Fix M4 dataset sequence length mismatch

[leviAckermea]

问题描述

M4 数据集在加载时存在序列长度不一致的问题，导致readme训练与评测中第二个部分模型训练。
bash ./scripts/short_term_forecast/TimesNet_M4.sh
执行bug：

修改内容

1. 调整 data_provider/m4.py 中数据解析逻辑，统一序列长度处理方式。
2. 优化 {insert_element_0_YGRhdGFfcHJvdmlkZXIvZGF0YV9sb2FkZXIucHlg} 中的滑窗截取逻辑，兼容 M4 数据集的长度特性。

验证方式

已通过 {insert_element_1_YHNjcmlwdHMvc2hvcnRfdGVybV9mb3JlY2FzdC9UaW1lc05ldF9NNC5zaGA=} 脚本测试，可正常运行 M4 数据集的短期预测任务，无长度不匹配报错。

[leviAckermea]

执行bug如下：ValueError: setting an array element with a sequence. The requested array has an inhomogeneous shape after 1 dimensions. The detected shape was (48000,) + inhomogeneous part.

[leehom0123-cmyk]

Epoch: 10 cost time: 143.77790188789368
Epoch: 10, Steps: 26 | Train Loss: 14.3322291 Vali Loss: 33.2273196 Test Loss: 33.2273196
EarlyStopping counter: 1 out of 3
Updating learning rate to 1.953125e-06

testing : short_term_forecast_m4_Hourly_TimesNet_m4_ftM_sl96_ll48_pl48_dm32_nh8_el2_dl1_df32_expand2_dc4_fc3_ebtimeF_dtTrue_Exp_0<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<
train 414
test 414
0
test shape: (414, 48, 1)
TimesNet
Traceback (most recent call last):
File "/home/bit/tslib/run.py", line 224, in
exp.test(setting)
File "/home/bit/tslib/exp/exp_short_term_forecasting.py", line 228, in test
smape_results, owa_results, mape, mase = m4_summary.evaluate()
^^^^^^^^^^^^^^^^^^^^^
File "/home/bit/tslib/utils/m4_summary.py", line 67, in evaluate
naive2_forecasts = np.array([v[~np.isnan(v)] for v in naive2_forecasts])
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
ValueError: setting an array element with a sequence. The requested array has an inhomogeneous shape after 1 dimensions. The detected shape was (100000,) + inhomogeneous part.

我测试了一下，还有以上错误。

[leviAckermea]

非常感谢你的来信 我已收到