enzymes_contfeats_gnnml3_tf.py

from __future__ import division
from __future__ import print_function

import time
import tensorflow as tf

from utils_tf import *
from models_tf import DSSGCN_GC_BATCH
from tensorflow import set_random_seed
import matplotlib.pyplot as plt
import scipy.io as sio
from sklearn.model_selection import StratifiedKFold
import numpy as np
import pandas as pd




# Settings
flags = tf.app.flags
FLAGS = flags.FLAGS

flags.DEFINE_float('learning_rate', 0.001, 'Initial learning rate.')
flags.DEFINE_integer('epochs', 400, 'Number of epochs to train.')
flags.DEFINE_list('hidden', [200,200,'meanmax',-100,-6], 'Number of units in each layer negative:denselayer, positive:ConvGraph layer ')
flags.DEFINE_list('activation_funcs', [tf.nn.relu,tf.nn.relu, None, tf.nn.relu,lambda x: x], 'Activation functions for layers  [tf.nn.relu, lambda x: x]')
flags.DEFINE_list('biases', [False,False,None,True,True], 'if apply bias on layers')
flags.DEFINE_list('isdroput_inp', [True,True,None,True,True], 'if apply dropout on layers'' input')
flags.DEFINE_list('isdroput_kernel', [True,True,None,False,False], 'if apply dropout on layers'' kernel')
flags.DEFINE_float('dropout', 0.1, 'Dropout rate (1 - keep probability).')
flags.DEFINE_float('weight_decay', 1e-4, 'Weight for L2 loss on embedding matrix.') 
flags.DEFINE_integer('numbatch', 3, 'number of update in each epoch')

# nfreq+1 supports of GNNML3 to be prepared
nfreq=3
# receptive field and bandwidth parameter
recfield=5
dv=1


nkernel=nfreq+1
bsize=flags.FLAGS.numbatch
a=sio.loadmat('dataset/enzymes/raw/enzymes.mat')
# list of adjacency matrix
A=a['A'][0]
# list of features
F=a['F'][0]
Y=a['Y'][0]



U=[];V=[]
dmax=0
dmin=1000

for i in range(0,len(A)):
    W=1.0*A[i]  
    d = W.sum(axis=0)
    dmax=max(dmax,d.max())
    dmin=min(dmin,d.min())
    # Laplacian matrix.     
    dis=1/np.sqrt(d)
    dis[np.isinf(dis)]=0
    dis[np.isnan(dis)]=0
    D=np.diag(dis)
    nL=np.eye(D.shape[0])-(W.dot(D)).T.dot(D)
    V1,U1 = np.linalg.eigh(nL) 
    V1[V1<0]=0
    U.append(U1)
    V.append(V1)


vmax=0
nmax=0
for v in V:    
    vmax=max(vmax,v.max())
    nmax=max(nmax,v.shape[0])

globalmax=vmax

A0=[];A1=[];A2=[]
ND=np.zeros((len(A),1)) 
FF=np.zeros((len(A),nmax,21+1)) 
YY=np.zeros((len(A),6))
SP=np.zeros((len(A),nkernel,nmax,nmax))



# prepare convolution supports

for i in range(0,len(A)):  
    n=F[i].shape[0]    
    FF[i,0:n,0:21]= F[i]#[:,0:3]  
    # add node degree as feature
    FF[i,0:n,-1]= A[i].sum(0)     
    ND[i,0]=n
    YY[i,Y[i]]=1  
    vmax= V[i].max() 
    if recfield==0:
        M=A[i]
    else:
        M=(A[i]+np.eye(n))
        for j in range(1,recfield):
            M=M.dot(M) 
    M=(M>0)
    
    SP[i,0,0:n,0:n]=np.eye(n)  
    freqcenter=np.linspace(V[i].min(),V[i].max(),nkernel-1)
    SP[i,0,0:n,0:n]=np.eye(n)       
    for ii in range(0,len(freqcenter)): 
        SP[i,ii+1,0:n,0:n]=M* (U[i].dot(np.diag(np.exp(-(dv*(V[i]-freqcenter[ii])**2))).dot(U[i].T))) 
              
    # SP[i,1,0:n,0:n]= M*(U[i].dot(np.diag(np.exp(-(1*(V[i]-0.0)**2))).dot(U[i].T)))
    # SP[i,2,0:n,0:n]= M*(U[i].dot(np.diag(np.exp(-(1*(V[i]-vmax*0.5)**2))).dot(U[i].T)))
    # SP[i,3,0:n,0:n]= M*(U[i].dot(np.diag(np.exp(-(1*(V[i]-vmax)**2))).dot(U[i].T)))   
     
num_supports=SP.shape[1]

def normalize_wrt_train(FF,ND,trid):
    tmp=np.zeros((0,FF[0].shape[1]))
    for i in trid:
        tmp=np.vstack((tmp,FF[i][0:int(ND[i]),:]))
    avg=tmp.mean(0)
    st=tmp.std(0)
    FFF=FF.copy()
    for i in range(0,len(FFF)):
        tmp2=(FFF[i][0:int(ND[i]),:]-avg)/st        
        FFF[i][0:int(ND[i]),:]=tmp2
    return FFF



for iter in range(0,20):
    seed = iter 
    np.random.seed(seed)
    tf.set_random_seed(seed)    

    tprediction=[]
    TS=[]
    NB=np.zeros((FLAGS.epochs,10))

    for fold in range(0,10):

        
        tsid=np.loadtxt('dataset/enzymes/raw/10fold_idx/test_idx-'+str(fold+1)+'.txt')
        trid=np.loadtxt('dataset/enzymes/raw/10fold_idx/train_idx-'+str(fold+1)+'.txt')
        trid=trid.astype(np.int)
        tsid=tsid.astype(np.int)
        
        FFF=normalize_wrt_train(FF,ND,trid)       
        

        placeholders = {        
                'support': tf.placeholder(tf.float32, shape=(None,num_supports,nmax,nmax)),
                'features': tf.placeholder(tf.float32, shape=(None,nmax, FFF.shape[2])),
                'labels': tf.placeholder(tf.float32, shape=(None, 6)),  
                'nnodes': tf.placeholder(tf.float32, shape=(None, 1)),               
                'dropout': tf.placeholder_with_default(0., shape=())       
            }

        model = DSSGCN_GC_BATCH(placeholders, input_dim=FFF.shape[2], logging=True)  

        sess = tf.Session()
        sess.run(tf.global_variables_initializer())
        

        feed_dictT = dict()
        feed_dictT.update({placeholders['labels']: YY[tsid,:]})    
        feed_dictT.update({placeholders['features']: FFF[tsid,:,:]})
        feed_dictT.update({placeholders['support']: SP[tsid,:,:,:]})    
        feed_dictT.update({placeholders['nnodes']: ND[tsid,]})     
        feed_dictT.update({placeholders['dropout']: 0})

        
        ytest=YY[tsid,:] 
        ind=np.round(np.linspace(0,len(trid),bsize+1))

        
        besttr=100
        for epoch in range(FLAGS.epochs):
            #otrid=trid.copy()
            np.random.shuffle(trid)
            ent=[]
            for i in range(0,bsize):
                feed_dictB = dict()
                bid=trid[int(ind[i]):int(ind[i+1])]
                feed_dictB.update({placeholders['labels']: YY[bid,:]})    
                feed_dictB.update({placeholders['features']: FFF[bid,:,:]})
                feed_dictB.update({placeholders['support']: SP[bid,:,:,:]})             
                feed_dictB.update({placeholders['nnodes']: ND[bid,]})            
                feed_dictB.update({placeholders['dropout']: FLAGS.dropout})
                outs = sess.run([model.opt_op,model.entropy], feed_dict=feed_dictB)
                ent.append(outs[1])
            
            outsT = sess.run([model.accuracy, model.loss, model.entropy,model.outputs], feed_dict=feed_dictT)
            
            vtest=np.sum(np.argmax(outsT[3],1)==np.argmax(ytest,1))            

            NB[epoch,fold]=vtest
            if np.mod(epoch + 1,1)==0 or epoch==0:
                print(fold," Epoch:", '%04d' % (epoch + 1),"train_xent=", "{:.5f}".format(np.mean(ent)), "test_loss=", "{:.5f}".format(outsT[1]), 
                "test_xent=", "{:.5f}".format(outsT[2]), "test_acc=", "{:.5f}".format(outsT[0]), " ntrue=", "{:.0f}".format(vtest))
        print(NB.sum(1).max()/(fold+1)/60)
    
    fname='logs/GNNML3_enzyms_fullfeat_'+ str(nkernel)+'_'+str(iter)+'.csv'
    pd.DataFrame(NB).to_csv(fname) 
print(NB.sum(1).max())