Simulation

ITSTGCN
Author

SEOYEON CHOI

Published

April 5, 2023

Simulation Study

Import

데이터 함쳐놓기

import pandas as pd

_a = pd.read_csv('./simulation_results/a.csv')
_b = pd.read_csv('./simulation_results/b.csv')
_c = pd.read_csv('./simulation_results/STGCN_ITSTGCN_random_epoch200.csv')

_df = pd.concat([_a,_b,_c],axis=0)

_df

_df.to_csv('./simulation_results/STGCN_ITSTGCN_random_epoch200_2.csv',index=False)
import matplotlib.pyplot as plt
import pandas as pd
import numpy as np

Fivenodes

random

  • lags = 2
    • GNAR 문서에 나온 대로 AR(2) 모형
  1. mrate = 0.8, filter = 12, epoch = 150
  2. mrate = 0.3, filter = 8, epoch = 50
  • interpolation = linear 또는 cubic, nearest 는 mse 너무 우리 방법에서 안 좋다.

block

random

df1 = pd.read_csv('./simulation_results/fivenodes/fivenodes_STGCN_ITSTGCN_random_epoch50.csv')
df2 = pd.read_csv('./simulation_results/fivenodes/fivenodes_STGCN_ITSTGCN_random_epoch100.csv')
df3 = pd.read_csv('./simulation_results/fivenodes/fivenodes_STGCN_ITSTGCN_random_epoch150.csv')
df4 = pd.read_csv('./simulation_results/fivenodes/fivenodes_STGCN_ITSTGCN_random_epoch200.csv')
df_gnar = pd.read_csv('./simulation_results/fivenodes/fivenodes_GNAR_random.csv')
data = pd.concat([df1,df2,df3,df4,df_gnar],axis=0)
data.query("method!='GNAR' and inter_method=='linear' and lags==2").plot.box(backend='plotly',x='mrate',color='method',y='mse',facet_col='epoch',facet_row='nof_filters',height=1200)

시뮬 예정(평균 시간, 평균mse)

0.7,0.75,0.8,0.85

12,16

150

# 1. mrate = 0.8, filter = 12, epoch = 150
data.query("mrate==0.8 and inter_method=='linear' and nof_filters==12 and epoch==150 and lags==2")['calculation_time'].mean(),data.query("mrate==0.8 and inter_method=='linear' and nof_filters==12 and epoch==150 and lags==2")['mse'].mean()
(109.59549897114435, 1.2304790377616883)
data.query("mrate==0.8 and inter_method=='linear' and nof_filters==12 and epoch==150 and lags==2").plot.box(backend='plotly',x='mrate',color='method',y='mse',facet_col='epoch',facet_row='nof_filters',height=400)

block

df1 = pd.read_csv('./simulation_results/fivenodes/fivenodes_STGCN_ITSTGCN_block_node1_epoch50.csv')
df2 = pd.read_csv('./simulation_results/fivenodes/fivenodes_STGCN_ITSTGCN_block_node1_epoch100.csv')
df3 = pd.read_csv('./simulation_results/fivenodes/fivenodes_STGCN_ITSTGCN_block_node1_epoch150.csv')
df4 = pd.read_csv('./simulation_results/fivenodes/fivenodes_STGCN_ITSTGCN_block_node1_epoch200.csv')
df5 = pd.read_csv('./simulation_results/fivenodes/fivenodes_STGCN_ITSTGCN_block_node2_epoch50.csv')
df6 = pd.read_csv('./simulation_results/fivenodes/fivenodes_STGCN_ITSTGCN_block_node2_epoch100.csv')
df7 = pd.read_csv('./simulation_results/fivenodes/fivenodes_STGCN_ITSTGCN_block_node2_epoch150.csv')
df8 = pd.read_csv('./simulation_results/fivenodes/fivenodes_GNAR_block_node1.csv')
df9 = pd.read_csv('./simulation_results/fivenodes/fivenodes_GNAR_block_node2.csv')
df1['block']=1
df2['block']=1
df3['block']=1
df4['block']=1
df5['block']=2
df6['block']=2
df7['block']=2
df8['block']=1
df9['block']=2
data2 = pd.concat([df1,df2,df3,df4,df5,df6,df7,df8,df9],axis=0)
data2.query("method=='GNAR' and block == 1")['mse'].mean(),data2.query("method=='GNAR' and block == 2")['mse'].mean()
(1.455923080444336, 1.5004450678825378)
data2.query("method=='GNAR' and inter_method == 'linear'")['mse'].mean(),data2.query("method=='GNAR' and inter_method == 'nearest'")['mse'].mean() # 차이 없음
(1.4813642161233085, 1.4813642161233085)
data2.query("epoch==50")['calculation_time'].mean(),data2.query("epoch==50")['calculation_time'].max()
(39.11611335332747, 56.8712797164917)
data2.query("epoch==150")['calculation_time'].mean(),data2.query("epoch==150")['calculation_time'].max()
(102.26520284502594, 152.8869686126709)
data2.query("method!='GNAR' and lags == 2 and inter_method=='nearest'").plot.box(backend='plotly',x='block',color='method',y='mse',facet_col='epoch',facet_row='nof_filters',height=800)
data2.query("inter_method=='linear' and epoch==150").plot.box(backend='plotly',x='block',color='method',y='mse',facet_col='lags',facet_row='nof_filters',height=1200)

시뮬 예정(평균 시간, 평균mse)

block 1,2 위 세팅 그대로

랜덤ㅁ 말고 block만

# 1. block = 2 interpolation = linear, filter = 12, epoch = 150
data2.query("block==1 and inter_method=='linear' and nof_filters==12 and epoch==50 and lags==2")['calculation_time'].mean(),data2.query("block==2 and inter_method=='linear' and nof_filters==12 and epoch==50 and lags==2")['mse'].mean()
(40.18422634204229, 1.2096982955932618)
data2.query("block==1 and inter_method=='linear' and nof_filters==12 and epoch==50 and lags==2").plot.box(backend='plotly',x='block',color='method',y='mse',facet_col='epoch',facet_row='nof_filters',height=400)

fivenodes simulation result

mindex= [[],[],[],list(range(50,150)),[]] # block 1
mindex= [list(range(10,100)),[],list(range(50,80)),[],[]] # node 2 30% mmissing

block 조건

df1 = pd.read_csv('./simulation_results/2023-04-09_23-37-17.csv') # GNAR random
df2 = pd.read_csv('./simulation_results/2023-04-10_07-06-32.csv') # STGCN, ITSTGCN random 70%, 75%
df3 = pd.read_csv('./simulation_results/2023-04-10_14-54-51.csv') # STGCN, ITSTGCN random 80%, 85%
df4 = pd.read_csv('./simulation_results/2023-04-10_15-54-03.csv') # GNAR block 1
df5 = pd.read_csv('./simulation_results/2023-04-10_15-56-27.csv') # GNAR block 2
df6 = pd.read_csv('./simulation_results/2023-04-10_23-44-52.csv') # STGCN, ITSTGCN block 1
df7 = pd.read_csv('./simulation_results/2023-04-11_04-40-00.csv') # STGCN, ITSTGCN block 2
df8 = pd.read_csv('./simulation_results/2023-04-14_21-21-34.csv') # S/TGCN, ITSTGCN missing 0
data = pd.concat([df1,df2, df3, df4,df5,df6,df7,df8],axis=0);data
dataset method mrate mtype lags nof_filters inter_method epoch mse calculation_time
0 five_nodes GNAR 0.70 rand 2 NaN linear NaN 1.406830 0.022885
1 five_nodes GNAR 0.75 rand 2 NaN linear NaN 1.406830 0.005927
2 five_nodes GNAR 0.80 rand 2 NaN linear NaN 1.406830 0.005557
3 five_nodes GNAR 0.85 rand 2 NaN linear NaN 1.406830 0.010217
4 five_nodes GNAR 0.70 rand 2 NaN linear NaN 1.406830 0.006891
... ... ... ... ... ... ... ... ... ... ...
235 fivenodes STGCN 0.00 NaN 2 16.0 NaN 150.0 1.162979 117.819705
236 fivenodes IT-STGCN 0.00 NaN 2 12.0 NaN 150.0 1.156077 122.355274
237 fivenodes IT-STGCN 0.00 NaN 2 12.0 NaN 150.0 1.162236 122.169977
238 fivenodes IT-STGCN 0.00 NaN 2 16.0 NaN 150.0 1.145952 123.042743
239 fivenodes IT-STGCN 0.00 NaN 2 16.0 NaN 150.0 1.158429 124.601893

1212 rows × 10 columns

data.to_csv('./simulation_results/Real_simulation/fivedones_Simulation.csv',index=False)
data = pd.read_csv('./simulation_results/Real_simulation/fivedones_Simulation.csv')
data.query("method=='GNAR' and mtype == 'rand'")['mse'].mean(),data.query("method=='GNAR' and mtype != 'rand'")['mse'].mean()
(1.4068299531936646, 1.4068299531936646)
data.query("method=='STGCN' and mtype == 'rand'")['mse'].mean(),data.query("method=='STGCN' and mtype != 'rand'")['mse'].mean()
(1.256219128270944, 3.429857851266861)
data.query("method=='IT-STGCN' and mtype == 'rand'")['mse'].mean(),data.query("method=='IT-STGCN' and mtype != 'rand'")['mse'].mean()
(1.223042539258798, 2.4890875375270842)

Baseline

data.query("method!='GNAR' and mrate==0").plot.box(backend='plotly',x='mrate',color='method',y='mse',facet_col='nof_filters',height=600)
data.query("method!='GNAR' and mtype =='rand' and (mrate==0.7 or mrate==0.75)").plot.box(backend='plotly',x='mrate',color='method',y='mse',facet_col='nof_filters',facet_row='inter_method',height=600)
data.query("method!='GNAR' and mtype =='rand'  and (mrate==0.8 or mrate==0.85)").plot.box(backend='plotly',x='mrate',color='method',y='mse',facet_col='nof_filters',facet_row='inter_method',height=600)
data.query("method!='GNAR' and mtype =='block' and inter_method=='linear' ").plot.box(backend='plotly',x='mrate',color='method',y='mse',facet_col='nof_filters',facet_row='inter_method',height=600)

chickenpox

random

  • 공식 패키지: lags 4 지정
  • mrate = 0.3
    • 결측값 비율 크니까 오차 많이 커지는 경향 있어서
  • nof_filters = 4
    • 차이 없어서
  • lags = 4, 8
    • 클 수록 작아지는 경향 있어서
  • GNAR보다 MSE는 낮음
  • cal_time
    • mean = 10
    • max = 21
  • block 은 임의로 한 노드만 해 본 결과임
data = pd.read_csv('./simulation_results/chickenpox_random.csv').sort_values(by='lags')
data.query("method!='GNAR'")['calculation_time'].mean(),data.query("method!='GNAR'")['calculation_time'].max(),data.query("method!='GNAR'")['calculation_time'].min()
(10.42619569649299, 21.886654376983643, 7.567165851593018)
data.query("method!='GNAR' and inter_method=='cubic' and mrate==0.3").plot.box(backend='plotly',x='mrate',color='method',y='mse',facet_col='lags',facet_row='nof_filters',height=1200)
data.query("method=='GNAR' and inter_method=='linear'").plot.box(backend='plotly',x='mrate',color='method',y='mse',facet_col='lags',height=600)

시뮬 예정(평균 시간, 평균mse)

epoch = 50

mrate = 0.3~0.5

filter 32 공식예제로 가기 하고 샆으면 3개 정도 추가로

# 1. mrate = 0.3, filter = 4, epoch = 50, lags = 4
data.query("method !='GNAR' and mrate==0.3 and inter_method=='cubic' and nof_filters==4 and lags==2")['calculation_time'].mean(),data.query("method != 'GNAR' and mrate==0.3 and inter_method=='cubic' and nof_filters==4 and lags==2")['mse'].mean()
(10.115000387032827, 1.0320488701264063)
data.query("method !='GNAR' and mrate==0.3 and inter_method=='cubic' and nof_filters==4 and lags==2").plot.box(backend='plotly',x='mrate',color='method',y='mse',facet_col='lags',facet_row='nof_filters',height=400)

block

data = pd.read_csv('./simulation_results/chickenpox_block.csv')
data.query("method != 'GNAR' and lags!=4 and lags!=6 and inter_method !='linear'").plot.box(backend='plotly',x='nof_filters',color='method',y='mse',facet_col='lags',facet_row='inter_method',height=600)
data.query("method=='GNAR'").plot.box(backend='plotly',x='mrate',color='inter_method',y='mse',facet_col='lags',height=600)

시뮬 예정(평균 시간, 평균mse)

block, rand 다

공식예제 수 따라

epoch 50

나중에 시간 남으면 100

data.query("inter_method=='cubic' and nof_filters==4 and lags==8").plot.box(backend='plotly',x='mrate',color='method',y='mse',facet_col='lags',facet_row='nof_filters',height=400)

chikenpox simulation result

my_list = [[] for _ in range(20)] #chickenpox
another_list = list(range(100,400))
my_list[1] = another_list
my_list[3] = another_list
my_list[5] = another_list
my_list[7] = another_list
my_list[9] = another_list
my_list[11] = another_list
my_list[13] = another_list
my_list[15] = another_list
mindex = my_list

block 30% missing을 위한 조건

df1 = pd.read_csv('./simulation_results/2023-04-11_06-56-35.csv') # GNAR random
df2 = pd.read_csv('./simulation_results/2023-04-11_07-01-42.csv') # GNAR block
df3 = pd.read_csv('./simulation_results/2023-04-11_18-20-22.csv') # STGCN, ITSTGCN random 30%
df4 = pd.read_csv('./simulation_results/2023-04-12_05-44-19.csv') # STGCN, ITSTGCN random 40%
df5 = pd.read_csv('./simulation_results/2023-04-12_17-03-28.csv') # STGCN, ITSTGCN random 50%
df6 = pd.read_csv('./simulation_results/2023-04-13_18-59-17.csv') # STGCN, ITSTGCN block cubic
df7 = pd.read_csv('./simulation_results/2023-04-14_00-57-11.csv') # STGCN, ITSTGCN block linear
df8 = pd.read_csv('./simulation_results/2023-04-14_12-55-58.csv') # STGCN, ITSTGCN 0% missing
data = pd.concat([df1,df2,df3,df4,df5,df6,df7,df8],axis=0);data
dataset method mrate mtype lags nof_filters inter_method epoch mse calculation_time
0 chickenpox GNAR 0.3 rand 4 NaN linear NaN 1.427494 0.070639
1 chickenpox GNAR 0.3 rand 4 NaN cubic NaN 1.427494 0.072070
2 chickenpox GNAR 0.4 rand 4 NaN linear NaN 1.427494 0.087900
3 chickenpox GNAR 0.4 rand 4 NaN cubic NaN 1.427494 0.094206
4 chickenpox GNAR 0.5 rand 4 NaN linear NaN 1.427494 0.096730
... ... ... ... ... ... ... ... ... ... ...
355 chickenpox IT-STGCN 0.0 NaN 4 16.0 NaN 50.0 1.014494 116.511157
356 chickenpox IT-STGCN 0.0 NaN 4 24.0 NaN 50.0 1.001220 117.726670
357 chickenpox IT-STGCN 0.0 NaN 4 24.0 NaN 50.0 1.002661 117.931265
358 chickenpox IT-STGCN 0.0 NaN 4 32.0 NaN 50.0 1.017814 123.757436
359 chickenpox IT-STGCN 0.0 NaN 4 32.0 NaN 50.0 1.014889 124.369595

1986 rows × 10 columns

data.to_csv('./simulation_results/Real_simulation/chikenpox_Simulation.csv',index=False)
data = pd.read_csv('./simulation_results/Real_simulation/chikenpox_Simulation.csv')
data.query("method=='GNAR' and mtype == 'rand'")['mse'].mean(),data.query("method=='GNAR' and mtype != 'rand'")['mse'].mean()
(1.4274942874908447, 1.4274942874908447)
data.query("method=='STGCN' and mtype == 'rand'")['mse'].mean(),data.query("method=='STGCN' and mtype != 'rand'")['mse'].mean()
(1.0746942153683414, 1.0201175289021598)
data.query("method=='IT-STGCN' and mtype == 'rand'")['mse'].mean(),data.query("method=='IT-STGCN' and mtype != 'rand'")['mse'].mean()
(1.0245469553603066, 1.0210112863116794)
data.query("method!='GNAR' and mrate ==0 ").plot.box(backend='plotly',x='mrate',color='method',y='mse',facet_col='nof_filters',height=600)
data.query("method!='GNAR' and mtype =='rand' ").plot.box(backend='plotly',x='mrate',color='method',y='mse',facet_col='nof_filters',facet_row='inter_method',height=800)
data.query("method!='GNAR' and mtype =='block' ").plot.box(backend='plotly',x='mrate',color='method',y='mse',facet_col='nof_filters',facet_row='inter_method',height=800)

Pedalme

데이터셋 미싱이 0.8일때

The number of derivatives at boundaries does not match: expected 1, got 0+0

해당 오류, 즉 미분되지 않는 오류가 생긴다.

  • 공식 패키지: lags 4 지정
  • mrate = 0.3
  • nof_filters = 12
    • 필터 클수록 mse 안정적으로 보임!
  • lags = 4, 8
    • lags 클 수록 커지는 경향이 있지만,
    • lags 크니까 GNAR보다 mse 평균적으로 낮게 보인다.
  • GNAR보다 MSE는 낮음
  • cal_time
    • mean = 1초도 안 된다!
    • max = 6초!

random

data = pd.read_csv('./simulation_results/pedalme_random.csv');data
dataset method mrate mtype lags nof_filters inter_method epoch mse calculation_time
0 pedalme GNAR 0.0 NaN 2 NaN NaN NaN 1.151634 0.009696
1 pedalme IT-STGCN 0.0 NaN 2 4.0 NaN 5.0 1.255845 1.595633
2 pedalme IT-STGCN 0.0 NaN 2 4.0 NaN 5.0 1.216815 1.721916
3 pedalme IT-STGCN 0.0 NaN 2 12.0 NaN 5.0 1.249243 1.420279
4 pedalme IT-STGCN 0.0 NaN 2 12.0 NaN 5.0 1.237032 0.910204
... ... ... ... ... ... ... ... ... ... ...
2527 pedalme STGCN 0.3 rand 8 4.0 linear 5.0 1.429247 0.560238
2528 pedalme STGCN 0.3 rand 8 4.0 cubic 5.0 1.431005 0.564748
2529 pedalme IT-STGCN 0.7 rand 8 12.0 linear 5.0 1.489404 0.543194
2530 pedalme IT-STGCN 0.3 rand 8 4.0 cubic 5.0 1.372652 0.628229
2531 pedalme GNAR 0.7 rand 8 NaN linear NaN 1.382030 0.020379

2532 rows × 10 columns

data.query("method!='GNAR'")['calculation_time'].mean(),data.query("method!='GNAR'")['calculation_time'].max(),data.query("method!='GNAR'")['calculation_time'].min()
(0.9260026927347537, 6.205296277999878, 0.4241352081298828)
data.query("mtype=='rand' and method!='GNAR' and lags!=4").plot.box(backend='plotly',x='mrate',color='method',y='mse',facet_col='lags',facet_row='nof_filters',height=1200)
data.query("mtype=='rand' and method!='GNAR' and mrate==0.7").plot.box(backend='plotly',x='inter_method',color='method',y='mse',facet_col='lags',facet_row='nof_filters',height=1200)
data.query("mtype=='rand' and method!='GNAR' and mrate==0.3").plot.box(backend='plotly',x='inter_method',color='method',y='mse',facet_col='lags',facet_row='nof_filters',height=1200)
data.query("method=='GNAR'").plot.box(backend='plotly',x='mtype',color='method',y='mse',facet_col='lags',height=600)

시뮬 예정(평균 시간, 평균mse)

lags 4,8

mrate 0.3~0.6

# 1. mrate = 0.3, filter = 4, epoch = 50, lags = 4
data.query("method !='GNAR' and mrate==0.3 and inter_method=='cubic' and nof_filters==12 and lags==8")['calculation_time'].mean(),data.query("method !='GNAR' and mrate==0.3 and inter_method=='cubic' and nof_filters==12 and lags==8")['mse'].mean()
(0.8366350531578064, 1.3687758445739746)
data.query("method !='GNAR' and mrate==0.3 and inter_method=='cubic' and nof_filters==12 and lags==8").plot.box(backend='plotly',x='mrate',color='method',y='mse',facet_col='lags',facet_row='nof_filters',height=400)

block

data = pd.read_csv('./simulation_results/pedalme_block.csv');data
dataset method mrate mtype lags nof_filters inter_method epoch mse calculation_time
0 pedalme IT-STGCN 0.047619 block 2 4.0 cubic 5.0 1.229210 0.758090
1 pedalme STGCN 0.047619 block 2 12.0 linear 5.0 1.223644 0.681700
2 pedalme STGCN 0.047619 block 2 12.0 cubic 5.0 1.237086 0.684113
3 pedalme STGCN 0.047619 block 2 4.0 linear 5.0 1.225114 0.659210
4 pedalme STGCN 0.047619 block 2 4.0 cubic 5.0 1.216191 0.664208
... ... ... ... ... ... ... ... ... ... ...
715 pedalme IT-STGCN 0.045977 block 8 4.0 cubic 5.0 1.425474 0.640063
716 pedalme STGCN 0.045977 block 8 12.0 cubic 5.0 1.302402 0.718187
717 pedalme STGCN 0.045977 block 8 12.0 linear 5.0 1.336038 0.719500
718 pedalme IT-STGCN 0.045977 block 8 12.0 linear 5.0 1.311962 0.831888
719 pedalme IT-STGCN 0.045977 block 8 12.0 cubic 5.0 1.315647 0.667004

720 rows × 10 columns

missing rate 조정하기 30~50% 여러개 block 해서

data.query("method!='GNAR'").plot.box(backend='plotly',x='inter_method',color='method',y='mse',facet_col='lags',facet_row='nof_filters',height=1200)

시뮬 예정(평균 시간, 평균mse)

data.query("inter_method=='linear' and nof_filters==12 and lags==4").plot.box(backend='plotly',x='mrate',color='method',y='mse',facet_col='lags',facet_row='nof_filters',height=400)

pedalme simulation result

block1

my_list = [[] for _ in range(15)] #pedalme
another_list = list(range(5,25))
my_list[1] = another_list
my_list[3] = another_list
my_list[5] = another_list
my_list[7] = another_list
my_list[9] = another_list
my_list[11] = another_list
mindex = my_list

block 30% missing을 위한 조건

block2

my_list = [[] for _ in range(15)] #pedalme
another_list = list(range(10,25))
my_list[2] = another_list
my_list[4] = another_list
my_list[5] = another_list
my_list[11] = another_list
mindex = my_list

block 30% missing을 위한 조건

df1 = pd.read_csv('./simulation_results/2023-04-13_20-37-59.csv') # STGCN, ITSTGCN random 0%,30%, 40%
df2 = pd.read_csv('./simulation_results/2023-04-13_21-29-38.csv') # STGCN, ITSTGCN random 50%, 60%
df3 = pd.read_csv('./simulation_results/2023-04-15_01-08-16.csv') # GNAR random 30%, 40%, 50%, 60%
df4 = pd.read_csv('./simulation_results/2023-04-13_21-56-36.csv') # GNAR block 30%
df5 = pd.read_csv('./simulation_results/2023-04-15_01-08-16.csv') # STGCN, ITSTGCN block 1 
df6 = pd.read_csv('./simulation_results/2023-04-15_01-38-46.csv') # STGCN, ITSTGCN block 2
# df7 = pd.read_csv('./simulation_results/2023-04-11_04-40-00.csv') # STGCN, ITSTGCN block linear
data = pd.concat([df1,df2,df3,df4,df5,df6],axis=0);data
dataset method mrate mtype lags nof_filters inter_method epoch mse calculation_time
0 pedalme STGCN 0.300000 rand 4 12.0 cubic 50.0 1.105710 6.291031
1 pedalme STGCN 0.300000 rand 4 12.0 linear 50.0 1.589615 6.157079
2 pedalme STGCN 0.300000 rand 8 12.0 cubic 50.0 1.750380 5.378390
3 pedalme STGCN 0.300000 rand 8 12.0 linear 50.0 1.359417 5.414476
4 pedalme STGCN 0.000000 NaN 4 12.0 NaN 50.0 1.247342 6.153305
... ... ... ... ... ... ... ... ... ... ...
115 pedalme IT-STGCN 0.137931 block 8 12.0 cubic 50.0 1.272855 6.657242
116 pedalme STGCN 0.142857 block 4 12.0 cubic 50.0 1.165904 6.297507
117 pedalme STGCN 0.137931 block 8 12.0 cubic 50.0 1.285807 5.538206
118 pedalme IT-STGCN 0.142857 block 4 12.0 cubic 50.0 1.270656 7.295285
119 pedalme IT-STGCN 0.137931 block 8 12.0 cubic 50.0 1.463783 6.675699

1692 rows × 10 columns

data.to_csv('./simulation_results/Real_simulation/pedalme_Simulation.csv',index=False)
data = pd.read_csv('./simulation_results/Real_simulation/pedalme_Simulation.csv')
data.query("method=='GNAR' and mtype == 'rand'")['mse'].mean(),data.query("method=='GNAR' and mtype != 'rand'")['mse'].mean()
(1.3026788234710691, 1.34235417842865)
data.query("method=='STGCN' and mtype == 'rand'")['mse'].mean(),data.query("method=='STGCN' and mtype != 'rand'")['mse'].mean()
(1.415487505743901, 1.3578179611100092)
data.query("method=='IT-STGCN' and mtype == 'rand'")['mse'].mean(),data.query("method=='IT-STGCN' and mtype != 'rand'")['mse'].mean()
(1.4283250387758017, 1.336261150572035)

Baseline

data.query("method!='GNAR' and mrate ==0 ").plot.box(backend='plotly',x='epoch',color='method',y='mse',facet_col='lags',height=800)
data.query("method!='GNAR' and mtype =='rand' ").plot.box(backend='plotly',x='mrate',color='method',y='mse',facet_col='lags',facet_row='inter_method',height=800)
data.query("method!='GNAR' and mtype =='block' ").plot.box(backend='plotly',x='mrate',color='method',y='mse',facet_col='lags',facet_row='inter_method',height=600)

Wikimath

  • 공식 패키지: lags 8 지정
  • 오히려 cuboic 보다 linear가 더 잘 맞추는 경향
  • mrate = 0.3
    • 크면 이상치 심하게 나와서 작게 잡기
  • nof_filters = 12, 16
    • 필터 크니까 mse 내려감
  • lags = 2,
  • cal_time
    • mean = 71s
    • max = 212s

random

df1 = pd.read_csv('./simulation_results/2023-04-15_16-58-03.csv')
df2 = pd.read_csv('./simulation_results/2023-04-15_17-01-39.csv')
df3 = pd.read_csv('./simulation_results/2023-04-15_17-07-23.csv')
df4 = pd.read_csv('./simulation_results/2023-04-15_17-13-13.csv')
df5 = pd.read_csv('./simulation_results/2023-04-15_17-29-49.csv')
data = pd.concat([df1,df2,df3,df4,df5],axis=0)
data.query("method=='STGCN'").sort_values(['mrate','lags','nof_filters'])
dataset method mrate mtype lags nof_filters inter_method epoch mse calculation_time
0 wikimath STGCN 0.3 rand 4 12 linear 1 0.863623 25.504817
0 wikimath STGCN 0.3 rand 4 12 cubic 1 0.847675 27.086116
0 wikimath STGCN 0.4 rand 2 12 linear 1 0.912734 30.048937
1 wikimath STGCN 0.4 rand 2 12 cubic 1 0.916843 27.104823
0 wikimath STGCN 0.4 rand 4 12 linear 1 0.907305 24.776503
1 wikimath STGCN 0.4 rand 4 12 cubic 1 0.854127 24.608104
2 wikimath STGCN 0.4 rand 8 12 linear 1 0.788011 24.233431
3 wikimath STGCN 0.4 rand 8 12 cubic 1 0.795219 24.228026
0 wikimath STGCN 0.5 rand 4 12 linear 1 0.914080 26.301605
1 wikimath STGCN 0.5 rand 4 12 cubic 1 0.975948 27.855870 
data.query("method!='STGCN'").sort_values(['mrate','lags','nof_filters'])
dataset method mrate mtype lags nof_filters inter_method epoch mse calculation_time
1 wikimath IT-STGCN 0.3 rand 4 12 linear 1 0.908916 28.928112
1 wikimath IT-STGCN 0.3 rand 4 12 cubic 1 0.856639 29.759748
4 wikimath IT-STGCN 0.4 rand 2 12 linear 1 0.864580 29.660712
5 wikimath IT-STGCN 0.4 rand 2 12 cubic 1 0.926426 30.838968
2 wikimath IT-STGCN 0.4 rand 4 12 linear 1 0.871146 29.008776
3 wikimath IT-STGCN 0.4 rand 4 12 cubic 1 0.905354 30.405766
6 wikimath IT-STGCN 0.4 rand 8 12 linear 1 0.822462 32.329447
7 wikimath IT-STGCN 0.4 rand 8 12 cubic 1 0.817621 29.447260
2 wikimath IT-STGCN 0.5 rand 4 12 linear 1 0.878943 31.140878
3 wikimath IT-STGCN 0.5 rand 4 12 cubic 1 1.002361 28.461372 
data.query("method!='GNAR'")['calculation_time'].mean(),data.query("method!='GNAR'")['calculation_time'].max(),data.query("method!='GNAR'")['calculation_time'].min()
data.query("mtype=='rand' and mrate != 0.9 and method!='GNAR' and inter_method=='cubic'").plot.box(backend='plotly',x='mrate',color='method',y='mse',facet_col='lags',facet_row='nof_filters',height=1200)
data.query("mtype=='rand' and method!='GNAR' and inter_method=='linear'").plot.box(backend='plotly',x='mrate',color='method',y='mse',facet_col='lags',facet_row='nof_filters',height=1200)

시뮬 예정(평균 시간, 평균mse)

data.query("method !='GNAR' and mrate==0.3 and inter_method=='cubic' and nof_filters==12 and lags==8")['calculation_time'].mean(),data.query("method !='GNAR' and mrate==0.3 and inter_method=='cubic' and nof_filters==12 and lags==8")['mse'].mean()
data.query("method !='GNAR' and mrate==0.3 and inter_method=='linear' and nof_filters==12 and lags==8")['calculation_time'].mean(),data.query("method !='GNAR' and mrate==0.3 and inter_method=='linear' and nof_filters==12 and lags==8")['mse'].mean()
data.query("method !='GNAR' and mrate==0.3 and nof_filters==12 and lags==8").plot.box(backend='plotly',x='mrate',color='method',y='mse',facet_col='lags',facet_row='nof_filters',height=400)

block

data = pd.read_csv('./simulation_results/wiki_block.csv');data
data.query("method!='GNAR'")['calculation_time'].mean(),data.query("method!='GNAR'")['calculation_time'].max(),data.query("method!='GNAR'")['calculation_time'].min()
data.query("method!='GNAR' and inter_method=='cubic'").plot.box(backend='plotly',x='mrate',color='method',y='mse',facet_col='lags',facet_row='nof_filters',height=1200)
data.query("inter_method=='linear'").plot.box(backend='plotly',x='mrate',color='method',y='mse',facet_col='lags',facet_row='nof_filters',height=1200)

시뮬 예정(평균 시간, 평균mse)

data.query("inter_method=='linear' and nof_filters==12 and lags==8")['calculation_time'].mean(),data.query("inter_method=='linear' and nof_filters==12 and lags==8")['mse'].mean()
data.query("inter_method=='linear' and nof_filters==12 and lags==8").plot.box(backend='plotly',x='mrate',color='method',y='mse',facet_col='lags',facet_row='nof_filters',height=400)

Windmilmedium

df1 = pd.read_csv('./simulation_results/2023-04-15_09-06-12.csv') # GNAR
df2 = pd.read_csv('./simulation_results/2023-04-15_09-19-44.csv') # STGCN IT-STGCN
df3 = pd.read_csv('./simulation_results/2023-04-15_09-28-32.csv') # STGCN IT-STGCN
df4 = pd.read_csv('./simulation_results/2023-04-15_09-36-55.csv') # STGCN IT-STGCN
df5 = pd.read_csv('./simulation_results/2023-04-15_09-54-30.csv') # STGCN IT-STGCN
df6 = pd.read_csv('./simulation_results/2023-04-15_10-03-08.csv') # STGCN IT-STGCN
df7 = pd.read_csv('./simulation_results/2023-04-15_10-15-48.csv') # STGCN IT-STGCN
df8 = pd.read_csv('./simulation_results/2023-04-15_10-25-19.csv') # STGCN IT-STGCN
df9 = pd.read_csv('./simulation_results/2023-04-15_10-34-48.csv') # STGCN IT-STGCN
df10 = pd.read_csv('./simulation_results/2023-04-15_10-43-02.csv') # STGCN IT-STGCN
data = pd.concat([df1,df2,df3,df4,df5,df6,df7,df8,df9,df10],axis=0)
data.query("method=='GNAR'")
dataset method mrate mtype lags nof_filters inter_method epoch mse calculation_time
0 windmilmedium GNAR 0.1 rand 4 NaN cubic NaN 1.410524 3.150951
1 windmilmedium GNAR 0.1 rand 4 NaN cubic NaN 1.410524 2.926645
2 windmilmedium GNAR 0.1 rand 4 NaN cubic NaN 1.410524 2.406094 
data.query("method!='GNAR' and method=='STGCN'").sort_values(['mrate','nof_filters'])
dataset method mrate mtype lags nof_filters inter_method epoch mse calculation_time
0 windmilmedium STGCN 0.1 rand 2 12.0 cubic 1.0 1.038629 85.205226
0 windmilmedium STGCN 0.4 rand 2 12.0 cubic 1.0 1.325845 85.721810
0 windmilmedium STGCN 0.5 rand 4 12.0 cubic 1.0 1.568322 87.435629
0 windmilmedium STGCN 0.8 rand 2 4.0 linear 1.0 1.256851 79.950040
0 windmilmedium STGCN 0.8 rand 2 8.0 linear 1.0 1.259840 83.667886
0 windmilmedium STGCN 0.8 rand 2 12.0 cubic 1.0 1.874045 85.171979
0 windmilmedium STGCN 0.8 rand 2 12.0 linear 1.0 1.427910 84.167046
0 windmilmedium STGCN 0.8 rand 2 16.0 linear 1.0 1.449077 93.408642
0 windmilmedium STGCN 0.8 rand 2 32.0 linear 1.0 1.339085 88.795324 
data.query("method!='GNAR' and method!='STGCN'").sort_values(['mrate','nof_filters'])
dataset method mrate mtype lags nof_filters inter_method epoch mse calculation_time
1 windmilmedium IT-STGCN 0.1 rand 2 12.0 cubic 1.0 0.991687 282.159899
1 windmilmedium IT-STGCN 0.4 rand 2 12.0 cubic 1.0 1.150386 282.688800
1 windmilmedium IT-STGCN 0.5 rand 4 12.0 cubic 1.0 1.232030 273.543777
1 windmilmedium IT-STGCN 0.8 rand 2 4.0 linear 1.0 1.217331 295.937213
1 windmilmedium IT-STGCN 0.8 rand 2 8.0 linear 1.0 1.162310 259.498000
1 windmilmedium IT-STGCN 0.8 rand 2 12.0 cubic 1.0 1.772746 262.409024
1 windmilmedium IT-STGCN 0.8 rand 2 12.0 linear 1.0 1.182181 287.886245
1 windmilmedium IT-STGCN 0.8 rand 2 16.0 linear 1.0 1.430025 306.211073
1 windmilmedium IT-STGCN 0.8 rand 2 32.0 linear 1.0 1.203084 309.623751

linear, 0.7~0.9, 4,8,

data.to_csv('./simulation_results/Real_simulation/windmilmedium_Simulation.csv',index=False)
data.query("method!='GNAR'")['calculation_time'].mean(),data.query("method!='GNAR'")['calculation_time'].max(),data.query("method!='GNAR'")['calculation_time'].min()

random

block

Windmilsmall

df1 = pd.read_csv('./simulation_results/2023-04-15_10-59-28.csv') # GNAR
df2 = pd.read_csv('./simulation_results/2023-04-15_11-15-05.csv') # STGCN IT-STGCN
df3 = pd.read_csv('./simulation_results/2023-04-15_11-30-48.csv') # STGCN IT-STGCN
df4 = pd.read_csv('./simulation_results/2023-04-15_11-46-45.csv') # STGCN IT-STGCN
df5 = pd.read_csv('./simulation_results/2023-04-15_12-02-19.csv') # STGCN IT-STGCN
df6 = pd.read_csv('./simulation_results/2023-04-15_15-00-32.csv') # STGCN IT-STGCN
# df7 = pd.read_csv('./simulation_results/2023-04-15_10-15-48.csv') # STGCN IT-STGCN
# df8 = pd.read_csv('./simulation_results/2023-04-15_10-25-19.csv') # STGCN IT-STGCN
# df9 = pd.read_csv('./simulation_results/2023-04-15_10-34-48.csv') # STGCN IT-STGCN
# df10 = pd.read_csv('./simulation_results/2023-04-15_10-43-02.csv') # STGCN IT-STGCN
data = pd.concat([df1,df2,df3,df4,df5,df6],axis=0)
data.query("method=='GNAR'").sort_values(['mrate','nof_filters'])
dataset method mrate mtype lags nof_filters inter_method epoch mse calculation_time
0 windmilsmall GNAR 0.3 rand 4 NaN cubic NaN 1.640339 1.222944
1 windmilsmall GNAR 0.3 rand 4 NaN linear NaN 1.640339 0.954795
6 windmilsmall GNAR 0.3 rand 4 NaN cubic NaN 1.640339 0.761808
7 windmilsmall GNAR 0.3 rand 4 NaN linear NaN 1.640339 0.790776
12 windmilsmall GNAR 0.3 rand 4 NaN cubic NaN 1.640339 0.987039
13 windmilsmall GNAR 0.3 rand 4 NaN linear NaN 1.640339 0.976644
2 windmilsmall GNAR 0.5 rand 4 NaN cubic NaN 1.640339 0.901695
3 windmilsmall GNAR 0.5 rand 4 NaN linear NaN 1.640339 0.939233
8 windmilsmall GNAR 0.5 rand 4 NaN cubic NaN 1.640339 0.986533
9 windmilsmall GNAR 0.5 rand 4 NaN linear NaN 1.640339 0.988925
14 windmilsmall GNAR 0.5 rand 4 NaN cubic NaN 1.640339 0.692504
15 windmilsmall GNAR 0.5 rand 4 NaN linear NaN 1.640339 0.954345
4 windmilsmall GNAR 0.7 rand 4 NaN cubic NaN 1.640339 0.938546
5 windmilsmall GNAR 0.7 rand 4 NaN linear NaN 1.640339 0.926433
10 windmilsmall GNAR 0.7 rand 4 NaN cubic NaN 1.640339 0.830638
11 windmilsmall GNAR 0.7 rand 4 NaN linear NaN 1.640339 0.947263
16 windmilsmall GNAR 0.7 rand 4 NaN cubic NaN 1.640339 0.939917
17 windmilsmall GNAR 0.7 rand 4 NaN linear NaN 1.640339 0.950329 
data.query("method!='GNAR' and method=='STGCN'").sort_values(['mrate','nof_filters'])
dataset method mrate mtype lags nof_filters inter_method epoch mse calculation_time
0 windmilsmall STGCN 0.3 rand 4 12.0 linear 1.0 1.200871 73.057425
1 windmilsmall STGCN 0.3 rand 4 12.0 cubic 1.0 1.442769 82.757819
0 windmilsmall STGCN 0.5 rand 4 12.0 linear 1.0 1.332673 73.738043
1 windmilsmall STGCN 0.5 rand 4 12.0 cubic 1.0 1.935493 75.276489
0 windmilsmall STGCN 0.5 rand 8 12.0 linear 1.0 1.287041 72.605574
1 windmilsmall STGCN 0.5 rand 8 12.0 cubic 1.0 1.954025 72.804517
0 windmilsmall STGCN 0.5 rand 4 32.0 linear 1.0 1.320776 74.728671
1 windmilsmall STGCN 0.5 rand 4 32.0 cubic 1.0 1.865486 75.110273
0 windmilsmall STGCN 0.7 rand 4 12.0 linear 1.0 1.470515 74.648811
1 windmilsmall STGCN 0.7 rand 4 12.0 cubic 1.0 2.182147 71.730788 
data.query("method!='GNAR' and method!='STGCN'").sort_values(['mrate','nof_filters'])
dataset method mrate mtype lags nof_filters inter_method epoch mse calculation_time
2 windmilsmall IT-STGCN 0.3 rand 4 12.0 linear 1.0 0.997696 255.592267
3 windmilsmall IT-STGCN 0.3 rand 4 12.0 cubic 1.0 1.049607 268.626712
2 windmilsmall IT-STGCN 0.5 rand 4 12.0 linear 1.0 1.103726 253.085496
3 windmilsmall IT-STGCN 0.5 rand 4 12.0 cubic 1.0 1.459947 278.678849
2 windmilsmall IT-STGCN 0.5 rand 8 12.0 linear 1.0 1.110951 266.898954
3 windmilsmall IT-STGCN 0.5 rand 8 12.0 cubic 1.0 1.314662 270.144128
2 windmilsmall IT-STGCN 0.5 rand 4 32.0 linear 1.0 1.153238 259.596158
3 windmilsmall IT-STGCN 0.5 rand 4 32.0 cubic 1.0 1.341482 253.815536
2 windmilsmall IT-STGCN 0.7 rand 4 12.0 linear 1.0 1.215104 283.833449
3 windmilsmall IT-STGCN 0.7 rand 4 12.0 cubic 1.0 2.141516 279.526782

linear, lags 4,8. filter 8,12

random

block