TimesFM：根據解碼器架構的時序預測模型應用

import numpy as np
import pandas as pd
import timesfm
import matplotlib.pyplot as plt
from sklearn.metrics import mean_squared_error, mean_absolute_percentage_error, r2_score
import os

# 設定環境變數以避免使用特定硬體加速
os.environ['XLA_PYTHON_CLIENT_PREALLOCATE'] = 'false'
os.environ['JAX_PMAP_USE_TENSORSTORE'] = 'false'

# 使用 pandas 讀取 CSV 資料並轉換日期格式
Y_df = pd.read_csv('AirPassengersDataset.csv')
Y_df['ds'] = pd.to_datetime(Y_df['ds'])
Y_df = Y_df.reset_index(drop=True)
Y_df.head()

# 分割資料集為訓練和測試集
Y_train_df = Y_df[Y_df.ds <= '1959-12-31']
Y_test_df = Y_df[Y_df.ds > '1959-12-31']

# 組態使用 CPU 作為後端計算資源
timesfm_backend = "cpu"
from jax._src import config

config.update("jax_platforms", {"cpu": "cpu", "gpu": "cuda", "tpu": ""}[timesfm_backend])

# 初始化 TimesFM 模型並定義引數
tfm = timesfm.TimesFm(
    context_len=128,
    horizon_len=12,
    input_patch_len=32,
    output_patch_len=128,
    num_layers=20,
    model_dims=1280,
    backend=timesfm_backend,
)

# 載入預訓練模型
tfm.load_from_checkpoint(repo_id="google/timesfm-1.0-200m")

# 生成預測結果
timesfm_forecast = tfm.forecast_on_df(
    inputs=Y_train_df,
    freq="MS",
    value_name="y",
    num_jobs=-1,
)

timesfm_forecast = timesfm_forecast[["ds","timesfm"]]
timesfm_forecast.head()

# 評估模型效能並計算錯誤指標
def calculate_error_metrics(y_true, y_pred):
    mse = mean_squared_error(y_true, y_pred)
    rmse = np.sqrt(mse)
    mape = mean_absolute_percentage_error(y_true, y_pred)
    r2 = r2_score(y_true, y_pred)
    adjusted_r2 = r2 - ( (len(y_true) - len(y_pred)) / (len(y_true) - len(y_pred) - len(tfm.params)) ) * (r2 - r)

    return {
        "MSE": mse,
        "RMSE": rmse,
        "MAPE": mape,
        "R^2": r2,
        "Adjusted R^2": adjusted_r2,
    }

error_metrics = calculate_error_metrics(Y_test_df[['y']], timesfm_forecast['timesfm'])
print(error_metrics)

# 改變索引並繪製圖表以比較真實值與預測值
Y_train_df.set_index('ds', inplace=True)
timesfm_forecast.set_index('ds', inplace=True)
Y_test_df.set_index('ds', inplace=True)

plt.figure(figsize=(20,5))
y_past = Y_train_df["y"][-50:]
y_pred = timesfm_forecast['timesfm']
y_test = Y_test_df["y"]

plt.plot(y_past, label="過去的時間序列值")
plt.plot(timesfm_forecast, label="預測")
plt.plot(y_test, label="實際時間序列值")

plt.title('航班乘客數預測', fontsize=15)
plt.ylabel('月度乘客人數', fontsize=15)
plt.xlabel('時間戳 [t]', fontsize=15)
plt.tight_layout()
plt.xticks(rotation=90)
plt.legend()
plt.show()

df = pd.read_csv('EPF_FR_BE.csv')
df[df['unique_id'] == 'FR'].head()

from collections import defaultdict

def get_batched_data_fn(batch_size: int = 64, context_len: int = 64, horizon_len: int = 3):
    examples = defaultdict(list)

    for country in ("FR", "BE"):
        sub_df = df[df["unique_id"] == country]

        for start in range(horizon_len + context_len):
            examples["country"].append(country)

            examples["inputs"].append(sub_df["y"][start:(start + context_len)].tolist())
            examples["gen_forecast"].append(sub_df["gen_forecast"][start:(start + context_len + horizon_len)].tolist())
            examples["week_day"].append(sub_df["week_day"][start:(start + context_len + horizon_len)].tolist())

            examples["outputs"].append(sub_df["y"][start + context_len:(start + context_len + horizon_len)].tolist())

    def data_fn():
        for i in range(len(examples[list(examples.keys())[0]])):
            yield {k: v[i:i + batch_size] for k, v in examples.items()}

    return data_fn

batched_data_fn = get_batched_data_fn()

timesfm_backend = "cpu"
from jax._src import config

config.update("jax_platforms", {"cpu": "cpu", "gpu": "cuda", "tpu": ""}[timesfm_backend])

model = timesfm.TimesFm(
    context_len=context_length,
    horizon_len=horizon_length,
)