def train_svm(self, dataset, kernel='linear', C=1.0, gamma=0.0):
"""
Train classifiers for class
"""
print '%d trains class %s'%(mpi.comm_rank, self.cls)
t = time.time()
pos = dataset.get_pos_samples_for_class(self.cls)
neg = dataset.get_neg_samples_for_class(self.cls)
pos_gist = self.gist_table[pos, :]
neg_gist = self.gist_table[neg, :]
x = np.concatenate((pos_gist, neg_gist))
y = [1]*pos.shape[0] + [-1]*neg.shape[0]
print '%d compute svm for %s'%(mpi.comm_rank, self.cls)
svm_filename = config.get_gist_svm_filename(self.cls, dataset)
print svm_filename
self.svm = train_svm(x, y, kernel, C, gamma)
print '\ttook', time.time()-t,'sec'
print 'the score on train-data is %f'%self.svm.score(x,y)
table_t = self.svm.predict_proba(x)
y2 = np.array(y)
y2 = (y2+1)/2 # switch to 0/1
ap,_,_ = Evaluation.compute_cls_pr(table_t[:,1], y2)
print 'ap on train: %f'%ap
cPickle.dump(self.svm, open(svm_filename, 'w'))
return ap
def train_svm(self, dataset, kernel='linear', C=1.0, gamma=0.0):
"""
Train classifiers for class
"""
print '%d trains class %s' % (mpi.comm_rank, self.cls)
t = time.time()
pos = dataset.get_pos_samples_for_class(self.cls)
neg = dataset.get_neg_samples_for_class(self.cls)
pos_gist = self.gist_table[pos, :]
neg_gist = self.gist_table[neg, :]
x = np.concatenate((pos_gist, neg_gist))
y = [1] * pos.shape[0] + [-1] * neg.shape[0]
print '%d compute svm for %s' % (mpi.comm_rank, self.cls)
svm_filename = config.get_gist_svm_filename(self.cls, dataset)
print svm_filename
self.svm = train_svm(x, y, kernel, C, gamma)
print '\ttook', time.time() - t, 'sec'
print 'the score on train-data is %f' % self.svm.score(x, y)
table_t = self.svm.predict_proba(x)
y2 = np.array(y)
y2 = (y2 + 1) / 2 # switch to 0/1
ap, _, _ = Evaluation.compute_cls_pr(table_t[:, 1], y2)
print 'ap on train: %f' % ap
cPickle.dump(self.svm, open(svm_filename, 'w'))
return ap
def gist_evaluate_svms(d_train, d_val):
gist_scores = np.zeros((len(d_val.images), len(d_val.classes)))
gist_table = np.load(config.get_gist_dict_filename(d_train.name))
kernels = ['rbf', 'linear', 'poly']
Cs = [1, 10, 100]
gammas = [0, 0.3, 1]
setts = list(itertools.product(kernels, Cs, gammas))
val_gt = d_val.get_cls_ground_truth()
for cls_idx in range(len(d_val.classes)):
cls = d_val.classes[cls_idx]
gist = GistClassifier(cls, d_train, gist_table=gist_table, d_val=d_val)
filename = config.get_gist_crossval_filename(d_train, cls)
# doing some crossval right here!!!
for set_idx in range(mpi.comm_rank, len(setts), mpi.comm_size):
sett = setts[len(setts) - 1 - set_idx]
kernel = sett[0]
C = sett[1]
gamma = sett[2]
train_ap = gist.train_svm(d_train, kernel, C, gamma)
val_gist_table = np.load(config.get_gist_dict_filename(d_val.name))
gist_scores = gist.svm.predict_proba(val_gist_table)[:, 1]
val_ap, _, _ = Evaluation.compute_cls_pr(gist_scores,
val_gt.subset_arr(cls))
w = open(filename, 'a')
w.write('%s C=%d gamma=%f - train: %f, val: %f\n' %
(kernel, C, gamma, train_ap, val_ap))
w.close()
print 'ap on val: %f' % val_ap
print '%d at safebarrier' % mpi.comm_rank
safebarrier(comm)
gist_scores = comm.reduce(gist_scores)
if mpi.comm_rank == 0:
print gist_scores
filename = config.get_gist_classifications_filename(d_val)
cPickle.dump(gist_scores, open(filename, 'w'))
res = Evaluation.compute_cls_pr(gist_scores, val_gt.arr)
print res
def gist_evaluate_svms(d_train, d_val):
gist_scores = np.zeros((len(d_val.images), len(d_val.classes)))
gist_table = np.load(config.get_gist_dict_filename(d_train.name))
kernels = ['rbf', 'linear', 'poly']
Cs = [1,10,100]
gammas = [0,0.3,1]
setts = list(itertools.product(kernels, Cs, gammas))
val_gt = d_val.get_cls_ground_truth()
for cls_idx in range(len(d_val.classes)):
cls = d_val.classes[cls_idx]
gist = GistClassifier(cls, d_train, gist_table=gist_table, d_val=d_val)
filename = config.get_gist_crossval_filename(d_train, cls)
# doing some crossval right here!!!
for set_idx in range(mpi.comm_rank, len(setts), mpi.comm_size):
sett = setts[len(setts)-1-set_idx]
kernel = sett[0]
C = sett[1]
gamma = sett[2]
train_ap = gist.train_svm(d_train, kernel, C, gamma)
val_gist_table = np.load(config.get_gist_dict_filename(d_val.name))
gist_scores = gist.svm.predict_proba(val_gist_table)[:,1]
val_ap,_,_ = Evaluation.compute_cls_pr(gist_scores, val_gt.subset_arr(cls))
w = open(filename, 'a')
w.write('%s C=%d gamma=%f - train: %f, val: %f\n'%(kernel, C, gamma, train_ap, val_ap))
w.close()
print 'ap on val: %f'%val_ap
print '%d at safebarrier'%mpi.comm_rank
safebarrier(comm)
gist_scores = comm.reduce(gist_scores)
if mpi.comm_rank == 0:
print gist_scores
filename = config.get_gist_classifications_filename(d_val)
cPickle.dump(gist_scores, open(filename,'w'))
res = Evaluation.compute_cls_pr(gist_scores, val_gt.arr)
print res
def train(self, pos, neg, kernel, C):
y = [1] * pos.shape[0] + [-1] * neg.shape[0]
x = np.concatenate((pos, neg))
model = train_svm(x, y, kernel, C)
self.svm = model
print "model.score(C=%d): %f" % (C, model.score(x, y))
table_t = svm_proba(x, model)
y2 = np.array(y)
y2 = (y2 + 1) / 2 # switch to 0/1
ap, _, _ = Evaluation.compute_cls_pr(table_t[:, 1], y2)
print "ap on train set: %f" % ap
filename = config.get_classifier_filename(self, self.cls, self.train_dataset)
self.svm = model
self.save_svm(model, filename)
return model
def train(self, pos, neg, kernel, C):
y = [1] * pos.shape[0] + [-1] * neg.shape[0]
x = np.concatenate((pos, neg))
model = train_svm(x, y, kernel, C)
self.svm = model
print 'model.score(C=%d): %f' % (C, model.score(x, y))
table_t = svm_proba(x, model)
y2 = np.array(y)
y2 = (y2 + 1) / 2 # switch to 0/1
ap, _, _ = Evaluation.compute_cls_pr(table_t[:, 1], y2)
print 'ap on train set: %f' % ap
filename = config.get_classifier_filename(self, self.cls,
self.train_dataset)
self.svm = model
self.save_svm(model, filename)
return model
def eval_cls(self, ext_detector):
print 'evaluate svm for %s'%self.cls
dataset = ext_detector.dataset
assert(dataset.name in ['full_pascal_val','full_pascal_test'])
print dataset.name
table_cls = np.zeros(len(dataset.images))
for img_idx, image in enumerate(dataset.images):
print '%d eval on img %d/%d'%(mpi.comm_rank, img_idx, len(dataset.images))
img_dets, _ = ext_detector.detect(image, astable=True)
img_scores = img_dets.subset_arr('score')
score = self.classify_image(img_scores)
table_cls[img_idx] = score
ap, _,_ = Evaluation.compute_cls_pr(table_cls, dataset.get_cls_ground_truth().subset_arr(self.cls))
print 'ap on val for %s: %f'%(self.cls, ap)
return table_cls
def eval_cls(self, ext_detector):
print 'evaluate svm for %s' % self.cls
dataset = ext_detector.dataset
assert (dataset.name in ['full_pascal_val', 'full_pascal_test'])
print dataset.name
table_cls = np.zeros(len(dataset.images))
for img_idx, image in enumerate(dataset.images):
print '%d eval on img %d/%d' % (mpi.comm_rank, img_idx,
len(dataset.images))
img_dets, _ = ext_detector.detect(image, astable=True)
img_scores = img_dets.subset_arr('score')
score = self.classify_image(img_scores)
table_cls[img_idx] = score
ap, _, _ = Evaluation.compute_cls_pr(
table_cls,
dataset.get_cls_ground_truth().subset_arr(self.cls))
print 'ap on val for %s: %f' % (self.cls, ap)
return table_cls