DCRNN_Simulation Tables_reshape

ITSTGCN
Author

SEOYEON CHOI

Published

June 13, 2023

Simulation Tables

import

import pandas as pd
data_fivenodes = pd.read_csv('./simulation_results/Real_simulation_reshape/DCRNN_fivedones_Simulation.csv')
data_chickenpox = pd.read_csv('./simulation_results/Real_simulation_reshape/DCRNN_chikenpox_Simulation.csv')
data_pedal = pd.read_csv('./simulation_results/Real_simulation_reshape/DCRNN_pedalme_Simulation.csv')
data_pedal2 = pd.read_csv('./simulation_results/Real_simulation_reshape/DCRNN_pedalme_Simulation_itstgcnsnd.csv')
data__wiki = pd.read_csv('./simulation_results/Real_simulation_reshape/DCRNN_wikimath.csv')
data_wiki_GSO = pd.read_csv('./simulation_results/Real_simulation_reshape/DCRNN_wikimath_GSO_st.csv')
data_windmillsmall = pd.read_csv('./simulation_results/Real_simulation_reshape/DCRNN_windmillsmall.csv')
data_monte = pd.read_csv('./simulation_results/Real_simulation_reshape/DCRNN_monte.csv')
data = pd.concat([data_fivenodes,data_chickenpox,data_pedal,data__wiki,data_windmillsmall,data_monte]);data
dataset method mrate mtype lags nof_filters inter_method epoch mse calculation_time
0 fivenodes STGCN 0.7 rand 2 2 linear 50 1.282193 13.957623
1 fivenodes STGCN 0.7 rand 2 2 nearest 50 1.268960 14.714340
2 fivenodes IT-STGCN 0.7 rand 2 2 linear 50 1.258667 20.865089
3 fivenodes IT-STGCN 0.7 rand 2 2 nearest 50 1.227342 19.442389
4 fivenodes STGCN 0.7 rand 2 2 linear 50 1.259042 15.236686
... ... ... ... ... ... ... ... ... ... ...
415 monte IT-STGCN 0.7 rand 4 12 nearest 50 1.066786 688.814203
416 monte STGCN 0.7 rand 4 12 nearest 50 1.148506 449.724371
417 monte IT-STGCN 0.7 rand 4 12 nearest 50 1.048256 815.432926
418 monte STGCN 0.7 rand 4 12 nearest 50 1.167802 709.441629
419 monte IT-STGCN 0.7 rand 4 12 nearest 50 1.012204 684.955110

2940 rows × 10 columns

data.to_csv('./simulation_results/Real_simulation_reshape/Final_Simulation_DCRNN.csv',index=False)
pedal_wiki_GSO = pd.concat([data_pedal2,data_wiki_GSO])
pedal_wiki_GSO.to_csv('./simulation_results/Real_simulation_reshape/Final_Simulation_DCRNN_pedal_wiki_GSO.csv',index=False)

Fivenodes

Baseline

pd.merge(data.query("dataset=='fivenodes' and mtype!='rand' and mtype!='block'").groupby(['nof_filters','method','lags'])['mse'].mean().reset_index(),
         data.query("dataset=='fivenodes' and mtype!='rand' and mtype!='block'").groupby(['nof_filters','method','lags'])['mse'].std().reset_index(),
         on=['method','nof_filters','lags']).rename(columns={'mse_x':'mean','mse_y':'std'}).round(3)
nof_filters method lags mean std
0 2 IT-STGCN 2 1.228 0.041
1 2 STGCN 2 1.230 0.042

Random

pd.merge(data.query("dataset=='fivenodes' and mtype=='rand'").groupby(['mrate','nof_filters','method','lags'])['mse'].mean().reset_index(),
         data.query("dataset=='fivenodes' and mtype=='rand'").groupby(['mrate','nof_filters','method','lags'])['mse'].std().reset_index(),
         on=['method','nof_filters','mrate','lags']).rename(columns={'mse_x':'mean','mse_y':'std'}).round(3).query("nof_filters==12")
mrate nof_filters method lags mean std

Block

pd.merge(data.query("dataset=='fivenodes' and mtype=='block'").groupby(['mrate','nof_filters','method'])['mse'].mean().reset_index(),
         data.query("dataset=='fivenodes' and mtype=='block'").groupby(['mrate','nof_filters','method'])['mse'].std().reset_index(),
         on=['method','nof_filters','mrate']).rename(columns={'mse_x':'mean','mse_y':'std'}).round(3)
mrate nof_filters method mean std
0 0.125 2 IT-STGCN 1.227 0.030
1 0.125 2 STGCN 1.254 0.046

ChickenpoxDatasetLoader(lags=4)

Baseline

pd.merge(data.query("dataset=='chickenpox' and mtype!='rand' and mtype!='block'").groupby(['nof_filters','method'])['mse'].mean().reset_index(),
         data.query("dataset=='chickenpox' and mtype!='rand' and mtype!='block'").groupby(['nof_filters','method'])['mse'].std().reset_index(),
         on=['method','nof_filters']).rename(columns={'mse_x':'mean','mse_y':'std'}).round(3).query("nof_filters==16")
nof_filters method mean std
0 16 IT-STGCN 0.726 0.007
1 16 STGCN 0.727 0.011

Random

pd.merge(data.query("dataset=='chickenpox' and mtype=='rand'").groupby(['mrate','inter_method','nof_filters','method'])['mse'].mean().reset_index(),
         data.query("dataset=='chickenpox' and mtype=='rand'").groupby(['mrate','inter_method','nof_filters','method'])['mse'].std().reset_index(),
         on=['method','inter_method','mrate','nof_filters']).rename(columns={'mse_x':'mean','mse_y':'std'}).round(3)
mrate inter_method nof_filters method mean std
0 0.3 linear 16 IT-STGCN 0.797 0.010
1 0.3 linear 16 STGCN 1.032 0.039
2 0.8 linear 16 IT-STGCN 1.467 0.076
3 0.8 linear 16 STGCN 2.287 0.074

Block

pd.merge(data.query("dataset=='chickenpox' and mtype=='block'").groupby(['inter_method','mrate','nof_filters','method'])['mse'].mean().reset_index(),
         data.query("dataset=='chickenpox' and mtype=='block'").groupby(['inter_method','mrate','nof_filters','method'])['mse'].std().reset_index(),
         on=['method','inter_method','mrate','nof_filters']).rename(columns={'mse_x':'mean','mse_y':'std'})
inter_method mrate nof_filters method mean std
0 linear 0.28777 16 IT-STGCN 0.739812 0.007356
1 linear 0.28777 16 STGCN 0.812195 0.006422
2 nearest 0.28777 16 IT-STGCN 0.738336 0.007345
3 nearest 0.28777 16 STGCN 0.832292 0.009452

PedalMeDatasetLoader (lags=4)

Baseline

pd.merge(data.query("dataset=='pedalme' and mtype!='rand' and mtype!='block'").groupby(['lags','nof_filters','method'])['mse'].mean().reset_index(),
         data.query("dataset=='pedalme' and mtype!='rand' and mtype!='block'").groupby(['lags','nof_filters','method'])['mse'].std().reset_index(),
         on=['method','lags','nof_filters']).rename(columns={'mse_x':'mean','mse_y':'std'}).round(3).query("lags==4")
lags nof_filters method mean std
0 4 8 IT-STGCN 1.131 0.015
1 4 8 STGCN 1.131 0.015

Random

pd.merge(data.query("dataset=='pedalme' and mtype=='rand'").groupby(['mrate','lags','inter_method','method'])['mse'].mean().reset_index(),
         data.query("dataset=='pedalme' and mtype=='rand'").groupby(['mrate','lags','inter_method','method'])['mse'].std().reset_index(),
         on=['method','mrate','lags','inter_method']).rename(columns={'mse_x':'mean','mse_y':'std'}).round(3)
mrate lags inter_method method mean std
0 0.3 4 linear IT-STGCN 1.190 0.029
1 0.3 4 linear STGCN 1.277 0.064
2 0.3 4 nearest IT-STGCN 1.179 0.035
3 0.3 4 nearest STGCN 1.278 0.060
4 0.6 4 linear IT-STGCN 1.314 0.072
5 0.6 4 linear STGCN 1.551 0.092
6 0.6 4 nearest IT-STGCN 1.303 0.078
7 0.6 4 nearest STGCN 1.509 0.068

Block

pd.merge(data.query("dataset=='pedalme' and mtype=='block'").groupby(['mrate','lags','inter_method','method'])['mse'].mean().reset_index(),
         data.query("dataset=='pedalme' and mtype=='block'").groupby(['mrate','lags','inter_method','method'])['mse'].std().reset_index(),
         on=['method','mrate','lags','inter_method']).rename(columns={'mse_x':'mean','mse_y':'std'}).round(3).query("lags==4")
mrate lags inter_method method mean std
0 0.286 4 linear IT-STGCN 1.154 0.014
1 0.286 4 linear STGCN 1.248 0.019
2 0.286 4 nearest IT-STGCN 1.150 0.014
3 0.286 4 nearest STGCN 1.304 0.021

W_st

pd.merge(data_pedal2.query("mtype=='rand'").groupby(['mrate','lags','inter_method','method'])['mse'].mean().reset_index(),
         data_pedal2.query("mtype=='rand'").groupby(['mrate','lags','inter_method','method'])['mse'].std().reset_index(),
         on=['method','mrate','lags','inter_method']).rename(columns={'mse_x':'mean','mse_y':'std'}).round(3).query("lags==4")
mrate lags inter_method method mean std
0 0.3 4 linear IT-STGCN 1.153 0.036
1 0.3 4 linear STGCN 1.263 0.053
2 0.3 4 nearest IT-STGCN 1.154 0.038
3 0.3 4 nearest STGCN 1.269 0.068
4 0.6 4 linear IT-STGCN 1.241 0.079
5 0.6 4 linear STGCN 1.506 0.065
6 0.6 4 nearest IT-STGCN 1.208 0.079
7 0.6 4 nearest STGCN 1.552 0.087 
pd.merge(data_pedal2.query("mtype=='block'").groupby(['mrate','lags','inter_method','method'])['mse'].mean().reset_index(),
         data_pedal2.query("mtype=='block'").groupby(['mrate','lags','inter_method','method'])['mse'].std().reset_index(),
         on=['method','mrate','lags','inter_method']).rename(columns={'mse_x':'mean','mse_y':'std'}).round(3).query("lags==4")
mrate lags inter_method method mean std
0 0.286 4 linear IT-STGCN 1.145 0.013
1 0.286 4 linear STGCN 1.295 0.019
2 0.286 4 nearest IT-STGCN 1.143 0.011
3 0.286 4 nearest STGCN 1.310 0.019

WikiMathsDatasetLoader (lags=8)

Baseline

pd.merge(data.query("dataset=='wikimath' and mrate==0").groupby(['lags','nof_filters','method'])['mse'].mean().reset_index(),
         data.query("dataset=='wikimath' and mrate==0").groupby(['lags','nof_filters','method'])['mse'].std().reset_index(),
         on=['lags','nof_filters','method']).rename(columns={'mse_x':'mean','mse_y':'std'}).round(3)
lags nof_filters method mean std
0 8 12 IT-STGCN 0.582 0.006
1 8 12 STGCN 0.580 0.006

Random

pd.merge(data.query("dataset=='wikimath' and mtype=='rand'").groupby(['mrate','lags','method'])['mse'].mean().reset_index(),
         data.query("dataset=='wikimath' and mtype=='rand'").groupby(['mrate','lags','method'])['mse'].std().reset_index(),
         on=['method','mrate','lags']).rename(columns={'mse_x':'mean','mse_y':'std'}).round(3)
mrate lags method mean std
0 0.3 8 IT-STGCN 0.588 0.007
1 0.3 8 STGCN 0.603 0.010
2 0.5 8 IT-STGCN 0.590 0.006
3 0.5 8 STGCN 0.652 0.015
4 0.6 8 IT-STGCN 0.592 0.005
5 0.6 8 STGCN 0.688 0.011
6 0.8 8 IT-STGCN 0.672 0.007
7 0.8 8 STGCN 0.846 0.031

Block

pd.merge(data.query("dataset=='wikimath' and mtype=='block'").groupby(['mrate','lags','method'])['mse'].mean().reset_index(),
         data.query("dataset=='wikimath' and mtype=='block'").groupby(['mrate','lags','method'])['mse'].std().reset_index(),
         on=['method','mrate','lags']).rename(columns={'mse_x':'mean','mse_y':'std'})
mrate lags method mean std
0 0.119837 8 IT-STGCN 0.582594 0.006427
1 0.119837 8 STGCN 0.578406 0.004975

missing values on the same nodes

pd.merge(data_wiki_GSO.groupby(['mrate','lags','method'])['mse'].mean().reset_index(),
        data_wiki_GSO.groupby(['mrate','lags','method'])['mse'].std().reset_index(),
         on=['method','mrate','lags']).rename(columns={'mse_x':'mean','mse_y':'std'}).round(3)
mrate lags method mean std
0 0.512 8 IT-STGCN 0.592 0.005
1 0.512 8 STGCN 0.665 0.015

WindmillOutputSmallDatasetLoader (lags=8)

Baseline

pd.merge(data.query("dataset=='windmillsmall' and mrate==0").groupby(['lags','method'])['mse'].mean().reset_index(),
         data.query("dataset=='windmillsmall' and mrate==0").groupby(['lags','method'])['mse'].std().reset_index(),
         on=['method','lags']).rename(columns={'mse_x':'mean','mse_y':'std'}).round(3)
lags method mean std
0 8 IT-STGCN 0.988 0.003
1 8 STGCN 0.987 0.002

Random

pd.merge(data.query("dataset=='windmillsmall' and mtype=='rand'").groupby(['mrate','lags','method'])['mse'].mean().reset_index(),
         data.query("dataset=='windmillsmall' and mtype=='rand'").groupby(['mrate','lags','method'])['mse'].std().reset_index(),
         on=['method','mrate','lags']).rename(columns={'mse_x':'mean','mse_y':'std'}).round(3)
mrate lags method mean std
0 0.7 8 IT-STGCN 1.117 0.034
1 0.7 8 STGCN 1.348 0.057

Block

pd.merge(data.query("dataset=='windmillsmall' and mtype=='block'").groupby(['mrate','lags','method'])['mse'].mean().reset_index(),
         data.query("dataset=='windmillsmall' and mtype=='block'").groupby(['mrate','lags','method'])['mse'].std().reset_index(),
         on=['method','mrate','lags']).rename(columns={'mse_x':'mean','mse_y':'std'}).round(3)
mrate lags method mean std
0 0.081 8 IT-STGCN 0.983 0.002
1 0.081 8 STGCN 0.994 0.005

Montevideobus (lags=4)

Baseline

pd.merge(data.query("dataset=='monte' and mrate==0").groupby(['lags','method'])['mse'].mean().reset_index(),
         data.query("dataset=='monte' and mrate==0").groupby(['lags','method'])['mse'].std().reset_index(),
         on=['method','lags']).rename(columns={'mse_x':'mean','mse_y':'std'}).round(3)
lags method mean std
0 4 IT-STGCN 0.936 0.002
1 4 STGCN 0.936 0.002

Random

pd.merge(data.query("dataset=='monte' and mtype=='rand'").groupby(['mrate','lags','inter_method','method'])['mse'].mean().reset_index(),
         data.query("dataset=='monte' and mtype=='rand'").groupby(['mrate','lags','inter_method','method'])['mse'].std().reset_index(),
         on=['mrate','inter_method','method','mrate','lags']).rename(columns={'mse_x':'mean','mse_y':'std'})
mrate lags inter_method method mean std
0 0.3 4 nearest IT-STGCN 0.937868 0.001369
1 0.3 4 nearest STGCN 1.007506 0.005817
2 0.5 4 nearest IT-STGCN 0.947735 0.002210
3 0.5 4 nearest STGCN 1.119226 0.024755
4 0.7 4 nearest IT-STGCN 1.030943 0.017530
5 0.7 4 nearest STGCN 1.214012 0.033267
6 0.8 4 nearest IT-STGCN 1.111060 0.036307
7 0.8 4 nearest STGCN 1.225077 0.072743

Block

pd.merge(data.query("dataset=='monte' and mtype=='block'").groupby(['mrate','lags','inter_method','method'])['mse'].mean().reset_index(),
         data.query("dataset=='monte' and mtype=='block'").groupby(['mrate','lags','inter_method','method'])['mse'].std().reset_index(),
         on=['method','mrate','inter_method','lags']).rename(columns={'mse_x':'mean','mse_y':'std'})
mrate lags inter_method method mean std
0 0.149142 4 nearest IT-STGCN 0.940344 0.001323
1 0.149142 4 nearest STGCN 0.955944 0.003010

Check

import itstgcnDCRNN
import torch
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
import random
class Eval_csy:
    def __init__(self,learner,train_dataset):
        self.learner = learner
        # self.learner.model.eval()
        try:self.learner.model.eval()
        except:pass
        self.train_dataset = train_dataset
        self.lags = self.learner.lags
        rslt_tr = self.learner(self.train_dataset) 
        self.X_tr = rslt_tr['X']
        self.y_tr = rslt_tr['y']
        self.f_tr = torch.concat([self.train_dataset[0].x.T,self.y_tr],axis=0).float()
        self.yhat_tr = rslt_tr['yhat']
        self.fhat_tr = torch.concat([self.train_dataset[0].x.T,self.yhat_tr],axis=0).float()
from plotnine import *
T = 500
t = np.arange(T)/T * 5

x = 1*np.sin(2*t)+np.sin(4*t)+1.5*np.sin(7*t)
eps_x  = np.random.normal(size=T)*0
y = x.copy()
for i in range(2,T):
    y[i] = 0.35*x[i-1] - 0.15*x[i-2] + 0.5*np.cos(0.4*t[i]) 
eps_y  = np.random.normal(size=T)*0
x = x
y = y
plt.plot(t,x,color='C0',lw=5)
plt.plot(t,x+eps_x,alpha=0.5,color='C0')
plt.plot(t,y,color='C1',lw=5)
plt.plot(t,y+eps_y,alpha=0.5,color='C1')
_node_ids = {'node1':0, 'node2':1}

_FX1 = np.stack([x+eps_x,y+eps_y],axis=1).tolist()

_edges1 = torch.tensor([[0,1]]).tolist()

data_dict1 = {'edges':_edges1, 'node_ids':_node_ids, 'FX':_FX1}

# save_data(data_dict1, './data/toy_example1.pkl')

data1 = pd.DataFrame({'x':x,'y':y,'xer':x,'yer':y})

# save_data(data1, './data/toy_example_true1.csv')

loader1 = itstgcnDCRNN.DatasetLoader(data_dict1)
dataset = loader1.get_dataset(lags=4)

mindex = itstgcn.rand_mindex(dataset,mrate=0) dataset_miss = itstgcn.miss(dataset,mindex,mtype=‘rand’)

mindex = [random.sample(range(0, T), int(T*0.8)),[np.array(list(range(20,30)))]]
dataset_miss = itstgcnDCRNN.miss(dataset,mindex,mtype='block')
/home/csy/Dropbox/blog/posts/GCN/itstgcnDCRNN/utils.py:71: UserWarning: Creating a tensor from a list of numpy.ndarrays is extremely slow. Please consider converting the list to a single numpy.ndarray with numpy.array() before converting to a tensor. (Triggered internally at /opt/conda/conda-bld/pytorch_1682343998658/work/torch/csrc/utils/tensor_new.cpp:245.)
dataset_padded = itstgcnDCRNN.padding(dataset_miss,interpolation_method='linear')

- 학습

lrnr = itstgcnDCRNN.StgcnLearner(dataset_padded)
lrnr.learn(filters=32,epoch=10)
10/10
lrnr1 = itstgcnDCRNN.ITStgcnLearner(dataset_padded)
lrnr1.learn(filters=16,epoch=10)
10/10
evtor = Eval_csy(lrnr,dataset_padded)
evtor1 = Eval_csy(lrnr1,dataset_padded)
with plt.style.context('seaborn-white'):
    fig, ax = plt.subplots(figsize=(20,10))
    
    ax.plot(evtor.f_tr[:,0],'--o',color='black',alpha=0.5,label='Imputation')
    ax.plot(data1['x'][:],'-',color='grey',label='Complete Data')
    ax.plot(evtor.fhat_tr[:,0],color='brown',lw=3,label='STGCN')
    ax.plot(evtor1.fhat_tr[:,0],color='blue',lw=3,label='ITSTGCN')
    
    ax.legend(fontsize=20,loc='lower left',facecolor='white', frameon=True)
    ax.tick_params(axis='y', labelsize=20)
    ax.tick_params(axis='x', labelsize=20)

with plt.style.context('seaborn-white'):
    fig, ax = plt.subplots(figsize=(20,10))
    
    ax.plot(evtor.f_tr[:,1],'--o',color='black',alpha=0.5,label='Imputation')
    ax.plot(data1['y'][:],'-',color='grey',label='Complete Data')
    ax.plot(evtor.fhat_tr[:,1],color='brown',lw=3,label='STGCN')
    ax.plot(evtor1.fhat_tr[:,1],color='blue',lw=3,label='ITSTGCN')
    
    ax.legend(fontsize=20,loc='lower left',facecolor='white', frameon=True)
    ax.tick_params(axis='y', labelsize=20)
    ax.tick_params(axis='x', labelsize=20)

import itstgcnsnd
import torch
import numpy as np
loader1 = itstgcnsnd.DatasetLoader(data_dict1)
dataset = loader1.get_dataset(lags=2)

mindex = itstgcn.rand_mindex(dataset,mrate=0) dataset_miss = itstgcn.miss(dataset,mindex,mtype=‘rand’)

mindex = [random.sample(range(0, T), int(T*0.5)),[np.array(list(range(20,30)))]]
dataset_miss = itstgcnsnd.miss(dataset,mindex,mtype='block')
/home/csy/Dropbox/blog/posts/GCN/itstgcnsnd/utils.py:71: UserWarning: Creating a tensor from a list of numpy.ndarrays is extremely slow. Please consider converting the list to a single numpy.ndarray with numpy.array() before converting to a tensor. (Triggered internally at /opt/conda/conda-bld/pytorch_1682343998658/work/torch/csrc/utils/tensor_new.cpp:245.)
dataset_padded = itstgcnsnd.padding(dataset_miss,interpolation_method='linear')

- 학습

lrnr = itstgcnsnd.StgcnLearner(dataset_padded)
lrnr.learn(filters=32,epoch=5)
5/5
lrnr1 = itstgcnsnd.ITStgcnLearner(dataset_padded)
lrnr1.learn(filters=32,epoch=5)
5/5
evtor = Eval_csy(lrnr,dataset_padded)
evtor1 = Eval_csy(lrnr1,dataset_padded)
with plt.style.context('seaborn-white'):
    fig, ax = plt.subplots(figsize=(20,10))
    
    ax.plot(evtor.f_tr[:,0],'--o',color='black',alpha=0.5,label='Imputation')
    ax.plot(data1['x'][:],'-',color='grey',label='Complete Data')
    ax.plot(evtor.fhat_tr[:,0],color='brown',lw=3,label='STGCN')
    ax.plot(evtor1.fhat_tr[:,0],color='blue',lw=3,label='ITSTGCN')
    
    ax.legend(fontsize=20,loc='lower left',facecolor='white', frameon=True)
    ax.tick_params(axis='y', labelsize=20)
    ax.tick_params(axis='x', labelsize=20)

with plt.style.context('seaborn-white'):
    fig, ax = plt.subplots(figsize=(20,10))
    
    ax.plot(evtor.f_tr[:,1],'--o',color='black',alpha=0.5,label='Imputation')
    ax.plot(data1['y'][:],'-',color='grey',label='Complete Data')
    ax.plot(evtor.fhat_tr[:,1],color='brown',lw=3,label='STGCN')
    ax.plot(evtor1.fhat_tr[:,1],color='blue',lw=3,label='ITSTGCN')
    
    ax.legend(fontsize=20,loc='lower left',facecolor='white', frameon=True)
    ax.tick_params(axis='y', labelsize=20)
    ax.tick_params(axis='x', labelsize=20)

hyperparameter

import itstgcn

data_dict = itstgcn.load_data('./data/fivenodes.pkl')
loader = itstgcn.DatasetLoader(data_dict)

from torch_geometric_temporal.dataset import ChickenpoxDatasetLoader
loader1 = ChickenpoxDatasetLoader()

from torch_geometric_temporal.dataset import PedalMeDatasetLoader
loader2 = PedalMeDatasetLoader()

from torch_geometric_temporal.dataset import WikiMathsDatasetLoader
loader3 = WikiMathsDatasetLoader()

from torch_geometric_temporal.dataset import WindmillOutputSmallDatasetLoader
loader6 = WindmillOutputSmallDatasetLoader()

from torch_geometric_temporal.dataset import MontevideoBusDatasetLoader
loader10 = MontevideoBusDatasetLoader()
try:
    from tqdm import tqdm
except ImportError:
    def tqdm(iterable):
        return iterable
Dataset RecurrentGCN Method Missing Rate Filters Lags Mean SD
fivenodes GConvGRU IT-STGCN 0.7 12 2 1.167 0.059
fivenodes GConvGRU STGCN 0.7 12 2 2.077 0.252
chickenpox GConvGRU IT-STGCN 0.8 16 4 1.586 0.199
chickenpox GConvGRU STGCN 0.8 16 4 2.529 0.292
pedalme GConvGRU IT-STGCN 0.6 12 4 1.571 0.277
pedalme GConvGRU STGCN 0.6 12 4 1.753 0.239
wikimath GConvGRU IT-STGCN 0.8 12 8 0.687 0.021
wikimath GConvGRU STGCN 0.8 12 8 0.932 0.04
windmillsmall GConvGRU IT-STGCN 0.7 12 8 1.180 0.035
windmillsmall GConvGRU STGCN 0.7 12 8 1.636 0.088
monte GConvGRU IT-STGCN 0.8 12 4 1.096 0.019
monte GConvGRU STGCN 0.8 12 4 1.516 0.040 
import torch
import torch.nn.functional as F
from torch_geometric_temporal.nn.recurrent import DCRNN

# from torch_geometric_temporal.dataset import ChickenpoxDatasetLoader
from torch_geometric_temporal.signal import temporal_signal_split

# loader1 = ChickenpoxDatasetLoader()

dataset = loader.get_dataset(lags=2)
dataset1 = loader1.get_dataset(lags=4)
dataset2 = loader2.get_dataset(lags=4)
dataset3 = loader3.get_dataset(lags=8)
dataset6 = loader6.get_dataset(lags=8)
dataset10 = loader10.get_dataset(lags=4)

train_dataset, test_dataset = temporal_signal_split(dataset, train_ratio=0.2)
train_dataset1, test_dataset1 = temporal_signal_split(dataset1, train_ratio=0.2)
train_dataset2, test_dataset2 = temporal_signal_split(dataset2, train_ratio=0.2)
train_dataset3, test_dataset3 = temporal_signal_split(dataset3, train_ratio=0.2)
train_dataset6, test_dataset6 = temporal_signal_split(dataset6, train_ratio=0.2)
train_dataset10, test_dataset10 = temporal_signal_split(dataset10, train_ratio=0.2)
class RecurrentGCN(torch.nn.Module):
    def __init__(self, node_features, filters):
        super(RecurrentGCN, self).__init__()
        self.recurrent = DCRNN(node_features, filters, 1)
        self.linear = torch.nn.Linear(filters, 1)

    def forward(self, x, edge_index, edge_weight):
        h = self.recurrent(x, edge_index, edge_weight)
        h = F.relu(h)
        h = self.linear(h)
        return h

fivenodes Nodes = 2, Filters = 2

model = RecurrentGCN(node_features=2, filters=2)

optimizer = torch.optim.Adam(model.parameters(), lr=0.01)

model.train()

for epoch in tqdm(range(50)):
    cost = 0
    _b=[]
    _d=[]
    for time, snapshot in enumerate(train_dataset):
        y_hat = model(snapshot.x, snapshot.edge_index, snapshot.edge_attr).reshape(-1)
        cost = cost + torch.mean((y_hat-snapshot.y)**2)
        _b.append(y_hat)
        _d.append(cost)
    cost = cost / (time+1)
    cost.backward()
    optimizer.step()
    optimizer.zero_grad()
100%|██████████| 50/50 [00:04<00:00, 12.09it/s]
model.eval()
cost = 0
_a = []
_a1=[]
for time, snapshot in enumerate(test_dataset):
    y_hat = model(snapshot.x, snapshot.edge_index, snapshot.edge_attr).reshape(-1)
    cost = cost + torch.mean((y_hat-snapshot.y)**2)
    _a.append(y_hat)
    _a1.append(cost)
cost = cost / (time+1)
cost = cost.item()
print("MSE: {:.4f}".format(cost))
MSE: 1.2529
_c = [_a1[i].detach() for i in range(len(_a1))]

_e = [_d[i].detach() for i in range(len(_d))]
fig, (( ax1,ax2),(ax3,ax4),(ax5,ax6)) = plt.subplots(3,2,figsize=(30,20))

ax1.set_title('train node1')
ax1.plot([train_dataset.targets[i][0] for i in range(train_dataset.snapshot_count)])
ax1.plot(torch.tensor([_b[i].detach()[0] for i in range(train_dataset.snapshot_count)]))

ax2.set_title('test node1')
ax2.plot([test_dataset.targets[i][0] for i in range(test_dataset.snapshot_count)])
ax2.plot(torch.tensor([_a[i].detach()[0] for i in range(test_dataset.snapshot_count)]))

ax3.set_title('train node2')
ax3.plot([train_dataset.targets[i][1] for i in range(train_dataset.snapshot_count)])
ax3.plot(torch.tensor([_b[i].detach()[1] for i in range(train_dataset.snapshot_count)]))


ax4.set_title('test node2')
ax4.plot([test_dataset.targets[i][1] for i in range(test_dataset.snapshot_count)])
ax4.plot(torch.tensor([_a[i].detach()[1] for i in range(test_dataset.snapshot_count)]))

ax5.set_title('train cost')
ax5.plot(_e)

ax6.set_title('test cost')
ax6.plot(_c)

Chickenpox Nodes = 4, Filters = 16

model1 = RecurrentGCN(node_features=4, filters=16)

optimizer1 = torch.optim.Adam(model1.parameters(), lr=0.01)

model1.train()

for epoch in tqdm(range(50)):
    cost = 0
    _b=[]
    _d=[]
    for time, snapshot in enumerate(train_dataset1):
        y_hat = model1(snapshot.x, snapshot.edge_index, snapshot.edge_attr).reshape(-1)
        cost = cost + torch.mean((y_hat-snapshot.y)**2)
        _b.append(y_hat)
        _d.append(cost)
    cost = cost / (time+1)
    cost.backward()
    optimizer1.step()
    optimizer1.zero_grad()
100%|██████████| 50/50 [00:13<00:00,  3.82it/s]
model1.eval()
cost = 0
_a = []
_a1=[]
for time, snapshot in enumerate(test_dataset1):
    y_hat = model1(snapshot.x, snapshot.edge_index, snapshot.edge_attr).reshape(-1)
    cost = cost + torch.mean((y_hat-snapshot.y)**2)
    _a.append(y_hat)
    _a1.append(cost)
cost = cost / (time+1)
cost = cost.item()
print("MSE: {:.4f}".format(cost))
MSE: 0.6929
_e = [_d[i].detach() for i in range(len(_d))]

_c = [_a1[i].detach() for i in range(len(_a1))]
fig, (( ax1,ax2),(ax3,ax4),(ax5,ax6)) = plt.subplots(3,2,figsize=(30,20))

ax1.set_title('train node1')
ax1.plot([train_dataset1.targets[i][0] for i in range(train_dataset1.snapshot_count)])
ax1.plot(torch.tensor([_b[i].detach()[0] for i in range(train_dataset1.snapshot_count)]))

ax2.set_title('test node1')
ax2.plot([test_dataset1.targets[i][0] for i in range(test_dataset1.snapshot_count)])
ax2.plot(torch.tensor([_a[i].detach()[0] for i in range(test_dataset1.snapshot_count)]))

ax3.set_title('train node2')
ax3.plot([train_dataset1.targets[i][1] for i in range(train_dataset1.snapshot_count)])
ax3.plot(torch.tensor([_b[i].detach()[1] for i in range(train_dataset1.snapshot_count)]))


ax4.set_title('test node2')
ax4.plot([test_dataset1.targets[i][1] for i in range(test_dataset1.snapshot_count)])
ax4.plot(torch.tensor([_a[i].detach()[1] for i in range(test_dataset1.snapshot_count)]))

ax5.set_title('train cost')
ax5.plot(_e)

ax6.set_title('test cost')
ax6.plot(_c)

Pedalme Nodes = 4, Filters = 8

model2 = RecurrentGCN(node_features=4, filters=8)

optimizer2 = torch.optim.Adam(model2.parameters(), lr=0.01)

model2.train()
    
for epoch in tqdm(range(50)):
    cost = 0
    _b=[]
    _d=[]
    for time, snapshot in enumerate(train_dataset2):
        y_hat = model2(snapshot.x, snapshot.edge_index, snapshot.edge_attr).reshape(-1)
        cost = cost + torch.mean((y_hat-snapshot.y)**2)
        _b.append(y_hat)
        _d.append(cost)
    cost = cost / (time+1)
    cost.backward()
    optimizer2.step()
    optimizer2.zero_grad()
100%|██████████| 50/50 [00:00<00:00, 56.65it/s]
model2.eval()
cost = 0
_a = []
_a1=[]
for time, snapshot in enumerate(test_dataset2):
    y_hat = model2(snapshot.x, snapshot.edge_index, snapshot.edge_attr).reshape(-1)
    cost = cost + torch.mean((y_hat-snapshot.y)**2)
    _a.append(y_hat)
    _a1.append(cost)
cost = cost / (time+1)
cost = cost.item()
print("MSE: {:.4f}".format(cost))
MSE: 0.4573
_e = [_d[i].detach() for i in range(len(_d))]
_c = [_a1[i].detach() for i in range(len(_a1))]
fig, (( ax1,ax2),(ax3,ax4),(ax5,ax6)) = plt.subplots(3,2,figsize=(30,20))

ax1.set_title('train node1')
ax1.plot([train_dataset2.targets[i][0] for i in range(train_dataset2.snapshot_count)])
ax1.plot(torch.tensor([_b[i].detach()[0] for i in range(train_dataset2.snapshot_count)]))

ax2.set_title('test node1')
ax2.plot([test_dataset2.targets[i][0] for i in range(test_dataset2.snapshot_count)])
ax2.plot(torch.tensor([_a[i].detach()[0] for i in range(test_dataset2.snapshot_count)]))

ax3.set_title('train node2')
ax3.plot([train_dataset2.targets[i][1] for i in range(train_dataset2.snapshot_count)])
ax3.plot(torch.tensor([_b[i].detach()[1] for i in range(train_dataset2.snapshot_count)]))

ax4.set_title('test node2')
ax4.plot([test_dataset2.targets[i][1] for i in range(test_dataset2.snapshot_count)])
ax4.plot(torch.tensor([_a[i].detach()[1] for i in range(test_dataset2.snapshot_count)]))

ax5.set_title('train cost')
ax5.plot(_e)

ax6.set_title('test cost')
ax6.plot(_c)

Wikimaths Nodes = 8, Filters = 12

model3 = RecurrentGCN(node_features=8, filters=12)

optimizer3 = torch.optim.Adam(model3.parameters(), lr=0.01)

model3.train()
    
for epoch in tqdm(range(50)):
    cost = 0
    _b=[]
    _d=[]
    for time, snapshot in enumerate(train_dataset3):
        y_hat = model3(snapshot.x, snapshot.edge_index, snapshot.edge_attr).reshape(-1)
        cost = cost + torch.mean((y_hat-snapshot.y)**2)
        _b.append(y_hat)
        _d.append(cost)
    cost = cost / (time+1)
    cost.backward()
    optimizer3.step()
    optimizer3.zero_grad()
100%|██████████| 50/50 [03:00<00:00,  3.62s/it]
model3.eval()
cost = 0
_a = []
_a1=[]
for time, snapshot in enumerate(test_dataset3):
    y_hat = model3(snapshot.x, snapshot.edge_index, snapshot.edge_attr).reshape(-1)
    cost = cost + torch.mean((y_hat-snapshot.y)**2)
    _a.append(y_hat)
    _a1.append(cost)
cost = cost / (time+1)
cost = cost.item()
print("MSE: {:.4f}".format(cost))
MSE: 0.5240
_e = [_d[i].detach() for i in range(len(_d))]
_c = [_a1[i].detach() for i in range(len(_a1))]
fig, (( ax1,ax2),(ax3,ax4),(ax5,ax6)) = plt.subplots(3,2,figsize=(30,20))

ax1.set_title('train node1')
ax1.plot([train_dataset3.targets[i][0] for i in range(train_dataset3.snapshot_count)])
ax1.plot(torch.tensor([_b[i].detach()[0] for i in range(train_dataset3.snapshot_count)]))

ax2.set_title('test node1')
ax2.plot([test_dataset3.targets[i][0] for i in range(test_dataset3.snapshot_count)])
ax2.plot(torch.tensor([_a[i].detach()[0] for i in range(test_dataset3.snapshot_count)]))

ax3.set_title('train node2')
ax3.plot([train_dataset3.targets[i][1] for i in range(train_dataset3.snapshot_count)])
ax3.plot(torch.tensor([_b[i].detach()[1] for i in range(train_dataset3.snapshot_count)]))


ax4.set_title('test node2')
ax4.plot([test_dataset3.targets[i][1] for i in range(test_dataset3.snapshot_count)])
ax4.plot(torch.tensor([_a[i].detach()[1] for i in range(test_dataset3.snapshot_count)]))

ax5.set_title('train cost')
ax5.plot(_e)

ax6.set_title('test cost')
ax6.plot(_c)

Windmillsmall Nodes = 8, Filters = 4

model6 = RecurrentGCN(node_features=8, filters=4)

optimizer6 = torch.optim.Adam(model6.parameters(), lr=0.01)

model6.train()

for epoch in tqdm(range(10)):
    cost = 0
    _b=[]
    _d=[]
    for time, snapshot in enumerate(train_dataset6):
        y_hat = model6(snapshot.x, snapshot.edge_index, snapshot.edge_attr).reshape(-1)
        cost = cost + torch.mean((y_hat-snapshot.y)**2)
        _b.append(y_hat)
        _d.append(cost)
    cost = cost / (time+1)
    cost.backward()
    optimizer6.step()
    optimizer6.zero_grad()
100%|██████████| 10/10 [01:24<00:00,  8.40s/it]
model6.eval()
cost = 0
_a = []
_a1=[]
for time, snapshot in enumerate(test_dataset6):
    y_hat = model6(snapshot.x, snapshot.edge_index, snapshot.edge_attr).reshape(-1)
    cost = cost + torch.mean((y_hat-snapshot.y)**2)
    _a.append(y_hat)
    _a1.append(cost)
cost = cost / (time+1)
cost = cost.item()
print("MSE: {:.4f}".format(cost))
MSE: 0.9971
_e = [_d[i].detach() for i in range(len(_d))]
_c = [_a1[i].detach() for i in range(len(_a1))]
fig, (( ax1,ax2),(ax3,ax4),(ax5,ax6)) = plt.subplots(3,2,figsize=(30,20))

ax1.set_title('train node1')
ax1.plot([train_dataset6.targets[i][0] for i in range(train_dataset6.snapshot_count)])
ax1.plot(torch.tensor([_b[i].detach()[0] for i in range(train_dataset6.snapshot_count)]))

ax2.set_title('test node1')
ax2.plot([test_dataset6.targets[i][0] for i in range(test_dataset6.snapshot_count)])
ax2.plot(torch.tensor([_a[i].detach()[0] for i in range(test_dataset6.snapshot_count)]))

ax3.set_title('train node2')
ax3.plot([train_dataset6.targets[i][1] for i in range(train_dataset6.snapshot_count)])
ax3.plot(torch.tensor([_b[i].detach()[1] for i in range(train_dataset6.snapshot_count)]))


ax4.set_title('test node2')
ax4.plot([test_dataset6.targets[i][1] for i in range(test_dataset6.snapshot_count)])
ax4.plot(torch.tensor([_a[i].detach()[1] for i in range(test_dataset6.snapshot_count)]))

ax5.set_title('train cost')
ax5.plot(_e)

ax6.set_title('test cost')
ax6.plot(_c)

Monte Nodes = 4, Filters = 12

model10 = RecurrentGCN(node_features=4, filters=12)

optimizer10 = torch.optim.Adam(model10.parameters(), lr=0.01)

model10.train()

for epoch in tqdm(range(50)):
    cost = 0
    _b=[]
    _d=[]
    for time, snapshot in enumerate(train_dataset10):
        y_hat = model10(snapshot.x, snapshot.edge_index, snapshot.edge_attr).reshape(-1)
        cost = cost + torch.mean((y_hat-snapshot.y)**2)
        _b.append(y_hat)
        _d.append(cost)
    cost = cost / (time+1)
    cost.backward()
    optimizer10.step()
    optimizer10.zero_grad()
model10.eval()
cost = 0
_a = []
_a1=[]
for time, snapshot in enumerate(test_dataset10):
    y_hat = model10(snapshot.x, snapshot.edge_index, snapshot.edge_attr).reshape(-1)
    cost = cost + torch.mean((y_hat-snapshot.y)**2)
    _a.append(y_hat)
    _a1.append(cost)
cost = cost / (time+1)
cost = cost.item()
print("MSE: {:.4f}".format(cost))
MSE: 0.9282
_e = [_d[i].detach() for i in range(len(_d))]

_c = [_a1[i].detach() for i in range(len(_a1))]
fig, (( ax1,ax2),(ax3,ax4),(ax5,ax6)) = plt.subplots(3,2,figsize=(30,20))

ax1.set_title('train node1')
ax1.plot([train_dataset10.targets[i][0] for i in range(train_dataset10.snapshot_count)])
ax1.plot(torch.tensor([_b[i].detach()[0] for i in range(train_dataset10.snapshot_count)]))

ax2.set_title('test node1')
ax2.plot([test_dataset10.targets[i][0] for i in range(test_dataset10.snapshot_count)])
ax2.plot(torch.tensor([_a[i].detach()[0] for i in range(test_dataset10.snapshot_count)]))

ax3.set_title('train node2')
ax3.plot([train_dataset10.targets[i][10] for i in range(train_dataset10.snapshot_count)])
ax3.plot(torch.tensor([_b[i].detach()[10] for i in range(train_dataset10.snapshot_count)]))


ax4.set_title('test node2')
ax4.plot([test_dataset10.targets[i][10] for i in range(test_dataset10.snapshot_count)])
ax4.plot(torch.tensor([_a[i].detach()[10] for i in range(test_dataset10.snapshot_count)]))

ax5.set_title('train cost')
ax5.plot(_e)

ax6.set_title('test cost')
ax6.plot(_c)