# load labels
labels = imt.readlabels(trfile)
m = len(labels)
# split to train and cv sets
trsplit = int(np.round(m*splitsize))
iterm=range(m)
nparams = len(paramc)*len(paramg)
scores = np.zeros(nparams)
# load in training set
# note: this takes up a bit of memory
print('Loading training data...')
trsetfull = imt.loadimgfromcsv(trfile,colstart=1)/255.0
print('...done')
trset = trsetfull[:trsplit]
cvset = trsetfull[trsplit:]
trlabels = labels[:trsplit]
cvlabels = labels[trsplit:]
c=5.0
g=0.05
#counter=0
#paramholder=np.zeros([nparams,2])
#bestscore=-1
#for c in paramc:
# for g in paramg:
# model=SVC(C=c,gamma=g)
# print('Fitting model with params c: %f, g: %f ...'%(c,g))
# model=model.fit(trset,trlabels)
layermodels = []
starttime = time.clock()
for i in range(len(hlayers)):
print('Training hidden layer %d' % i)
# create denoising autoencoder, consider more options
da = nnt.dAE(alayers[i],
alayers[i + 1],
noise=noisel[i],
errtype=etype)
for j in range(epochs):
print('Starting epoch %d...' % j)
for z in range(nblocks):
# load up images
if i == 0:
x = imt.loadimgfromcsv(trfile,
range(blocksplits[z],
blocksplits[z + 1]),
colstart=1) / 255.0
else:
x = imt.loadimgfromcsv(acttmpfile,
range(blocksplits[z],
blocksplits[z + 1]),
colstart=0,
headlines=0)
nx = x.shape[0]
for zz in range(nx):
da.GD(x[zz].reshape(x.shape[1], 1), alpha=lr)
epochend = time.clock()
etime = (epochend - starttime) / 60
print('...finished in %2f minutes' % etime)
if i < len(hlayers) - 1:
# save out activations
# load labels
labels = imt.readlabels(trfile)
m = len(labels)
# split to train and cv sets
trsplit = int(np.round(m * splitsize))
iterm = range(m)
nparams = len(paramc) * len(paramg)
scores = np.zeros(nparams)
# load in training set
# note: this takes up a bit of memory
print('Loading training data...')
trsetfull = imt.loadimgfromcsv(trfile, colstart=1) / 255.0
print('...done')
trset = trsetfull[:trsplit]
cvset = trsetfull[trsplit:]
trlabels = labels[:trsplit]
cvlabels = labels[trsplit:]
c = 5.0
g = 0.05
#counter=0
#paramholder=np.zeros([nparams,2])
#bestscore=-1
#for c in paramc:
# for g in paramg:
# model=SVC(C=c,gamma=g)
# print('Fitting model with params c: %f, g: %f ...'%(c,g))
# model=model.fit(trset,trlabels)
try:
layermodels=pickle.load(open('data/pretune.p','r'))
except:
layermodels=[]
starttime=time.clock()
for i in range(len(hlayers)):
print('Training hidden layer %d'%i)
# create denoising autoencoder, consider more options
da=nnt.dAE(alayers[i],alayers[i+1],noise=noisel[i],errtype=etype)
for j in range(epochs):
print('Starting epoch %d...'%j)
for z in range(nblocks):
# load up images
if i==0:
x=imt.loadimgfromcsv(trfile,range(blocksplits[z],blocksplits[z+1]),colstart=1)/255.0
else:
x=imt.loadimgfromcsv(acttmpfile,range(blocksplits[z],blocksplits[z+1]),colstart=0,headlines=0)
nx=x.shape[0]
for zz in range(nx):
da.GD(x[zz].reshape(x.shape[1],1),alpha=lr)
epochend=time.clock()
etime=(epochend-starttime)/60
print('...finished in %2f minutes' %etime)
if i<len(hlayers)-1:
# save out activations
print('Saving out activations...')
writeto=acttmpfileroot+'_%d.csv'%i
f=open(writeto,'wb')
csvf=csv.writer(f)
for c in range(nblocks):
