def main():
D, cp, ca = loadData()
#num = 500
#clfP,clfA = fitting(D[:num,:],cp[:num],ca[:num])
clfP, clfA = fitting(D, cp, ca)
joblib.dump(clfP, 'svmmodels/test/SVMl_clfP.pkl')
joblib.dump(clfA, 'svmmodels/test/SVMl_clfA.pkl')
def makefit():
log('training linear model')
t = time.time()
D, cp, ca = loadData()
clfP, clfA, tt = fitting(D, cp, ca)
dirname = 'linearModel'
if not os.path.isdir('svmmodels/' + dirname):
os.mkdir('svmmodels/' + dirname)
joblib.dump(clfP, 'svmmodels/' + dirname + '/clfP.pkl')
joblib.dump(clfA, 'svmmodels/' + dirname + '/clfA.pkl')
log('model trained and saved in %.2f sec' % (time.time() - t))
def train_and_save(NUM=200):
tD = np.load('data_4096s/ALLDATAR.npy')
cp = (tD[:NUM, 2] == 1) | (tD[:NUM, 2] == 3)
ca = (tD[:NUM, 2] == 2) | (tD[:NUM, 2] == 3)
D = tD[:NUM, 3:]
kwargs = {'kernel': 'linear', 'C': 1.0, 'probability': True}
clfP = svm.SVC(**kwargs)
clfA = svm.SVC(**kwargs)
log('fitting plastic')
t = time.time()
clfP.fit(D, cp)
ttime = time.time() - t
log('in %f sec' % (time.time() - t))
log('fitting animals')
t = time.time()
clfA.fit(D, ca)
ttime += (time.time() - t)
log('in %f sec' % (time.time() - t))
joblib.dump(clfP, 'svmmodels/allprob/clfP.pkl')
joblib.dump(clfA, 'svmmodels/allprob/clfA.pkl')
# Train the svm on this labeled data and pickle it.
print "Pickling data and labels..."
Dfile = open(
"/home/ely/projects/faces/data/features/poselet_data/poselet_data_" +
str(num_poselets_trained) + ".pkl", 'w')
Lfile = open(
"/home/ely/projects/faces/data/features/poselet_labels/poselet_labels_"
+ str(num_poselets_trained) + ".pkl", 'w')
pickle.dump(DATA, Dfile)
pickle.dump(LABELS, Lfile)
Dfile.close()
Lfile.close()
print "Training and pickling the SVM..."
clf = train_svm(DATA, LABELS)
#classifier_output_file = open("/home/ely/projects/faces/data/classifiers/poselets/poselet_" + str(num_poselets_trained) + "-svm.pkl",'w')
classifier_output_file = "/home/ely/projects/faces/data/classifiers/poselets/poselet_" + str(
num_poselets_trained) + "-svm.pkl"
joblib.dump(clf, classifier_output_file)
#classifier_output_file.close()
poselet_end_time = time.time() - poselet_start_time
print "Finished training poselet classifier %d of %d..." % (
num_poselets_trained, total_number_of_poselets - 1)
print "Required %f seconds." % (poselet_end_time)
######################################################
print "Pickling data and labels..."
Dfile = open("/home/ely/projects/faces/data/features/poselet_data/poselet_data_"+str(num_poselets_trained)+".pkl",'w')
Lfile = open("/home/ely/projects/faces/data/features/poselet_labels/poselet_labels_"+str(num_poselets_trained)+".pkl",'w')
pickle.dump(DATA,Dfile)
pickle.dump(LABELS,Lfile)
Dfile.close()
Lfile.close()
print "Training and pickling the SVM..."
clf = train_svm(DATA,LABELS)
#classifier_output_file = open("/home/ely/projects/faces/data/classifiers/poselets/poselet_" + str(num_poselets_trained) + "-svm.pkl",'w')
classifier_output_file = "/home/ely/projects/faces/data/classifiers/poselets/poselet_" + str(num_poselets_trained) + "-svm.pkl"
joblib.dump(clf, classifier_output_file)
#classifier_output_file.close()
poselet_end_time = time.time() - poselet_start_time
print "Finished training poselet classifier %d of %d..."%(num_poselets_trained,total_number_of_poselets-1)
print "Required %f seconds."%(poselet_end_time)
######################################################
def batchtrain(self, njob=1, phase=None, shared_memory='no', verbose='on'):
"""Batch training which is called for rought training as well as
finetuning.
"""
t0 = time()
nnodes = getattr(self, 'nnodes')
dlen = getattr(self, 'dlen')
dim = getattr(self, 'dim')
mapsize = getattr(self, 'mapsize')
#############################################
# Seting the parameters
initmethod = getattr(self, 'initmethod')
mn = np.min(mapsize)
if mn == 1:
mpd = float(nnodes * 10) / float(dlen)
else:
mpd = float(nnodes) / float(dlen)
ms = max(mapsize[0], mapsize[1])
if mn == 1:
ms = ms / 5.
# Based on somtoolbox, Matlab
# case 'train', sTrain.trainlen = ceil(50 * mpd);
# case 'rough', sTrain.trainlen = ceil(10 * mpd);
# case 'finetune', sTrain.trainlen = ceil(40 * mpd);
if phase == 'rough':
# training length
trainlen = int(np.ceil(10 * mpd))
# radius for updating
if initmethod == 'random':
# trainlen = int(np.ceil(15 * mpd))
radiusin = max(1, np.ceil(ms / 2.))
radiusfin = max(1, radiusin / 8.)
elif initmethod == 'pca':
radiusin = max(1, np.ceil(ms / 8.))
radiusfin = max(1, radiusin / 4.)
elif phase == 'finetune':
# train lening length
trainlen = int(np.ceil(40 * mpd))
# radius for updating
if initmethod == 'random':
# trainlen = int(np.ceil(50 * mpd))
radiusin = max(1, ms / 8.) # from radius fin in rough training
radiusfin = max(1, radiusin / 16.)
elif initmethod == 'pca':
radiusin = max(1, np.ceil(ms / 8.) / 4)
radiusfin = 1 # max(1, ms / 128)
radius = np.linspace(radiusin, radiusfin, trainlen)
##################################################
UD2 = getattr(self, 'UD2')
New_Codebook_V = np.empty((nnodes, dim))
New_Codebook_V = getattr(self, 'codebook')
# print 'data is in shared memory?', shared_memory
if shared_memory == 'yes':
data = getattr(self, 'data')
Data_folder = tempfile.mkdtemp()
data_name = os.path.join(Data_folder, 'data')
dump(data, data_name)
data = load(data_name, mmap_mode='r')
else:
data = getattr(self, 'data')
# X2 is part of euclidean distance (x-y)^2 = x^2 +y^2 - 2xy that we use for
# each data row in bmu finding.
# Since it is a fixed value we can skip it during bmu finding for each data
# point, but later we need it calculate quantification error.
X2 = np.einsum('ij, ij->i', data, data)
if verbose == 'on':
print '%s training...' % phase
print 'radius_ini: {}, radius_final: {}, trainlen: {}'.format(
radiusin, radiusfin, trainlen)
for i in range(trainlen):
# In case of Guassian neighborhood
H = np.exp(-1.0 * UD2 / (2.0 * radius[i]**2)).reshape(nnodes, nnodes)
t1 = time()
bmu = None
bmu = self.para_bmu_find(data, New_Codebook_V, njb=njob)
if verbose == 'on':
print
# Updating the codebook
t2 = time()
New_Codebook_V = self.update_codebook_voronoi(data, bmu, H, radius)
# print 'updating nodes: ', round (time() - t2, 3)
if verbose == 'on':
print "epoch: {} ---> elapsed time: {}, quantization error: {}".format(
i + 1, round(time() - t1, 3), np.mean(np.sqrt(bmu[1] + X2)))
self.codebook = New_Codebook_V
bmu[1] = np.sqrt(bmu[1] + X2)
self.bmu = bmu
def save_model(self):
joblib.dump(self.model, os.sep.join([settings.RELATED_SEARCH_MODEL_BASE_DIR, self.model_file ]) )