消融实验是否真正进行，变长Patch划分是否真的有用，PaEn_loss怎么正常的进行参数学习

[cooker-coder]

在模型主干部分，将Pane_loss手动设置为0.且不使用Patch变长划分方法。得到如下实验结果:

完整实验日志：
D:\Anaconda\envs\iTransformer\python.exe D:\project\HDMixer-master\run_longExp.py --is_training 1 --model_id test --model HDMixer
Args in experiment:
Namespace(mix_time=1, mix_variable=1, mix_channel=1, deform_patch=0, deform_range=0.25, lambda_=0.1, r=0.01, mlp_ratio=1.0, window_size=6, shift_size=3, weight_decay=0.001, random_seed=2021, is_training=1, model_id='test', model='HDMixer', data='ETTh1', root_path='./dataset/', data_path='ETTh1.csv', features='M', target='OT', freq='h', checkpoints='./checkpoints/', seq_len=336, label_len=48, pred_len=96, fc_dropout=0.3, head_dropout=0.0, patch_len=16, stride=8, padding_patch='end', revin=1, affine=0, subtract_last=0, decomposition=0, kernel_size=25, individual=0, embed_type=0, enc_in=7, dec_in=7, c_out=7, d_model=16, n_heads=4, e_layers=1, d_layers=1, d_ff=32, moving_avg=25, factor=1, distil=True, dropout=0.8, embed='timeF', activation='gelu', output_attention=True, do_predict=True, num_workers=0, itr=1, train_epochs=50, batch_size=512, patience=20, learning_rate=0.005, des='Exp', loss='mse', lradj='type3', pct_start=0.3, use_amp=False, use_gpu=True, gpu=0, use_multi_gpu=False, devices='4,5,6,7', test_flop=False)
Use GPU: cuda:0
41
42

start training : test_HDMixer_ETTh1_ftM_sl336_ll48_pl96_dm16_nh4_el1_dl1_df32_fc1_ebtimeF_dtTrue_Exp_0>>>>>>>>>>>>>>>>>>>>>>>>>>
train 8209
val 2785
test 2785

Epoch: 1 cost time: 0.4792213439941406
Epoch: 1, Steps: 16 | Train Loss: 0.6970009 Vali Loss: 1.0869455 Test Loss: 0.5323638
Validation loss decreased (inf --> 1.086946). Saving model ...
Updating learning rate to 0.005

Epoch: 2 cost time: 0.2640256881713867
Epoch: 2, Steps: 16 | Train Loss: 0.6260976 Vali Loss: 0.9772700 Test Loss: 0.5195729
Validation loss decreased (1.086946 --> 0.977270). Saving model ...
Updating learning rate to 0.005

Epoch: 3 cost time: 0.26801466941833496
Epoch: 3, Steps: 16 | Train Loss: 0.4422899 Vali Loss: 0.7530154 Test Loss: 0.4579403
Validation loss decreased (0.977270 --> 0.753015). Saving model ...
Updating learning rate to 0.005

Epoch: 4 cost time: 0.26700329780578613
Epoch: 4, Steps: 16 | Train Loss: 0.3889118 Vali Loss: 0.7376963 Test Loss: 0.3982546
Validation loss decreased (0.753015 --> 0.737696). Saving model ...
Updating learning rate to 0.0045000000000000005

Epoch: 5 cost time: 0.27027344703674316
Epoch: 5, Steps: 16 | Train Loss: 0.3626329 Vali Loss: 0.6873276 Test Loss: 0.3911011
Validation loss decreased (0.737696 --> 0.687328). Saving model ...
Updating learning rate to 0.004050000000000001

Epoch: 6 cost time: 0.2700023651123047
Epoch: 6, Steps: 16 | Train Loss: 0.3479787 Vali Loss: 0.6780385 Test Loss: 0.3767796
Validation loss decreased (0.687328 --> 0.678038). Saving model ...
Updating learning rate to 0.0036450000000000007

Epoch: 7 cost time: 0.26932525634765625
Epoch: 7, Steps: 16 | Train Loss: 0.3428182 Vali Loss: 0.6805835 Test Loss: 0.3700825
EarlyStopping counter: 1 out of 20
Updating learning rate to 0.0032805

Epoch: 8 cost time: 0.2701568603515625
Epoch: 8, Steps: 16 | Train Loss: 0.3396684 Vali Loss: 0.6660573 Test Loss: 0.3697884
Validation loss decreased (0.678038 --> 0.666057). Saving model ...
Updating learning rate to 0.0029524500000000006

Epoch: 9 cost time: 0.28374457359313965
Epoch: 9, Steps: 16 | Train Loss: 0.3381216 Vali Loss: 0.6794029 Test Loss: 0.3658020
EarlyStopping counter: 1 out of 20
Updating learning rate to 0.002657205

Epoch: 10 cost time: 0.26601290702819824
Epoch: 10, Steps: 16 | Train Loss: 0.3369397 Vali Loss: 0.6707555 Test Loss: 0.3663085
EarlyStopping counter: 2 out of 20
Updating learning rate to 0.0023914845000000003

Epoch: 11 cost time: 0.26800107955932617
Epoch: 11, Steps: 16 | Train Loss: 0.3363151 Vali Loss: 0.6673336 Test Loss: 0.3646697
EarlyStopping counter: 3 out of 20
Updating learning rate to 0.0021523360500000006

Epoch: 12 cost time: 0.26697278022766113
Epoch: 12, Steps: 16 | Train Loss: 0.3359175 Vali Loss: 0.6708089 Test Loss: 0.3654612
EarlyStopping counter: 4 out of 20
Updating learning rate to 0.0019371024450000004

Epoch: 13 cost time: 0.2676119804382324
Epoch: 13, Steps: 16 | Train Loss: 0.3356292 Vali Loss: 0.6697249 Test Loss: 0.3642505
EarlyStopping counter: 5 out of 20
Updating learning rate to 0.0017433922005000006

Epoch: 14 cost time: 0.2702009677886963
Epoch: 14, Steps: 16 | Train Loss: 0.3353219 Vali Loss: 0.6727983 Test Loss: 0.3644657
EarlyStopping counter: 6 out of 20
Updating learning rate to 0.0015690529804500003

Epoch: 15 cost time: 0.2720067501068115
Epoch: 15, Steps: 16 | Train Loss: 0.3349026 Vali Loss: 0.6730704 Test Loss: 0.3643808
EarlyStopping counter: 7 out of 20
Updating learning rate to 0.0014121476824050007

Epoch: 16 cost time: 0.2844867706298828
Epoch: 16, Steps: 16 | Train Loss: 0.3349119 Vali Loss: 0.6706899 Test Loss: 0.3641191
EarlyStopping counter: 8 out of 20
Updating learning rate to 0.0012709329141645004

Epoch: 17 cost time: 0.26850175857543945
Epoch: 17, Steps: 16 | Train Loss: 0.3348218 Vali Loss: 0.6739876 Test Loss: 0.3640708
EarlyStopping counter: 9 out of 20
Updating learning rate to 0.0011438396227480504

Epoch: 18 cost time: 0.27217555046081543
Epoch: 18, Steps: 16 | Train Loss: 0.3344863 Vali Loss: 0.6649176 Test Loss: 0.3637382
Validation loss decreased (0.666057 --> 0.664918). Saving model ...
Updating learning rate to 0.0010294556604732454

Epoch: 19 cost time: 0.26722288131713867
Epoch: 19, Steps: 16 | Train Loss: 0.3343521 Vali Loss: 0.6679214 Test Loss: 0.3641189
EarlyStopping counter: 1 out of 20
Updating learning rate to 0.0009265100944259208

Epoch: 20 cost time: 0.2639956474304199
Epoch: 20, Steps: 16 | Train Loss: 0.3341514 Vali Loss: 0.6696994 Test Loss: 0.3638243
EarlyStopping counter: 2 out of 20
Updating learning rate to 0.0008338590849833288

Epoch: 21 cost time: 0.2739288806915283
Epoch: 21, Steps: 16 | Train Loss: 0.3343266 Vali Loss: 0.6666123 Test Loss: 0.3635328
EarlyStopping counter: 3 out of 20
Updating learning rate to 0.0007504731764849959

Epoch: 22 cost time: 0.28064680099487305
Epoch: 22, Steps: 16 | Train Loss: 0.3341014 Vali Loss: 0.6711146 Test Loss: 0.3637099
EarlyStopping counter: 4 out of 20
Updating learning rate to 0.0006754258588364964

Epoch: 23 cost time: 0.27042412757873535
Epoch: 23, Steps: 16 | Train Loss: 0.3342334 Vali Loss: 0.6717736 Test Loss: 0.3640175
EarlyStopping counter: 5 out of 20
Updating learning rate to 0.0006078832729528468

Epoch: 24 cost time: 0.2941441535949707
Epoch: 24, Steps: 16 | Train Loss: 0.3342039 Vali Loss: 0.6736273 Test Loss: 0.3632407
EarlyStopping counter: 6 out of 20
Updating learning rate to 0.0005470949456575622

Epoch: 25 cost time: 0.2786116600036621
Epoch: 25, Steps: 16 | Train Loss: 0.3339713 Vali Loss: 0.6671935 Test Loss: 0.3636402
EarlyStopping counter: 7 out of 20
Updating learning rate to 0.0004923854510918059

Epoch: 26 cost time: 0.2823636531829834
Epoch: 26, Steps: 16 | Train Loss: 0.3339507 Vali Loss: 0.6698967 Test Loss: 0.3636934
EarlyStopping counter: 8 out of 20
Updating learning rate to 0.00044314690598262535

Epoch: 27 cost time: 0.2686622142791748
Epoch: 27, Steps: 16 | Train Loss: 0.3337385 Vali Loss: 0.6671231 Test Loss: 0.3634807
EarlyStopping counter: 9 out of 20
Updating learning rate to 0.0003988322153843628

Epoch: 28 cost time: 0.2706575393676758
Epoch: 28, Steps: 16 | Train Loss: 0.3337661 Vali Loss: 0.6692212 Test Loss: 0.3635621
EarlyStopping counter: 10 out of 20
Updating learning rate to 0.0003589489938459265

Epoch: 29 cost time: 0.2756321430206299
Epoch: 29, Steps: 16 | Train Loss: 0.3339321 Vali Loss: 0.6690661 Test Loss: 0.3636963
EarlyStopping counter: 11 out of 20
Updating learning rate to 0.00032305409446133386

Epoch: 30 cost time: 0.2682504653930664
Epoch: 30, Steps: 16 | Train Loss: 0.3339566 Vali Loss: 0.6705001 Test Loss: 0.3636905
EarlyStopping counter: 12 out of 20
Updating learning rate to 0.0002907486850152005

Epoch: 31 cost time: 0.2700002193450928
Epoch: 31, Steps: 16 | Train Loss: 0.3339658 Vali Loss: 0.6708741 Test Loss: 0.3634060
EarlyStopping counter: 13 out of 20
Updating learning rate to 0.0002616738165136805

Epoch: 32 cost time: 0.2805604934692383
Epoch: 32, Steps: 16 | Train Loss: 0.3339972 Vali Loss: 0.6701417 Test Loss: 0.3635286
EarlyStopping counter: 14 out of 20
Updating learning rate to 0.00023550643486231244

Epoch: 33 cost time: 0.273252010345459
Epoch: 33, Steps: 16 | Train Loss: 0.3336993 Vali Loss: 0.6769592 Test Loss: 0.3633735
EarlyStopping counter: 15 out of 20
Updating learning rate to 0.0002119557913760812

Epoch: 34 cost time: 0.2915949821472168
Epoch: 34, Steps: 16 | Train Loss: 0.3338362 Vali Loss: 0.6752858 Test Loss: 0.3634149
EarlyStopping counter: 16 out of 20
Updating learning rate to 0.00019076021223847308

Epoch: 35 cost time: 0.28769588470458984
Epoch: 35, Steps: 16 | Train Loss: 0.3337940 Vali Loss: 0.6715749 Test Loss: 0.3633811
EarlyStopping counter: 17 out of 20
Updating learning rate to 0.00017168419101462577
Epoch: 36 cost time: 0.29315876960754395
Epoch: 36, Steps: 16 | Train Loss: 0.3337394 Vali Loss: 0.6721165 Test Loss: 0.3634064
EarlyStopping counter: 18 out of 20
Updating learning rate to 0.00015451577191316317
Epoch: 37 cost time: 0.29315996170043945
Epoch: 37, Steps: 16 | Train Loss: 0.3334903 Vali Loss: 0.6692940 Test Loss: 0.3634848
EarlyStopping counter: 19 out of 20
Updating learning rate to 0.00013906419472184688

Epoch: 38 cost time: 0.2873845100402832
Epoch: 38, Steps: 16 | Train Loss: 0.3336127 Vali Loss: 0.6707417 Test Loss: 0.3634249
EarlyStopping counter: 20 out of 20
Early stopping

testing : test_HDMixer_ETTh1_ftM_sl336_ll48_pl96_dm16_nh4_el1_dl1_df32_fc1_ebtimeF_dtTrue_Exp_0<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<
test 2785
mse:0.3637380599975586, mae:0.3900781273841858, rse:0.5743133425712585

结果优于论文中的HDMixer呈现出来的准确度.

[strongerux]

指标好也是一坨，现在的时序模型太水了，test画出来趋势都对不上，纯瞎蒙一个值接近ground truth就完事，NIPS2019那篇Shape and Time Distortion Loss for Training Deep Time Series Forecasting Models 形状跑出来很好不知道为什么没人跟，长期预测我宁愿信趋势都不会信精确值，都在刷一点没用的点，跑着玩玩就行了不要太较真哈哈。

[guangyuyang001]

时序上的可变形感觉有点像伪命题，今年的顶会也有好几篇但是实际效果...