learn_rate=8e-3
updts_old=0.
epoch=0.
updts_new=0.
epochs=40
cost=np.zeros(epochs)
num_layers=1
modes=np.arange(1,4)
for k in range(epochs):
#np.random.shuffle(modes)
nu_max=np.asarray(np.arange(1980,3000,4))
np.random.shuffle(nu_max)
for j in range(nu_max.size):
start_time=timeit.default_timer()
if (updts_old==0.):
[c,updts_old]=train_spectra(nu_max[j],modes[0],nhidden,learn_rate,num_layers=num_layers,W=0,bhid=0,bvis=0)
#cost[i-min_modes,j]=c.eval()
else:
Wlist=[]
blist=[]
bvislist=[]
for item in range(len(updts_old)):
Wlist.append(updts_old[item][0][1])
blist.append(updts_old[item][1][1])
bvislist.append(updts_old[item][2][1])
updts_old=0.
[c,updts_old]=train_spectra(nu_max[j],modes[0],nhidden,learn_rate,num_layers=num_layers,W=Wlist,bhid=blist,bvis=bvislist)
#cost[i-min_modes,j]=c.eval()
end_time=timeit.default_timer()
print("Sample %d completed in %d s time"%(j,(end_time-start_time)))
cost[k]+=np.mean([item.eval() for item in c])
def train_func():
import train_deep_spectra
from train_deep_spectra import train_spectra
import numpy as np
import pandas as pd
import cPickle as cp
import timeit
#from theano import shared
#import os
#procno=int(os.environ["PBS_VNODENAME"])
num_modes=3
min_modes=12
#cost=np.zeros(num_modes*nu_max.size)
#cost=cost.reshape(num_modes,nu_max.size)
import theano
nhidden=1000
learn_rate=3e-3
updts_old=0.
epoch=0.
updts_new=0.
epochs=20
cost_perepoch=np.zeros(epochs)
num_layers=1
modes=np.arange(1,4)
freqs=np.arange(1956,3000,4)
sig_noise=np.load('/home/rakesh/sig_noise.npy')
sig=np.load('/home/rakesh/sig.npy')
sig_size=np.size(sig[0,:])
sig_t=theano.shared(np.asarray(sig,dtype=theano.config.floatX),borrow=True)
sig_noise_t=theano.shared(np.asarray(sig_noise,dtype=theano.config.floatX),borrow=True)
cost=[]
for k in range(epochs):
#np.random.shuffle(modes)
start_time=timeit.default_timer()
nu_max=np.asarray(np.arange(1956,3000,4))
np.random.shuffle(nu_max)
for j in range(nu_max.size):
#for i in range(num_modes):
p=np.ravel(np.where(freqs==nu_max[j]))
start_sample_time=timeit.default_timer()
if (updts_old==0.):
#st=timeit.default_timer()
[c,updts_old]=train_spectra(sig_t[p[0],:],sig_noise_t[p[0],:],sig_size,nu_max[j],modes[0],nhidden,learn_rate,num_layers=num_layers,W=0,bhid=0,bvis=0)
#en=timeit.default_timer()
#print("takes %3.4f s in one train"%(en-st))
#cost[i-min_modes,j]=c.eval()
else:
Wlist=[]
blist=[]
bvislist=[]
for item in range(len(updts_old)):
Wlist.append(updts_old[item][0][1])
blist.append(updts_old[item][1][1])
bvislist.append(updts_old[item][2][1])
updts_old=0.
#st=timeit.default_timer()
[c,updts_old]=train_spectra(sig_t[p[0],:],sig_noise_t[p[0],:],sig_size,nu_max[j],modes[0],nhidden,learn_rate,num_layers=num_layers,W=Wlist,bhid=blist,bvis=bvislist)
#en=timeit.default_timer()
#print("takes %3.4f s in one train"%(en-st))
#cost[k]+=np.mean([item.eval() for item in c])
#cost[k]+=c
cost.append(c)
end_sample_time=timeit.default_timer()
print("Sample %d takes %3.4f s time"%(j,(end_sample_time-start_sample_time)))
cost_perepoch[k]+=np.mean([item[0].eval() for item in cost])
end_time=timeit.default_timer()
print("Completed training in epoch %3.4fs"%(end_time-start_time))
df=pd.DataFrame(cost_perepoch/k)
df.to_csv('/home/rakesh/Code/Code_gpuopt/Results/cost.csv')
f=file('/home/rakesh/Code/Code_gpuopt/Results/nnparams.sav','wb')
cp.dump(updts_old,f,protocol=cp.HIGHEST_PROTOCOL)
f.close()
sig_t=theano.shared(np.asarray(sig,dtype=theano.config.floatX),borrow=True)
sig_noise_t=theano.shared(np.asarray(sig_noise,dtype=theano.config.floatX),borrow=True)
cost=[]
for k in range(epochs):
#np.random.shuffle(modes)
start_time=timeit.default_timer()
nu_max=np.asarray(np.arange(1956,3000,4))
np.random.shuffle(nu_max)
for j in range(nu_max.size):
#for i in range(num_modes):
p=np.ravel(np.where(freqs==nu_max[j]))
start_sample_time=timeit.default_timer()
if (updts_old==0.):
#st=timeit.default_timer()
[c,updts_old]=train_spectra(sig_t[p[0],:],sig_noise_t[p[0],:],sig_size,nu_max[j],modes[0],nhidden,learn_rate,num_layers=num_layers,W=0,bhid=0,bvis=0)
#en=timeit.default_timer()
#print("takes %3.4f s in one train"%(en-st))
#cost[i-min_modes,j]=c.eval()
else:
Wlist=[]
blist=[]
bvislist=[]
for item in range(len(updts_old)):
Wlist.append(updts_old[item][0][1])
blist.append(updts_old[item][1][1])
bvislist.append(updts_old[item][2][1])
updts_old=0.
#st=timeit.default_timer()
[c,updts_old]=train_spectra(sig_t[p[0],:],sig_noise_t[p[0],:],sig_size,nu_max[j],modes[0],nhidden,learn_rate,num_layers=num_layers,W=Wlist,bhid=blist,bvis=bvislist)
#en=timeit.default_timer()
def run_train(nnparams=0):
import train_deep_spectra
from train_deep_spectra import train_spectra
import numpy as np
import pandas as pd
import cPickle as cp
import timeit
from theano import tensor as T
#from theano import shared
#import os
#procno=int(os.environ["PBS_VNODENAME"])
num_modes=3
min_modes=12
#cost=np.zeros(num_modes*nu_max.size)
#cost=cost.reshape(num_modes,nu_max.size)
import theano
nhidden=1000
learn_rate=3e-3
updts_old=nnparams
epoch=0.
updts_new=0.
epochs=1
cost_perepoch=np.zeros(epochs)
num_layers=1
modes=np.arange(1,4)
freqs=np.arange(1956,3000,4)
sig_noise=np.load('/home/rakesh/sig_noise.npy')
sig=np.load('/home/rakesh/sig.npy')
sig_size=np.size(sig[0,:])
sig_t=theano.shared(np.asarray(sig,dtype=theano.config.floatX),borrow=True)
sig_noise_t=theano.shared(np.asarray(sig_noise,dtype=theano.config.floatX),borrow=True)
for k in range(epochs):
#np.random.shuffle(modes)
cost=[]
start_time=timeit.default_timer()
nu_max=np.asarray(np.arange(1956,3000,4))
np.random.shuffle(nu_max)
for j in range(nu_max.size):
#for i in range(num_modes):
p=np.ravel(np.where(freqs==nu_max[j]))
start_sample_time=timeit.default_timer()
if (updts_old==0.):
#st=timeit.default_timer()
[c,updts_old]=train_spectra(sig_t[p[0],:],sig_noise_t[p[0],:],sig_size,nu_max[j],modes[0],nhidden,learn_rate,num_layers=num_layers,W=0,bhid=0,bvis=0)
#en=timeit.default_timer()
#print("takes %3.4f s in one train"%(en-st))
#cost[i-min_modes,j]=c.eval()
else:
Wlist=[]
blist=[]
bvislist=[]
for item in range(len(updts_old)):
Wlist.append(updts_old[item][0][1])
blist.append(updts_old[item][1][1])
bvislist.append(updts_old[item][2][1])
updts_old=0.
#st=timeit.default_timer()
[c,updts_old]=train_spectra(sig_t[p[0],:],sig_noise_t[p[0],:],sig_size,nu_max[j],modes[0],nhidden,learn_rate,num_layers=num_layers,W=Wlist,bhid=blist,bvis=bvislist)
#en=timeit.default_timer()
#print("takes %3.4f s in one train"%(en-st))
#cost[k]+=np.mean([item.eval() for item in c])
#cost[k]+=c
cost.append(c[0])
end_sample_time=timeit.default_timer()
print("Sample %f takes %3.4f s time"%(j,(end_sample_time-start_sample_time)))
# cost_perepoch[k]+=np.mean([item.eval() for item in cost])
end_time=timeit.default_timer()
#print(sum(cost))
#cost_perepoch[k]=sum(cost)/(j+1)
print("Completed training in epoch %d in %3.4fs"%((k+1),(end_time-start_time)))
return(np.mean(cost),updts_old)