def mytrain():
global bb0
global ww1
print "Training model:"
min_err = 0
min_err_epoch = 0
times_reduce = 0
for i in range(epoch):
start_time = time.time()
index = 1
for j in range(n_batch):
fi,x,y = get_fi_h1_y(train_file,index,batch_size)
index += batch_size
gx,ww1,ww2, ww3,bb1,bb2,bb3 = train(x,y)
b_size = len(fi)
for t in range(b_size):
ft = fi[t]
gxt = gx[t]
xt=x[t]
bb0=bb0-lr*gxt*xt*(1-xt)
for feat in ft:
ww0[feat]=ww0[feat]-lr *gxt*xt*(1-xt)
train_time = time.time() - start_time
mins = int(train_time / 60)
secs = int(train_time % 60)
print 'training: ' + str(mins) + 'm ' + str(secs) + 's'
start_time = time.time()
print_err(train_file,'\t\tTraining Err: \t' + str(i))# train error
train_time = time.time() - start_time
mins = int(train_time / 60)
secs = int(train_time % 60)
print 'training error: ' + str(mins) + 'm ' + str(secs) + 's'
start_time = time.time()
auc, rmse = auc_rmse(test_file)
test_time = time.time() - start_time
mins = int(test_time / 60)
secs = int(test_time % 60)
ut.log_p( 'Test Err:' + str(i) + '\t' + str(auc) + '\t' + str(rmse))
print 'test error: ' + str(mins) + 'm ' + str(secs) + 's'
#stop training when no improvement for a while
if auc>min_err:
min_err=auc
min_err_epoch=i
if times_reduce<3:
times_reduce+=1
else:
times_reduce-=1
if times_reduce<-2:
break
ut.log_p( 'Minimal test error is '+ str( min_err)+' , at EPOCH ' + str(min_err_epoch))
def print_err(file,msg=''):
auc,rmse=auc_rmse(file)
ut.log_p( msg + '\t' + str(auc) + '\t' + str(rmse))
if advertiser == '2997':
lr=0.05
if advertiser== '3386': #number of batches
x_dim=0
if sys.argv[2]=='mod' and advertiser=='2997':
lr=0.1
lambda1=0.00
ut.log_p('X:'+str(x_dim) + ' | Hidden 1:'+str(hidden1)+ ' | Hidden 2:'+str(hidden2)+
' | L rate:'+str(lr)+ ' | activation1:'+ str(acti_type)+
' | lambda:'+str(lambda1)
)
ww3=numpy.zeros(hidden2)
# ww3=rng.uniform(-0.05,0.05,hidden2)
bb3=0.
arr=[]
arr.append(x_dim)
arr.append(hidden1)
arr.append(hidden2)
# ww1,bb1,ww2,bb2=da.get_da_weights(train_file,arr,ncases=train_size,batch_size=100000)
# pickle.dump( (ww1,bb1,ww2,bb2), open( "2997_da_10.p", "wb" ))
lambda1=0.0000001
dropout=0.98
lambda1=0
elif advertiser=='3386':
train_size=ut.file_len(train_file) #training size
test_size=ut.file_len(test_file)
x_dim=556897
elif advertiser=='all':
x_dim=937670
dropout=0.95
lambda1=0
lr=0.001
hidden0=500
hidden1=400
ut.log_p('drop_mlp4da.py|ad:'+advertiser+'|drop:'+str(dropout)+'|b_size:'+str(batch_size)+' | X:'+str(x_dim) + ' | Hidden 0:'+str(hidden0)+ ' | Hidden 1:'+str(hidden1)+ ' | Hidden 2:'+str(hidden2)+
' | L_r:'+str(lr)+ ' | activation1:'+ str(acti_type)+
' | lambda:'+str(lambda1)
)
# initialise parameters
arr=[]
arr.append(x_dim)
arr.append(hidden0)
arr.append(hidden1)
arr.append(hidden2)
ww0,bb0=ut.init_weight(x_dim,hidden0,'sigmoid')
ww1,bb1=ut.init_weight(hidden0,hidden1,'sigmoid')
ww2,bb2=ut.init_weight(hidden1,hidden2,'sigmoid')
# ww0,bb0,ww1,bb1,ww2,bb2=da.get_da_weights(train_file,arr,ncases=train_size,batch_size=100000)
# pickle.dump( (ww0,bb0,ww1,bb1,ww2,bb2), open( "2997_da_4l_10.p", "wb" ))
name = s[1 + k][0:s[1 + k].index(':')]
field = name_field[name]
feat_field[feat] = field
def feat_layer_one_index(feat, l):
return 1 + feat_field[feat] * k + l
def feats_to_layer_one_array(feats):
x = numpy.zeros(xdim)
x[0] = w_0
for feat in feats:
x[feat_layer_one_index(feat, 0):feat_layer_one_index(feat, k)] = feat_weights[feat]
return x
ut.log_p('drop_mlp3fm.py|ad:'+advertiser+'|drop:'+str(dropout)+'|b_size:'+str(batch_size)+' | X:'+str(xdim) + ' | Hidden 1:'+str(hidden1)+ ' | Hidden 2:'+str(hidden2)+
' | L_r:'+str(lr)+ ' | activation1:'+ str(acti_type)+
' | lambda:'+str(lambda1)
)
# initialise parameters
w=rng.uniform( low=-numpy.sqrt(6. / (xdim + hidden1)),
high=numpy.sqrt(6. / (xdim + hidden1)),
size=(xdim,hidden1))
if acti_type=='sigmoid':
ww1=numpy.asarray((w))
elif acti_type=='tanh':
ww1=numpy.asarray((w*4))
else:
ww1=numpy.asarray(rng.uniform(-1,1,size=(xdim,hidden1)))
bb1=numpy.zeros(hidden1)
def print_err(file,msg=''):
auc,rmse=get_err_bat(file)
ut.log_p( msg + '\t' + str(auc) + '\t' + str(rmse))
if len(sys.argv) > 1:
advertiser = sys.argv[1]
train_file = "../../make-ipinyou-data/" + advertiser + "/train.dl.txt" # training file
test_file = "../../make-ipinyou-data/" + advertiser + "/test.dl.txt" # test file
feats = ut.feats_len(train_file) # feature size
train_size = 312437 # ut.file_len(train_file) #training size
test_size = 156063 # ut.file_len(test_file) #test size
n_batch = train_size / batch_size # number of batches
ut.log_p(
"Hidden one:"
+ str(hidden1)
+ "|L rate:"
+ str(lr)
+ "|activation1:"
+ str(acti_type)
+ "|feats:"
+ str(feats)
+ "|lambda1:"
+ str(lambda1)
)
# initialise parameters
w = rng.uniform(
low=-numpy.sqrt(6.0 / (feats + hidden1)), high=numpy.sqrt(6.0 / (feats + hidden1)), size=(feats, hidden1)
)
if acti_type == "sigmoid":
ww1 = numpy.asarray((w))
elif acti_type == "tanh":
ww1 = numpy.asarray((w * 4))
else:
def print_err(file, msg=""):
auc, rmse = get_err_bat(file)
ut.log_p(msg + "\t" + str(auc) + "\t" + str(rmse))
def log_p(msg, m=""):
ut.log_p(msg, "drop_mlp4da" + str(advertiser))
# feats = ut.feats_len(train_file) #feature size
train_size = 312437 # ut.file_len(train_file) #training size
test_size = 156063 # ut.file_len(test_file) #test size
n_batch = train_size / batch_size # number of batches
o_fm = fm.DataFM(fm_model_file)
ut.log_p(
"X:"
+ str(o_fm.xdim)
+ " | Hidden 1:"
+ str(hidden1)
+ " | Hidden 2:"
+ str(hidden2)
+ " | L rate:"
+ str(lr)
+ " | activation1:"
+ str(acti_type)
+ " | lambda:"
+ str(lambda1)
)
# initialise parameters
w = rng.uniform(
low=-numpy.sqrt(6.0 / (o_fm.xdim + hidden1)),
high=numpy.sqrt(6.0 / (o_fm.xdim + hidden1)),
size=(o_fm.xdim, hidden1),
)
if acti_type == "sigmoid":
ww1 = numpy.asarray((w))
def log_p(msg,m=""):
ut.log_p(msg,"drop_mlp4da"+str(advertiser))
lambda1=0.0000001
elif advertiser=='3386':
x_dim=556897
elif advertiser=='2261':
lambda1=0
x_dim=333143
if sys.argv[2]=='mod' and advertiser=='2997':
train_file=train_file+'.mod4.txt'
train_size=ut.file_len(train_file)
print train_file
n_batch=train_size/batch_size
lr=0.002
lambda1=0.00
ut.log_p('ucl_mlp4da.py|'+'bat size:'+str(batch_size)+' | X:'+str(x_dim) + ' | Hidden 0:'+str(hidden0)+ ' | Hidden 1:'+str(hidden1)+ ' | Hidden 2:'+str(hidden2)+
' | L rate:'+str(lr)+ ' | activation1:'+ str(acti_type)+
' | lambda:'+str(lambda1)
)
# initialise parameters
arr=[]
arr.append(x_dim)
arr.append(hidden0)
arr.append(hidden1)
arr.append(hidden2)
ww0,bb0=ut.init_weight(x_dim,hidden0,'sigmoid')
ww1,bb1=ut.init_weight(hidden0,hidden1,'sigmoid')
ww2,bb2=ut.init_weight(hidden1,hidden2,'sigmoid')
# ww0,bb0,ww1,bb1,ww2,bb2=da.get_da_weights(train_file,arr,ncases=train_size,batch_size=100000)
# pickle.dump( (ww0,bb0,ww1,bb1,ww2,bb2), open( "2997_da_4l_10.p", "wb" ))
train_file='../../make-ipinyou-data/' + advertiser + '/train.fm.txt' #training file
test_file='../../make-ipinyou-data/' + advertiser + '/test.fm.txt' #test file
fm_model_file='../../make-ipinyou-data/' + advertiser + '/fm.model.txt' #fm model file
#feats = ut.feats_len(train_file) #feature size
train_size=312437 #ut.file_len(train_file) #training size
test_size=156063 #ut.file_len(test_file) #test size
n_batch=train_size/batch_size #number of batches
o_fm=fm.DataFM(fm_model_file)
ut.log_p('F:' + str(len(o_fm.feat_weights)) + ' | X:'+str(o_fm.xdim) + ' | Hidden1:'+str(hidden1)+ ' | Hidden2:'+str(hidden2)+
' | L rate:'+str(lr)+ ' | activation:'+ str(acti_type)+
' | lambda:'+str(lambda1)
)
# initialise parameters
w=rng.uniform( low=-numpy.sqrt(6. / (o_fm.xdim + hidden1)),
high=numpy.sqrt(6. / (o_fm.xdim + hidden1)),
size=(o_fm.xdim,hidden1))
if acti_type=='sigmoid':
ww1=numpy.asarray((w*4))
elif acti_type=='tanh':
ww1=numpy.asarray((w*4))
else:
ww1=numpy.asarray(rng.uniform(-1,1,size=(o_fm.xdim,hidden1)))
bb1=numpy.zeros(hidden1)
