def find_best_self(domain):
X_train = load_obj("%s/X_train"%domain)
y_train = load_obj("%s/y_train"%domain)
X_test = load_obj("%s/X_test"%domain)
y_test = load_obj("%s/y_test"%domain)
X_un = load_obj("%s/X_un"%domain)
thetas = [0.5,0.6,0.7,0.8,0.9]
best_acc = 0.0
best_clf =""
best_theta = 0.0
resFile = open("../work/params/%s_in_theta.csv"%domain,"w")
resFile.write("theta, acc\n")
for theta in thetas:
print "##############################"
print "start with theta=%s"%theta
print "##############################"
acc,clf_func = self_train(domain,X_train,y_train,X_test,y_test,X_un,theta=theta)
if best_acc<acc:
best_acc = acc
best_clf = clf_func
best_theta = theta
resFile.write("%f, %f\n"%(theta,acc))
resFile.flush()
resFile.close()
print "##############################"
print "best_theta:",best_theta,"best_acc:",best_acc
save_obj(best_clf,"%s/self_clf"%domain)
pass
def majority_vote(target):
X_test = load_obj("%s/X_test"%target)
y_test = load_obj("%s/y_test"%target)
domains = []
if "mlp" in target:
domains = ["mlp/books","mlp/dvd","mlp/electronics","mlp/kitchen"]
else:
if "large" not in target:
domains = ["books","dvd","electronics","kitchen"]
if target not in domains:
return
else:
domains =["large/baby","large/cell_phone","large/imdb","large/yelp2014"]
models = []
for source in domains:
if target == source:
continue
else:
print source
clf_func = load_obj("%s/self_clf"%source)
models.append(clf_func)
eclf = EnsembleVoteClassifier(clfs=models,refit=False)#weights=[1,1,1],
eclf.fit(X_test,y_test) # this line is not doing work
tmp_name = target.upper()[0] if "large" not in target else "large/"+target.upper()[6]
tmp_name = target.upper()[0] if "mlp" not in target else "mlp/"+target.upper()[4]
save_obj(eclf, '%s_eclf'%(tmp_name))
pred = eclf.predict(X_test)
acc = accuracy_score(y_test,pred) if "large" not in target else f1_score(y_test,pred,average='macro')
print 'self-train',acc
pass
def unlabel_sim(target):
tgt_un = load_obj("%s/X_un"%target)
# print target,tgt_un.shape
c_t = compute_centriod(tgt_un)
computed_tgt_sim = [cos_sim(x,c_t) for x in tgt_un]
save_obj(computed_tgt_sim,"%s/tgt_sim"%target)
pass
def generate_un():
source = "d1"
domains = ["books","dvd","electronics","kitchen"]
for target in domains:
src_un,tgt_un = set_up_unlabelled(source,target,False)
save_obj(tgt_un,"%s/X_un"%target)
pass
def majority_vote_mlp(target):
X_test = load_obj("%s/X_test"%target)
y_test = load_obj("%s/y_test"%target)
# domains = ["mlp/books","mlp/dvd","mlp/electronics","mlp/kitchen"]
data_name = ["books", "dvd", "electronics", "kitchen"]
X_joint = load_obj("%s/X_joint"%target)
y_joint = load_obj("%s/y_joint"%target)
temp_un = load_obj("%s/X_un"%target)
meta_sources = []
for i in range(len(data_name)):
if 'mlp/'+data_name[i] != target:
meta_sources.append(data_name[i])
# print meta_sources
models = []
for j in range(len(meta_sources)):
temp_X = X_joint[j]
temp_y = y_joint[j]
thetas = [0.5,0.6,0.7,0.8,0.9]
best_acc = 0.0
best_clf =""
best_theta = 0.0
resFile = open("../work/params/%s_theta_self-%s.csv"%(target,meta_sources[j].upper()[0]),"w")
resFile.write("theta, acc\n")
for theta in thetas:
print "##############################"
print "start with theta=%s"%theta
print "##############################"
acc,clf_func = self_train(target,temp_X,temp_y,X_test,y_test,temp_un,theta=theta)
if best_acc<acc:
best_acc = acc
best_clf = clf_func
best_theta = theta
resFile.write("%f, %f\n"%(theta,acc))
resFile.flush()
resFile.close()
print "##############################"
print "best_theta:",best_theta,"best_acc:",best_acc
models.append(best_clf)
eclf = EnsembleVoteClassifier(clfs=models,refit=False)#weights=[1,1,1],
eclf.fit(X_test,y_test) # this line is not doing work
# tmp_name = target.upper()[0] if "large" not in target else "large/"+target.upper()[6]
# tmp_name = 'mlp/'+target.upper()[4]
save_obj(eclf, "%s/self_clf"%target)
pred = eclf.predict(X_test)
# print pred
acc = accuracy_score(y_test,pred)
print 'self-train',acc
pass
def src_cost(target):
X_joint = load_obj("%s/X_joint"%target)
src_train = get_sents(X_joint)
tgt_un = load_obj("%s/X_un"%target)
c_t = compute_centriod(tgt_un)
# print src_train
sim = [cos_sim(x,c_t) for x in src_train]
# s = sum(sim)
# sim = [x/s for x in sim]
# print normalized_sim
sim = list(split_list(sim,3))
save_obj(sim,"%s/src_cost"%target)
pass
def find_theta(target):
print target
resFile = open("../work/params/%s_theta.csv"%target,"w")
resFile.write("theta, acc, method\n")
thetas = [0.5,0.6,0.7,0.8,0.9]
best_theta = 0.0
best_acc = 0.0
best_pos = ""
best_neg = ""
best_method = ""
for theta in thetas:
pos_star,neg_star,acc,method,sorting,pos_proba,neg_proba = predict_tops(target,theta=theta)
# print "S_L:",
# acc1 = test_confidence(target,option=1,theta=theta)
print "PL(T_L*):",acc,theta
# acc = test_confidence(target, pos_star,neg_star,theta=theta)
# print "S_L+T_L*:",
# acc3 = test_confidence(target,option=2,theta=theta)
resFile.write("%f, %f, %s, %s\n"%(theta,acc,method,sorting))
if best_acc<acc:
best_acc = acc
best_theta = theta
best_pos = pos_star
best_neg = neg_star
best_method = method
best_sorting = sorting
best_pos_proba = pos_proba
best_neg_proba = neg_proba
resFile.flush()
resFile.close()
print "####################################"
print "best_theta:",best_theta,"best_acc:",best_acc, "best_method:",best_method,best_sorting
save_obj(best_pos,"%s/pos_star"%target)
save_obj(best_neg,"%s/neg_star"%target)
save_obj(best_pos_proba,"%s/pos_proba"%target)
save_obj(best_neg_proba,"%s/neg_proba"%target)
pass
def joint_train(target):
X_test = load_obj("%s/X_test"%target)
y_test = load_obj("%s/y_test"%target)
domains = []
temp_X = []
temp_y = []
temp_un = []
if "mlp" in target:
domains = ["mlp/books","mlp/dvd","mlp/electronics","mlp/kitchen"]
temp_X = get_sents(load_obj("%s/X_joint"%target))
temp_y = get_sents(load_obj("%s/y_joint"%target))
for source in domains:
if target == source:
continue
else:
X_un = load_obj("%s/X_un"%source)
temp_un = concatenate(temp_un,X_un)
else:
if "large" not in target:
domains = ["books","dvd","electronics","kitchen"]
if target not in domains:
return
else:
domains =["large/baby","large/cell_phone","large/imdb","large/yelp2014"]
for source in domains:
if target == source:
continue
else:
print source
X_train = load_obj("%s/X_train"%source)
y_train = load_obj("%s/y_train"%source)
X_un = load_obj("%s/X_un"%source)
temp_X = concatenate(temp_X,X_train)
temp_y = concatenate(temp_y,y_train)
temp_un = concatenate(temp_un,X_un)
thetas = [0.5,0.6,0.7,0.8,0.9]
best_acc = 0.0
best_clf =""
best_theta = 0.0
resFile = open("../work/params/%s_theta_joint.csv"%domain,"w")
resFile.write("theta, acc\n")
for theta in thetas:
print "##############################"
print "start with theta=%s"%theta
print "##############################"
acc,clf_func = self_train(target,temp_X,temp_y,X_test,y_test,temp_un,theta=theta)
if best_acc<acc:
best_acc = acc
best_clf = clf_func
best_theta = theta
resFile.write("%f, %f\n"%(theta,acc))
resFile.flush()
resFile.close()
print "##############################"
print "best_theta:",best_theta,"best_acc:",best_acc
save_obj(best_clf,"%s/joint_clf"%target)
return acc,clf_func
def compute_psi(target,k=None):
pos_star = load_obj('%s/pos_star'%target)
neg_star = load_obj('%s/neg_star'%target)
star_matrix = concatenate(pos_star,neg_star)
# print star_matrix
X_joint = load_obj('%s/X_joint'%target)
y_joint = load_obj('%s/y_joint'%target)
# print np.array(X_joint).shape
src_cost = load_obj("%s/src_cost"%target)
# X_train = get_sents(X_joint)
# print np.matmul(star_matrix,X_train.T)
# psi_matrix = np.dot(star_matrix,X_train.T).T
#softmax(np.dot(star_matrix,X_train.T).T)
# print k
if k == None:
psi_matrix = []
for X_split in X_joint:
# print np.dot(star_matrix,np.array(X_split).T).T
# print np.array(X_split).shape
# temp = softmax(np.dot(star_matrix,np.array(X_split).T).T,axis=0)
temp = softmax(normalize(np.dot(star_matrix,np.array(X_split).T).T),axis=0)
psi_matrix.append(temp)
# print temp
save_obj(np.array(psi_matrix),"%s/psi_matrix"%(target))
# np.save("../data/%s/psi_matrix"%(target),np.array(psi_matrix))
else:
psi_matrix = []
X_psi = []
y_psi = []
cost_psi = []
X_index = []
for (X_split,y_split,cost_split) in zip(X_joint,y_joint,src_cost):
temp = normalize(np.dot(star_matrix,np.array(X_split).T).T)
# temp = np.dot(star_matrix,np.array(X_split).T).T
filtered,index = top_k(temp,k)
# print softmax(filtered,axis=0),index
psi_matrix.append(softmax(filtered,axis=0))
# print filtered,filtered.shape
X_temp = np.array(X_split)[index]
X_psi.append(X_temp)
y_temp = np.array(y_split)[index]
y_psi.append(y_temp)
cost_temp = np.array(cost_split)[index]
cost_psi.append(cost_temp)
X_index.append(index)
# print y_temp.shape
# print top_k(psi_matrix,k)
# print psi_matrix[0].sum(axis=0).shape,psi_matrix[0].sum(axis=0)
psi_matrix = np.array(psi_matrix)
X_psi = np.array(X_psi)
y_psi = np.array(y_psi)
cost_psi = np.array(cost_psi)
save_obj(psi_matrix,"%s/%s/psi_matrix"%(target,k))
save_obj(X_psi,"%s/%s/X_psi"%(target,k))
save_obj(y_psi,"%s/%s/y_psi"%(target,k))
save_obj(cost_psi,"%s/%s/src_cost_psi"%(target,k))
save_obj(X_index,"%s/%s/X_index"%(target,k))
# print sum([y for domain in y_psi for y in domain if y==1])
return np.array(psi_matrix)
def chen12(target="all"):
time_start = time.time()
amazon = np.load("../data/amazon.npz")
amazon_xx = coo_matrix(
(amazon['xx_data'], (amazon['xx_col'], amazon['xx_row'])),
shape=amazon['xx_shape'][::-1]).tocsc()
amazon_xx = amazon_xx[:, :5000]
amazon_yy = amazon['yy']
amazon_yy = (amazon_yy + 1) / 2
amazon_offset = amazon['offset'].flatten()
time_end = time.time()
print("Time used to process the Amazon data set = {} seconds.".format(
time_end - time_start))
print("Number of training instances = {}, number of features = {}.".format(
amazon_xx.shape[0], amazon_xx.shape[1]))
print("Number of nonzero elements = {}".format(amazon_xx.nnz))
print("amazon_xx shape = {}.".format(amazon_xx.shape))
print("amazon_yy shape = {}.".format(amazon_yy.shape))
data_name = ["books", "dvd", "electronics", "kitchen"]
num_data_sets = 4
data_insts, data_labels, num_insts = [], [], []
for i in range(num_data_sets):
data_insts.append(amazon_xx[amazon_offset[i]:amazon_offset[i + 1], :])
data_labels.append(amazon_yy[amazon_offset[i]:amazon_offset[i + 1], :])
print(
"Length of the {} data set label list = {}, label values = {}, label balance = {}"
.format(
data_name[i],
amazon_yy[amazon_offset[i]:amazon_offset[i + 1], :].shape[0],
np.unique(amazon_yy[amazon_offset[i]:amazon_offset[i + 1], :]),
np.sum(amazon_yy[amazon_offset[i]:amazon_offset[i + 1], :])))
num_insts.append(amazon_offset[i + 1] - amazon_offset[i])
# Randomly shuffle.
r_order = np.arange(num_insts[i])
np.random.shuffle(r_order)
data_insts[i] = data_insts[i][r_order, :]
data_labels[i] = data_labels[i][r_order, :]
print("Data sets: {}".format(data_name))
print("Number of total instances in the data sets: {}".format(num_insts))
num_trains = 2000
input_dim = amazon_xx.shape[1]
# convert to tf_idf vectors
# data_insts = tf_idf(data_insts)
for i in range(num_data_sets):
print data_name[i]
# Build source instances.
source_insts = []
source_labels = []
for j in range(num_data_sets):
if j != i:
source_insts.append(
data_insts[j][:num_trains, :].todense().astype(np.float32))
source_labels.append(
data_labels[j][:num_trains, :].ravel().astype(np.int64))
if target == data_name[i] or target == "all":
# Build test instances.
target_idx = i
target_insts = data_insts[i][num_trains:, :].todense().astype(
np.float32)
target_labels = data_labels[i][num_trains:, :].ravel().astype(
np.int64)
# train_data=np.concatenate((get_sents(source_insts),target_insts),axis=0)
# train_labels = np.concatenate((np.array(source_labels).flatten(),target_labels),axis=0)
train_data = get_sents(source_insts)
train_labels = np.array(source_labels).flatten()
# print target_insts
unlabel_data = data_insts[i][:num_trains, :].todense().astype(
np.float32)
train, test, X_un = mlp_vectors(train_data, train_labels,
get_sents(source_insts),
target_insts, unlabel_data)
mlp_source_insts = list(split_list(train, 3))
source_labels = list(
split_list(np.array(source_labels).flatten(), 3))
mlp_target_insts = test
clf = LogisticRegression().fit(train,
np.array(source_labels).flatten())
pred = clf.predict(test)
acc = accuracy_score(target_labels, pred)
print acc
if acc > 0.847:
save_obj(mlp_source_insts, "mlp/%s/X_joint" % data_name[i])
save_obj(source_labels, "mlp/%s/y_joint" % data_name[i])
save_obj(mlp_target_insts, "mlp/%s/X_test" % data_name[i])
save_obj(target_labels, "mlp/%s/y_test" % data_name[i])
save_obj(X_un, "mlp/%s/X_un" % data_name[i])
return acc
def recon():
# input and output folder
image_path = r'dataset'
save_path = 'output'
if not os.path.exists(save_path):
os.makedirs(save_path)
img_list = glob.glob(image_path + '/**/' + '*.png', recursive=True)
img_list += glob.glob(image_path + '/**/' + '*.jpg', recursive=True)
# read BFM face model
# transfer original BFM model to our model
if not os.path.isfile('BFM/BFM_model_front.mat'):
transfer_BFM09()
device = 'cuda:0' if torch.cuda.is_available() else 'cpu:0'
bfm = BFM(r'BFM/BFM_model_front.mat', device)
# read standard landmarks for preprocessing images
lm3D = bfm.load_lm3d()
model = resnet50_use().to(device)
model.load_state_dict(torch.load(r'models\params.pt'))
model.eval()
for param in model.parameters():
param.requires_grad = False
for file in img_list:
# load images and corresponding 5 facial landmarks
img, lm = load_img(file, file.replace('jpg', 'txt'))
# preprocess input image
input_img_org, lm_new, transform_params = Preprocess(img, lm, lm3D)
input_img = input_img_org.astype(np.float32)
input_img = torch.from_numpy(input_img).permute(0, 3, 1, 2)
# the input_img is BGR
input_img = input_img.to(device)
arr_coef = model(input_img)
coef = torch.cat(arr_coef, 1)
# reconstruct 3D face with output coefficients and face model
face_shape, face_texture, face_color, landmarks_2d, z_buffer, angles, translation, gamma = reconstruction(
coef, bfm)
fx, px, fy, py = estimate_intrinsic(landmarks_2d, transform_params,
z_buffer, face_shape, bfm, angles,
translation)
face_shape_t = transform_face_shape(face_shape, angles, translation)
face_color = face_color / 255.0
face_shape_t[:, :, 2] = 10.0 - face_shape_t[:, :, 2]
images = render_img(face_shape_t, face_color, bfm, 300, fx, fy, px, py)
images = images.detach().cpu().numpy()
images = np.squeeze(images)
path_str = file.replace(image_path, save_path)
path = os.path.split(path_str)[0]
if os.path.exists(path) is False:
os.makedirs(path)
from PIL import Image
images = np.uint8(images[:, :, :3] * 255.0)
# init_img = np.array(img)
# init_img[images != 0] = 0
# images += init_img
img = Image.fromarray(images)
img.save(file.replace(image_path, save_path).replace('jpg', 'png'))
face_shape = face_shape.detach().cpu().numpy()
face_color = face_color.detach().cpu().numpy()
face_shape = np.squeeze(face_shape)
face_color = np.squeeze(face_color)
save_obj(
file.replace(image_path,
save_path).replace('.jpg',
'_mesh.obj'), face_shape, bfm.tri,
np.clip(face_color, 0,
1.0)) # 3D reconstruction face (in canonical view)
from load_data import transfer_UV
from utils import process_uv
# loading UV coordinates
uv_pos = transfer_UV()
tex_coords = process_uv(uv_pos.copy())
tex_coords = torch.tensor(tex_coords,
dtype=torch.float32).unsqueeze(0).to(device)
face_texture = face_texture / 255.0
images = render_img(tex_coords, face_texture, bfm, 600, 600.0 - 1.0,
600.0 - 1.0, 0.0, 0.0)
images = images.detach().cpu().numpy()
images = np.squeeze(images)
# from PIL import Image
images = np.uint8(images[:, :, :3] * 255.0)
img = Image.fromarray(images)
img.save(
file.replace(image_path,
save_path).replace('.jpg', '_texture.png'))
def main():
parser = argparse.ArgumentParser()
parser.add_argument('-i',
'--filename-input',
type=str,
default=os.path.join('./data/obj/liu_mesh.obj'))
parser.add_argument('-o',
'--output-dir',
type=str,
default=os.path.join('./data/obj/vd092_mesh.obj'))
args = parser.parse_args()
# other settings
camera_distance = 3.0
elevation = -5
azimuth = 0
# load from Wavefront .obj file
mesh = sr.Mesh.from_obj(args.filename_input,
load_texture=True,
texture_res=5,
texture_type='vertex')
# create renderer with SoftRas
renderer = sr.SoftRenderer(camera_mode='look_at')
net = Net.ResNet([3, 4, 23, 3]).to(device)
im = torch.rand(1, 3, 200, 200).to(device)
out = net(im).to(device)
BFM_coeff = out
if not os.path.isfile('./BFM/BFM_model_front.mat'):
transferBFM09()
Face_model = BFM()
BFM_net = compute_bfm(
torch.tensor(Face_model.idBase, dtype=torch.float16),
torch.tensor(Face_model.exBase, dtype=torch.float16),
torch.tensor(Face_model.meanshape, dtype=torch.float16),
torch.tensor(Face_model.texBase, dtype=torch.float16),
torch.tensor(Face_model.meantex, dtype=torch.float16),
torch.tensor(Face_model.tri, dtype=torch.int32))
id_coeff = BFM_coeff[:, 0:80]
ex_coeff = BFM_coeff[:, 80:144]
tex_coeff = BFM_coeff[:, 144:224]
print(id_coeff)
vertices, textures, tri = BFM_net(id_coeff, ex_coeff, tex_coeff)
# draw object from different view
mesh.reset_()
elevation = BFM_coeff[:, 226]
# elevation = torch.sum(elevation)
azimuth = -90
renderer.transform.set_eyes_from_angles(camera_distance, 0, 180)
# images = renderer.render_mesh(mesh)
print(vertices)
print(mesh.faces)
faces = torch.tensor(Face_model.tri, dtype=torch.int32).to(device) - 1
faces = faces.unsqueeze(0)
print(faces)
images = renderer.forward(mesh.vertices,
mesh.faces,
mesh.textures,
texture_type='vertex')
print(images)
image = images.detach().cpu().numpy()[0, 0:3].transpose((1, 2, 0))
image = (image * 255).astype(np.uint8)
plt.figure(0)
plt.imshow(image)
plt.show()
img_name1 = 'D:/files/project/data/human_faces/CACD2000/CACD2000/17_Jennifer_Lawrence_0013.jpg'
img_name2 = 'D:/files/project/data/human_faces/CACD2000/CACD2000/17_Lily_Cole_0008.jpg'
predictor_model = './model/shape_predictor_68_face_landmarks.dat'
detector = dlib.get_frontal_face_detector() # dlib人脸检测器
predictor = dlib.shape_predictor(predictor_model)
img1 = cv2.imread(img_name1)
image2 = cv2.imread(img_name2)
image1 = cv2.cvtColor(img1, cv2.COLOR_BGR2RGB)
# image2 = cv2.cvtColor(img2, cv2.COLOR_BGR2RGB)
points1 = np.zeros((68, 2))
points2 = np.zeros((68, 2))
rects, scores, idx = detector.run(image, 2, 0)
faces = dlib.full_object_detections()
for rect in rects:
faces.append(predictor(image, rect))
i = 0
for landmark in faces:
for idx, point in enumerate(landmark.parts()):
points1[i, 0] = point.x
points1[i, 1] = point.y
i = i + 1
h, w, c = image.shape
show_image(image, points1)
rects, scores, idx = detector.run(image2, 2, 0)
faces = dlib.full_object_detections()
for rect in rects:
faces.append(predictor(image2, rect))
i = 0
for landmark in faces:
for idx, point in enumerate(landmark.parts()):
points2[i, 0] = point.x
points2[i, 1] = point.y
i = i + 1
h, w, c = image2.shape
show_image(image2, points2)
tr = trans.estimate_transform('affine', src=points1, dst=points2)
M = tr.params[0:2, :]
cv_img = cv2.warpAffine(image1, M, (image.shape[1], image.shape[0]))
show_image(image2, points2)
param = np.linalg.inv(tr.params)
theta = normalize_transforms(param[0:2, :], w, h)
to_tensor = torchvision.transforms.ToTensor()
tensor_img = to_tensor(image).unsqueeze(0)
theta = torch.Tensor(theta).unsqueeze(0)
grid = F.affine_grid(theta, tensor_img.size())
tensor_img = F.grid_sample(tensor_img, grid)
tensor_img = tensor_img.squeeze(0)
warp_img = convert_image_np(tensor_img)
show_image(warp_img, points2)
vertices = vertices[0].detach().cpu().numpy()
faces = faces[0].detach().cpu().numpy() + 1
textures = textures[0].detach().cpu().numpy()
load.save_obj('./123.obj', vertices, faces, textures)
def label_to_tensor(v):
# print len(v)
return Variable(torch.FloatTensor(v).view(len(v), -1).to(device))
def softmax(x):
"""Compute softmax values for each sets of scores in x."""
e_x = np.exp(x - np.max(x))
return e_x / e_x.sum(axis=0) # only difference
if __name__ == '__main__':
if len(sys.argv) > 2:
target = sys.argv[1]
test_index = int(sys.argv[2])
# option = sys.argv[3]
evd = compute_single_evidence(target, test_index, tops=2000)
collected = contruct_collection(evd, option='all')
save_obj(collected, 'collected-all')
collected = contruct_collection(evd)
save_obj(collected, 'collected-theta')
# elif len(sys.argv) >2:
# target = sys.argv[1]
# # k = int(sys.argv[2])
# # tops = int(sys.argv[3])
# test_index = int(sys.argv[2])
# evd = compute_single_evidence(target,test_index,tops=2000)
else:
print "<target,test_index>"
