squared_error, median_absolute_error, \10    r2_score11import gradio as gr12 13examples = [14    '1天'15]16 17 18class Conf:19    # epochs20    EPOCHS = 5021    # 时间序列长度22    SEQ_LEN = 5023    # 预测步数24    PREDICT_STEP = 725    # 测试训练集比例26    TRAIN_DATA_RATE = 0.827    # 批大小28    BATCH_SIZE = 5029    # 网络形状30    LAYERS = [1, 50, 100, 1]31 32 33def model_evaluation(y_true, y_pred):34    metrics = _cal_metrics(y_true, y_pred)35    for (k, v) in metrics.items():36        print(k + ": " + str(v))37 38 39def _cal_metrics(y_true, y_pred):40    """41    计算各个指标的值42    """43    re = _calc_re(y_true, y_pred)44    metrics = {45        "explained_variance_score":46            explained_variance_score(y_true, y_pred),47        "mean_absolute_error":48            mean_absolute_error(y_true, y_pred),49        "mean_squared_error":50            mean_squared_error(y_true, y_pred),51        "median_absolute_error":52            median_absolute_error(y_true, y_pred),53        "r2_score":54            r2_score(y_true, y_pred),55        "sum_relative_error":56            re[0],57        "mean_relative_error":58            re[1]59    }60 61    return metrics62 63 64def _calc_re(y_true, y_pred):65    """66    计算相对误差（Sum/Mean Relative Error）67    """68    return [((y_true - y_pred) / y_pred).sum().values, ((y_true - y_pred) / y_pred).mean().values]69 70 71def load_data(filename):72    """73    数据准备74    """75    data = pd.read_csv(filename).values76 77    result_0 = []78    for index in range(len(data) - Conf.SEQ_LEN - 1):79        result_0.append(data[index: index + Conf.SEQ_LEN + 1])80    # 数据标准化81 82    result = normalise_windows(result_0)83 84    result = np.array(result)85    result_0 = np.array(result_0)86 87    row = round(result.shape[0] * Conf.TRAIN_DATA_RATE)88    train = result[:int(row), :]89    np.random.shuffle(train)90 91    _X_train = train[:, :-1]92    _y_train = train[:, -1]93    _X_test = result[int(row):, :-1]94    _y_test = result[int(row):, -1]95    _y_test_p_0 = result_0[int(row):, 0]96 97    # 增加一列98    _X_train = _X_train[:, :, np.newaxis]99    _X_test = _X_test[:, :, np.newaxis]100 101    print(_X_train.shape)102    print(_X_test.shape)103    return [_X_train, _y_train, _X_test, _y_test, _y_test_p_0]104 105 106def normalise_windows(window_data):107    """108    对原始数据做标准化：n_i = (p_i/p)0 - 1)109    对应的反标准化公式为：p_i = p_0(n_i + 1)110    """111    normalised_data = []112    for window in window_data:113        normalised_window = [((float(p) / float(window[0])) - 1) for p in window]114        normalised_data.append(normalised_window)115    return normalised_data116 117 118def anti_normalise_windows(p_data, normalised_data):119    """120    对原始数据做标准化：n_i = (p_i/p_0 - 1)121    对应的反标准化公式为：p_i = p_0(n_i + 1)122    """123    anti_normalised_data = []124    for (n, p_0) in zip(normalised_data, p_data):125        anti_normalised_window = p_0 * (n + 1)126        anti_normalised_data.append(anti_normalised_window)127    return anti_normalised_data128 129 130def build_model(layers):131    """132    模型定义133    """134    model = Sequential()135 136    model.add(LSTM(units=layers[1], input_shape=(layers[1], layers[0]), return_sequences=True))137    model.add(Dropout(0.2))138 139    model.add(LSTM(layers[2], return_sequences=False))140    model.add(Dropout(0.2))141 142    model.add(Dense(units=layers[3]))143    model.add(Activation("tanh"))144 145    start = time.time()146    model.compile(loss="mse", optimizer="rmsprop")147    print("> Compilation Time : ", time.time() - start)148    return model149 150 151def predict_point_by_point(model, data):152    """153    每次预测1步154    """155    predict = model.predict(data)156    predict = np.reshape(predict, (len(predict),))157    return predict158 159 160def predict_by_len(model, data, predict_len):161    """162    预测predict_len步163    """164    predicted = []165    for i in range(predict_len):166        predicted.append(model.predict(data[np.newaxis, :, :])[0, 0])167        data = data[1:]168        data = np.insert(data, -1, predicted[-1], axis=0)169    return predicted170 171 172def plot_results(y_true, y_pred):173    fig = plt.figure(facecolor='white')174    ax = fig.add_subplot(111)175    ax.plot(y_true, label='True Data')176    plt.plot(y_pred, label='Prediction')177    plt.legend()178    return fig179 180 181global_start_time = time.time()182 183print('> Loading data... ')184 185X_train, y_train, X_test, y_test, y_test_p_0 = load_data('price.csv')186 187print('> Data Loaded. Compiling...')188 189model = build_model(Conf.LAYERS)190 191model.fit(X_train, y_train, batch_size=Conf.BATCH_SIZE, epochs=Conf.EPOCHS, validation_split=0.05)192 193# 预测一步194predicted = predict_point_by_point(model, X_test)195 196# 预测7天197predicted_len = predict_by_len(model, X_test[-1], 7)198predicted_len_0 = anti_normalise_windows(y_test_p_0, predicted_len)199 200# 7天绘图201fig = plt.figure(facecolor='white')202ax = fig.add_subplot(111)203ax.plot(predicted_len_0, label='Prediction')204plt.legend()205 206print('Training duration (s) : ', time.time() - global_start_time)207 208y_test_0 = anti_normalise_windows(y_test_p_0, y_test)209 210# 预测一步211predicted_0 = anti_normalise_windows(y_test_p_0, predicted)212# 预测一步绘图213fig1 = plot_results(y_test_0, predicted_0)214 215# 模型评估216model_evaluation(pd.DataFrame(y_test_0), pd.DataFrame(predicted_0))217 218 219def process(choice):220    """221    整个后端的输入输出都在这里体现，也就是模型的推理部分。222    """223    if choice == "1天":224        output = fig1225    else:226        output = fig227    # 对推理结果做特定处理，方便展示228    return output  # 返回处理后的结果229 230 231iface = gr.Interface(  # 定义这个gr的接口232    fn=process,  # 实现输入输出的函数233    inputs=gr.Dropdown(["1天", "7天"], label="选择天数"),  # 输入本文框234    outputs=gr.Plot(label="Plot"),  # 输出结果以什么样的形式展示，随意替换235    # outputs=gr.Textbox(),236    title="多周期库存算法",237    examples=examples,  # gradio直接给设计了examples这个属性238    # allow_flagging="never",239)240iface.launch(debug=True, share=False)  # 写好对象直接launch就行241