def save_model_and_predict():
print 'Saving model...'
save_file = file('CNNmodel.pkl', 'wb')
genVariables = [epoch,pat_idx,tim_idx,xi,yi,zi+numPred,itr+1,patience,best_validation_loss]
cPickle.dump(genVariables,save_file,protocol = cPickle.HIGHEST_PROTOCOL)
for i in xrange(len(params)):
cPickle.dump(params[i].get_value(borrow=True), save_file, protocol = cPickle.HIGHEST_PROTOCOL)
save_file.close()
for l in logcost:
myfile.write("%f\n"%l)
logcost = []
print('Predicting for test patient')
tpat = ReadPat.new(test_pat_num,test_tstamp,num_channels)
shared_data.set_value(numpy.asarray(tpat.data,dtype = theano.config.floatX),borrow = True)
shared_truth.set_values(numpy.asarray(tpat.truth,dtype = 'int32'),borrow = True)
pstart_time = time.clock()
Prediction = numpy.zeros(img_shape)
for ix in xrange(len(xvalues)):
for iy in xrange(len(yvalues)):
for iz in xrange(len(zvalues)):
errors,p_y_given_x,pred = test_model(xvalues[ix],yvalues[iy],zvalues[iz])
pred = pred.reshape([numPred,numPred,numPred])
Prediction[xvalues[ix]+offset/2:xvalues[ix]+plen-offset/2 +1,
yvalues[iy]+offset/2:yvalues[iy]+plen-offset/2 +1,
zvalues[iz]+offset/2:zvalues[iz]+plen-offset/2 +1] = pred
p_y_given_x = p_y_given_x.transpose(1,0).reshape([num_classes,numPred,numPred,numPred])
PostProbs[:,xvalues[ix]+offset/2:xvalues[ix]+plen-offset/2 +1,
yvalues[iy]+offset/2:yvalues[iy]+plen-offset/2 +1,
zvalues[iz]+offset/2:zvalues[iz]+plen-offset/2 +1] = p_y_given_x
def load_shareddata(pat_idx): data = numpy.zeros(TStamps[pat_idx],num_channels, img_shape[0],img_shape[1],img_shape[2]],dtype = theano.config.floatX) truth = numpy.zeros(TStamps[pat_idx],img_shape[0], img_shape[1],img_shape[2]],dtype = 'int32') for index in xrange(TStamps[pat_idx]): pat = ReadPat.new(pat_idx+1,index+1,num_channels) data[index,:,:,:,:] = numpy.asarray(pat.data,dtype = theano.config.floatX) truth[index,:,:,:] = numpy.asarray(pat.truth,dtype = 'int32') shared_data.set_value(data) truth_truth.set_value(truth)
x = (20,160)
y = (17,192)
z = (11,151)
img_shape = [181,217,181]
num_patients = 4
TStamps = [4,4,5,4]
patch_size = [19,19]
valid_pat_num = 5
valid_Tstamps = 4
test_pat_num = 1
test_Tstamp = 1
pat = readPatMS.new(1,1)
num_patches = np.shape(image.extract_patches_2d(pat.data[0,20:160,17:192,0],patch_size))[0]
train_patches = np.zeros([num_patches, num_channels, patch_size[0], patch_size[1]])
trpatches_truth = np.zeros([num_patches])
shared_data = theano.shared(numpy.asarray(train_patches,dtype = theano.config.floatX),borrow = True)
shared_truth = theano.shared(numpy.asarray(trpatches_truth,dtype = 'int32'),borrow = True)
rng = numpy.random.RandomState(23455)
#Define Theano Tensors
nz = T.lscalar()
x = T.ftensor4('x')
y = T.ivector('y')
epoch,pat_idx,tim_idx,xi,yi,zi,itr,patience,best_validation_loss = [0,0,0,0,0,0,0,patience,numpy.inf] layer3convW = None layer3convb = None layer2convW = None layer2convb = None layer1convW = None layer1convb = None layer0convW = None layer0convb = None ###################################### ##Loading Dataset to Shared space data = numpy.zeros([numpy.max(TStamps),num_channels,img_shape[0],img_shape[1],img_shape[2]]) truth = numpy.zeros([numpy.max(TStamps),img_shape[0],img_shape[1],img_shape[2]]) for ix in xrange(TStamps[0]): pat = ReadPat.new(1,ix+1,num_channels) data[ix,:,:,:,:] = pat.data truth[ix,:,:,:] = pat.truth shared_data = theano.shared(numpy.asarray(data,dtype = theano.config.floatX),borrow = True) shared_truth = theano.shared(numpy.asarray(truth,dtype = 'int32'),borrow = True) shape = tpat.truth.shape Prediction = numpy.zeros(shape) shape = numpy.insert(shape,0,num_classes) PostProbs = numpy.zeros(shape) ###################################### ##Build architecture p_shape = (plen,plen,plen) #ignore_border = True idx = T.lscalar()
savedModel = file('CNNmodel.pkl', 'rb')
genVariables = cPickle.load(savedModel)
epoch, pat_idx, tim_idx, ix, iy, iz, itr, best_validation_loss, best_itr = genVariables
layer3convW = cPickle.load(savedModel)
layer3convb = cPickle.load(savedModel)
layer2convW = cPickle.load(savedModel)
layer2convb = cPickle.load(savedModel)
layer1convW = cPickle.load(savedModel)
layer1convb = cPickle.load(savedModel)
layer0convW = cPickle.load(savedModel)
layer0convb = cPickle.load(savedModel)
######################################
##Loading Dataset to Shared space
tpat = readPatMS.new(test_pat_num, test_tstamp)
test_data = theano.shared(numpy.asarray(tpat.data, dtype=theano.config.floatX),
borrow=True)
test_truth = theano.shared(numpy.asarray(tpat.truth, dtype='int32'),
borrow=True)
p_shape = (plen, plen, plen)
#ignore_border = True
idx = T.lscalar()
idy = T.lscalar()
idz = T.lscalar()
time_idx = T.lscalar()
x = T.ftensor4('x')
z = T.itensor3('y')
y = z.reshape([