def show_reconstruction(model, img, n_images, args, num):
model.eval()
x = img
_, x_recon = model(x)
data_o = np.concatenate([x.cpu().data])
data_r = np.concatenate([x_recon.cpu().data])
data = np.concatenate([data_r])
img = combine_images(np.transpose(data_o, [0, 2, 3, 1]))
image = img * 255
Image.fromarray(image.astype(np.uint8)).save(args + str(num)+ "_" +"real.png")
img = combine_images(np.transpose(data_r, [0, 2, 3, 1]))
image = img * 255
Image.fromarray(image.astype(np.uint8)).save(args + str(num)+ "_" + "recon.png")
rel = cv2.imread(args + str(num)+ "_" + "real.png", 0)
mask = cv2.imread(args + str(num)+ "_" + "recon.png", 0)
#draw_mask_edge_on_image_cv2(rel, mask, str(num))
img = combine_images(np.transpose(data, [0, 2, 3, 1]))
image = img * 255
Image.fromarray(image.astype(np.uint8)).save(args + str(num)+ "_" + "recon.png")
# print()
# print('Reconstructed images are saved to %s/real_and_recon_color.png' % args.save_dir)
# print('-' * 70)
# plt.figure()
# plt.imshow(plt.imread(args.save_dir + "/real_and_recon.png", )) #matplot imshow() WILL AUTO CHANGE ONE CHANNEL IMG TO COLOR ONE
plt.show()
def start_message(message):
bot.send_message(message.chat.id, 'Привет, ты написал мне /start')
room_id = message.chat.id
#bot.send_message(message.chat.id, 'Привет, ты написал мне /start')
#bot.send_photo(chat_id=room_id, photo=open('images/resized-2C.png', 'rb'))
player_cards = get_random_cards()
score = get_score(player_cards)
bot.send_message(message.chat.id,
'Вы набрали: %s!' % score,
reply_markup=markup)
combine_images(player_cards, 'player.png')
bot.send_photo(chat_id=room_id, photo=open('player.png', 'rb'))
TEST = 'ffffff'
print(TEST)
def show_reconstruction(model, test_loader, n_images, args):
import matplotlib.pyplot as plt
from utils import combine_images
from PIL import Image
import numpy as np
distance = torch.zeros((1, )).cuda()
adversary = GradientSignAttack(model, loss_fn=caps_loss, eps=0.3)
model.eval()
for x, y in test_loader:
y = torch.zeros(y.size(0), 10).scatter_(1, y.view(-1, 1), 1.)
x, y = Variable(x.cuda()), Variable(y.cuda())
x_adv = adversary.perturb(x, args.lam_recon, y)
_, x_recon = model(x_adv)
distance = compute_distance(x_adv, x_recon, distance)
data = np.concatenate([x.cpu().data, x_recon.cpu().data])
img = combine_images(np.transpose(data, [0, 2, 3, 1]))
image = img * 255
Image.fromarray(image.astype(np.uint8)).save(args.save_dir +
"/real_and_recon.png")
print()
print('Reconstructed images are saved to %s/real_and_recon.png' %
args.save_dir)
print('-' * 70)
plt.imshow(plt.imread(args.save_dir + "/real_and_recon.png", ))
plt.show()
break
distance = distance[1:]
print(distance)
def show_reconstruction(self, test_loader, n_images):
import matplotlib.pyplot as plt
from utils import combine_images
from PIL import Image
import numpy as np
self.model.eval()
for x, _ in test_loader:
if self.use_cuda:
x = Variable(x[:min(n_images, x.size(0))].cuda(),
volatile=True)
else:
x = Variable(x[:min(n_images, x.size(0))], volatile=True)
_, x_recon = self.model(x)
data = np.concatenate([x.data.cpu(), x_recon.data.cpu()])
img = combine_images(np.transpose(data, [0, 2, 3, 1]))
image = img * 255
Image.fromarray(image.astype(np.uint8)).save(self.config.save_dir +
"/real_and_recon.png")
print()
print('Reconstructed images are saved to %s/real_and_recon.png' %
config.save_dir)
print('-' * 70)
plt.imshow(plt.imread(config.save_dir + "/real_and_recon.png", ))
plt.show()
break
def test_transformer(config, netG, train_iterators, monitor, param_file):
netG_A2B = netG['netG_A2B']
train_iterator_src, train_iterator_trg = train_iterators
# Load boundary image to get Variable shapes
bod_map_A = train_iterator_src.next()[0]
bod_map_B = train_iterator_trg.next()[0]
real_bod_map_A = nn.Variable(bod_map_A.shape)
real_bod_map_B = nn.Variable(bod_map_B.shape)
real_bod_map_A.persistent, real_bod_map_B.persistent = True, True
################### Graph Construction ####################
# Generator
with nn.parameter_scope('netG_transformer'):
with nn.parameter_scope('netG_A2B'):
fake_bod_map_B = netG_A2B(
real_bod_map_A, test=True,
norm_type=config["norm_type"]) # (1, 15, 64, 64)
fake_bod_map_B.persistent = True
# load parameters of networks
with nn.parameter_scope('netG_transformer'):
with nn.parameter_scope('netG_A2B'):
nn.load_parameters(param_file)
monitor_vis = nm.MonitorImage('result',
monitor,
interval=config["test"]["vis_interval"],
num_images=1,
normalize_method=lambda x: x)
# Test
i = 0
iter_per_epoch = train_iterator_src.size // config["test"]["batch_size"] + 1
if config["num_test"]:
num_test = config["num_test"]
else:
num_test = train_iterator_src.size
for _ in range(iter_per_epoch):
bod_map_A = train_iterator_src.next()[0]
bod_map_B = train_iterator_trg.next()[0]
real_bod_map_A.d, real_bod_map_B.d = bod_map_A, bod_map_B
# Generate fake images
fake_bod_map_B.forward(clear_buffer=True)
i += 1
images_to_visualize = [
real_bod_map_A.d, fake_bod_map_B.d, real_bod_map_B.d
]
visuals = combine_images(images_to_visualize)
monitor_vis.add(i, visuals)
if i > num_test:
break
def manipulate_latent(model, data, args):
x_true, y_true = data
index = np.argmax(y_true, 1) == args.digit
number = np.random.randint(low=0, high=sum(index) - 1)
x, y = x_true[index][number], y_true[index][number]
x, y = np.expand_dims(x, 0), np.expand_dims(y, 0)
noise = np.zeros([1, 10, 16])
x_recons = []
# Change params of vect in 0.05 steps. See also [1]
for dim in range(16):
for r in [
-0.25, -0.2, -0.15, -0.1, -0.05, 0, 0.05, 0.1, 0.15, 0.2, 0.25
]:
tmp = np.copy(noise)
tmp[:, :, dim] = r
x_recon = model.predict([x, y, tmp])
x_recons.append(x_recon)
x_recons = np.concatenate(x_recons)
img = utils.combine_images(x_recons, height=16)
image = img * 255
Image.fromarray(image.astype(
np.uint8)).save(args.save_dir + '/manipulate-%d.png' % args.digit)
print('Manipulated result saved to %s/manipulate-%d.png' %
(args.save_dir, args.digit))
def test_rotated_images(model, data):
data2 = load_rotated_mnist()
x_test, y_test = data2
print("x_test_shape :", x_test.shape)
print("y_test_shape :", y_test.shape)
y_pred, x_recon = model.predict(x_test, batch_size=100)
print("y_pred_test :", y_pred.shape)
print("x_recon_test :", x_recon.shape)
print('-' * 50)
print(
'Test acc:',
np.sum(np.argmax(y_pred, 1) == np.argmax(y_test, 1)) / y_test.shape[0])
import matplotlib.pyplot as plt
from utils import combine_images
from PIL import Image
img = combine_images(np.concatenate([x_test[:50], x_recon[:50]]))
image = img * 255
Image.fromarray(image.astype(np.uint8)).save("real_and_recon.png")
print()
print('Reconstructed images are saved to ./real_and_recon.png')
print('-' * 50)
plt.imshow(plt.imread("real_and_recon.png", ))
plt.show()
def manipulate_latent(model, data, args):
print('-' * 30 + 'Begin: manipulate' + '-' * 30)
x_test, y_test = data
index = np.argmax(y_test, 1) == args.digit
number = np.random.randint(low=0, high=sum(index) - 1)
x, y = x_test[index][number], y_test[index][number]
x, y = np.expand_dims(x, 0), np.expand_dims(y, 0)
noise = np.zeros([1, 2, 16]) # 10->2
x_recons = []
for dim in range(16):
for r in [
-0.25, -0.2, -0.15, -0.1, -0.05, 0, 0.05, 0.1, 0.15, 0.2, 0.25
]:
tmp = np.copy(noise)
tmp[:, :, dim] = r
x_recon = model.predict([x, y, tmp])
x_recons.append(x_recon)
x_recons = np.concatenate(x_recons)
img = combine_images(x_recons, height=16)
# image = img*255
# Image.fromarray(image.astype(np.uint8)).save(args.save_dir + '/manipulate-%d.png' % args.digit)
# Image.fromarray(img).save(args.save_dir + '/manipulate-%d.png' % args.digit)
# save_images(x_recons,image_manifold_size(x_recons.shape[0]), args.save_dir + '/manipulate-%d.png' % args.digit)
img_path = args.save_dir + '/manipulate-%d.png' % args.digit
scipy.misc.imsave(
img_path,
img) # directly save the combned reconstructed BIGGER image!
print('manipulated result saved to %s/manipulate-%d.png' %
(args.save_dir, args.digit))
print('-' * 30 + 'End: manipulate' + '-' * 30)
def show_reconstruction(model, test_loader, n_images, args):
import matplotlib.pyplot as plt
from utils import combine_images
from PIL import Image
import numpy as np
model.eval()
for x, _, _ in test_loader:
# x = x.float()
# x = x/256
x = x.view_as(torch.Tensor(100, 1, 98, 60))
x = Variable(x[:min(n_images, x.size(0))].cuda(), volatile=True)
_, x_recon = model(x)
data = np.concatenate([x.data, x_recon.data])
img = combine_images(np.transpose(data, [0, 2, 3, 1]))
image = img * 255
Image.fromarray(image.astype(np.uint8)).save(args.save_dir +
"/real_and_recon.png")
print()
print('Reconstructed images are saved to %s/real_and_recon.png' %
args.save_dir)
print('-' * 70)
plt.imshow(plt.imread(args.save_dir + "/real_and_recon.png", ))
plt.show()
break
def test(model, data, sample):
x_test, y_test = data
#print "x_test shape is: ",x_test.shape
angle = 0
sample = sample
one_hots = np.diag(np.ones(10))
x_test_new = np.array([x_test[sample] for i in xrange(9)])
for i in xrange(0,9):
a = mc.imrotate(x_test[sample,:,:,0],angle)
x_test_new[i,:,:,0] = a
angle = angle + 10
digitcaps, y_pred, x_recon = model.predict([x_test_new, one_hots], batch_size=9)
#print('-'*50)
#print('Test acc:', np.sum(np.argmax(y_pred, 1) == np.argmax(y_test, 1))*1.0/y_test.shape[0])
#print "shape is ",digitcaps.shape
label = np.argmax(y_test[sample])
#print "labels is: ",y_test[sample]," ",label
for i in xrange(0,9):
print np.dot(digitcaps[0,label,:],digitcaps[i,label,:])/(np.linalg.norm(digitcaps[0,label,:])*np.linalg.norm(digitcaps[i,label,:]))
#print digitcaps[:,label,:]
import matplotlib.pyplot as plt
from utils import combine_images
from PIL import Image
img = combine_images(np.concatenate([np.array([x_test_new[0]]),x_recon]))
image = img * 255
Image.fromarray(image.astype(np.uint8)).save("real_and_recon_fmnist_rot.png")
#print()
#print('Reconstructed images are saved to ./real_and_recon_fmnist_rot.png')
print('-'*50)
def manipulate_latent(model, data, args):
print('-' * 30 + 'Begin: manipulate' + '-' * 30)
x_test, y_test = data
index = np.argmax(y_test, 1) == args.digit
number = np.random.randint(low=0, high=sum(index) - 1)
x, y = x_test[index][number], y_test[index][number]
x, y = np.expand_dims(x, 0), np.expand_dims(y, 0)
noise = np.zeros([1, 10, 16])
x_recons = []
for dim in range(16):
for r in [
-0.25, -0.2, -0.15, -0.1, -0.05, 0, 0.05, 0.1, 0.15, 0.2, 0.25
]:
tmp = np.copy(noise)
tmp[:, :, dim] = r
x_recon = model.predict([x, y, tmp])
x_recons.append(x_recon)
x_recons = np.concatenate(x_recons)
img = combine_images(x_recons, height=16)
image = img * 255
Image.fromarray(image.astype(
np.uint8)).save(args.save_dir + '/manipulate-%d.png' % args.digit)
print('manipulated result saved to %s/manipulate-%d.png' %
(args.save_dir, args.digit))
print('-' * 30 + 'End: manipulate' + '-' * 30)
def test(model, data, args):
print('-' * 30 + 'Begin: test' + '-' * 30)
x_test, y_test = data
print('Testing on {0} images'.format(len(y_test)))
print(np.argmax(y_test, axis=1))
y_pred, x_recon = model.predict(x_test)
true_lab = []
pred_lab = []
TP = 0
FP = 0
TN = 0
FN = 0
for i in range(0, len(y_pred)):
t = np.argmax(y_test[i])
p = np.argmax(y_pred[i])
print("GT: " + str(t) + " Pred: " + str(p) + "--------- GT: " +
str(y_test[i]) + " Pred: " + str(y_pred[i]))
true_lab.append(t)
pred_lab.append(p)
if t == p == 1:
TP += 1
if p == 1 and t != p:
FP += 1
if t == p == 0:
TN += 1
if p == 0 and t != p:
FN += 1
print("GT: " + str(t) + " Pred: " + str(p) + "--------- GT: " +
str(y_test[i]) + " Pred: " + str(y_pred[i]))
from sklearn.metrics import roc_auc_score, accuracy_score, confusion_matrix
# conf_mat = confusion_matrix(y_test, y_pred)
print("Accuracy {0}".format(
accuracy_score(np.argmax(y_test, axis=1), np.argmax(y_pred, axis=1))))
roc = roc_auc_score(true_lab, pred_lab)
print("TP: {0}, FP: {1}, TN: {2}, FN: {3}".format(TP, FP, TN, FN))
print("Accuracy: {0}".format((TP + TN) / (TP + TN + FP + FN)))
print("AUC ROC : " + str(roc))
print("F1 score: {0}".format(2 * TP / (2 * TP + FP + FN)))
print("Sensitivity(TPR): {0}, Specificity(TNR): {1}".format(
TP / (TP + FN), TN / (FP + TN)))
print("PPV(Precision): {0}".format(TP / (TP + FP)))
print("NPV: {0}".format(TN / (TN + FN)))
img = utils.combine_images(np.concatenate([x_test[:50], x_recon[:50]]))
rescaled = (255.0 / img.max() * (img - img.min())).astype(np.uint8)
Image.fromarray(rescaled).save(args.save_dir + "/real_and_recon.png")
print()
print('Reconstructed images are saved to %s/real_and_recon.png' %
args.save_dir)
print('-' * 30 + 'End: test' + '-' * 30)
def test(model, data, args):
time_date = datetime.datetime.now().strftime("%y-%m-%d-%H-%M")
filename = time_date + '_capsule_network' + '_batch_size=' + str(
args.batch_size) + '_epochs=' + str(args.epochs) + '_ntr=' + str(
x_train.shape[0]) + '_nch=' + str(x_train.shape[3])
x_test, y_test = data
y_pred, x_recon = model.predict(x_test, batch_size=100)
print('-' * 30 + 'Begin: test' + '-' * 30)
img = combine_images(
np.concatenate([np.abs(x_test[:50]),
np.abs(x_recon[:50])]))
image = cm.hot(img) * 255
Image.fromarray(image.astype(np.uint8)).save(args.save_dir +
"/real_and_recon4.png")
print('Reconstructed images are saved to %s/real_and_recon4.png' %
args.save_dir)
print('-' * 30 + 'End: test' + '-' * 30)
predicted_classes = y_pred
predicted_classes = np.asarray(predicted_classes)
b = np.zeros_like(predicted_classes)
b[np.arange(len(predicted_classes)), predicted_classes.argmax(1)] = 1
predicted_classes = np.argmax(np.round(b), axis=1)
b = np.zeros_like(y_test)
b[np.arange(len(y_test)), y_test.argmax(1)] = 1
y_test = np.argmax(np.round(b), axis=1)
class_names = ['Unresolved', 'FRI', 'FRII']
from sklearn.metrics import classification_report
print(
classification_report(y_test,
predicted_classes,
target_names=class_names,
digits=4))
os.chdir(args.save_dir)
filename1 = filename + '_classification_report.txt'
f = open(filename1, 'w')
f.write(
classification_report(y_test,
predicted_classes,
target_names=class_names,
digits=4))
f = open(filename1, 'w')
f.write(
classification_report(y_test,
predicted_classes,
target_names=class_names,
digits=4))
############# Architecture
with open(filename + '_architecture.txt', 'w') as fh:
model.summary(print_fn=lambda x: fh.write(x + '\n'))
with open(filename + '_architecture.txt', 'w') as fh:
model.summary(print_fn=lambda x: fh.write(x + '\n'))
def show_reconstruction(model, test_loader, n_images, args):
model.eval()
for x, _ in test_loader:
x = Variable(x[:min(n_images, x.size(0))].cuda(), volatile=True)
_, x_recon = model(x)
print(x.shape)
data_o = np.concatenate([x.cpu().data])
data_r = np.concatenate([x_recon.cpu().data])
data = np.concatenate([data_r])
img = combine_images(np.transpose(data_o, [0, 2, 3, 1]))
image = img * 255
Image.fromarray(image.astype(np.uint8)).save(args.save_dir + "/real.png")
img = combine_images(np.transpose(data_r, [0, 2, 3, 1]))
image = img * 255
Image.fromarray(image.astype(np.uint8)).save(args.save_dir + "/recon.png")
rel = cv2.imread(args.save_dir + "/real.png", 0)
mask = cv2.imread(args.save_dir + "/recon.png", 0)
draw_mask_edge_on_image_cv2(rel, mask)
img = combine_images(np.transpose(data, [0, 2, 3, 1]))
image = img * 255
Image.fromarray(image.astype(np.uint8)).save(args.save_dir + "/recon.png")
print()
print('Reconstructed images are saved to %s/real_and_recon.png' % args.save_dir)
print('-' * 70)
im_gray = cv2.imread(args.save_dir + "/real_and_recon.png", cv2.IMREAD_GRAYSCALE)
im_color = cv2.applyColorMap(im_gray, cv2.COLORMAP_HSV) # cv2 applycolorap can change an one-channel to color one
plt.figure()
plt.imshow(im_color, cmap='jet_r')
Image.fromarray(im_color.astype(np.uint8)).save(args.save_dir + "/real_and_recon_color.png")
# print()
# print('Reconstructed images are saved to %s/real_and_recon_color.png' % args.save_dir)
# print('-' * 70)
# plt.figure()
# plt.imshow(plt.imread(args.save_dir + "/real_and_recon.png", )) #matplot imshow() WILL AUTO CHANGE ONE CHANNEL IMG TO COLOR ONE
plt.show()
break
def save_output_image(self, samples, image_name):
"""
Visualizing and saving images in the .png format
:param samples: images to be visualized
:param image_name: name of the saved .png file
"""
if not os.path.exists(args.save_dir + "/images"):
os.makedirs(args.save_dir + "/images")
img = combine_images(samples)
img = img * 255
Image.fromarray(img.astype(np.uint8)).save(args.save_dir + "/images/" +
image_name + ".png")
print(image_name, "Image saved.")
def test(config, netG, train_iterator, monitor, param_file):
# Load image and boundary image to get Variable shapes
img, bod_map, bod_map_resize = train_iterator.next()
real_img = nn.Variable(img.shape)
real_bod_map = nn.Variable(bod_map.shape)
real_bod_map_resize = nn.Variable(bod_map_resize.shape)
################### Graph Construction ####################
# Generator
with nn.parameter_scope('netG_decoder'):
fake_img = netG(real_bod_map, test=False)
fake_img.persistent = True
# load parameters of networks
with nn.parameter_scope('netG_decoder'):
nn.load_parameters(param_file)
monitor_vis = nm.MonitorImage('result',
monitor,
interval=config["test"]["vis_interval"],
num_images=4,
normalize_method=lambda x: x)
# Test
i = 0
iter_per_epoch = train_iterator.size // config["test"]["batch_size"] + 1
if config["num_test"]:
num_test = config["num_test"]
else:
num_test = train_iterator.size
for _ in range(iter_per_epoch):
img, bod_map, bod_map_resize = train_iterator.next()
real_img.d = img
real_bod_map.d = bod_map
real_bod_map_resize.d = bod_map_resize
# Generate fake image
fake_img.forward(clear_buffer=True)
i += 1
images_to_visualize = [real_bod_map_resize.d, fake_img.d, img]
visuals = combine_images(images_to_visualize)
monitor_vis.add(i, visuals)
if i > num_test:
break
def test(model, data, args):
x_test, y_test = data
y_pred, x_recon = model.predict(x_test, batch_size=100)
print('-'*30 + 'Begin: test' + '-'*30)
print('Test acc:', np.sum(np.argmax(y_pred, 1) == np.argmax(y_test, 1))/y_test.shape[0])
img = combine_images(np.concatenate([x_test[:50],x_recon[:50]]))
image = img * 255
Image.fromarray(image.astype(np.uint8)).save(args.save_dir + "/real_and_recon.png")
print()
print('Reconstructed images are saved to %s/real_and_recon.png' % args.save_dir)
print('-' * 30 + 'End: test' + '-' * 30)
plt.imshow(plt.imread(args.save_dir + "/real_and_recon.png"))
plt.show()
def test(model, data, batch_size, args):
x_test, y_test = data
#y_test = np.argmax(y_test,axis=1)
y_pred, x_recon = model.predict(x_test, batch_size=batch_size)
print('-' * 30 + 'Begin: test' + '-' * 30)
print(
'Test acc:',
np.sum(np.argmax(y_pred, 1) == np.argmax(y_test, 1)) / y_test.shape[0])
img = combine_images(np.concatenate([x_test[:50], x_recon[:50]]))
image = img * 255
Image.fromarray(image.astype(np.uint8)).save(args.save_dir +
"/real_and_recon.png")
print()
print('Reconstructed images are saved to %s/real_and_recon.png' %
args.save_dir)
print('-' * 30 + 'End: test' + '-' * 30)
plt.imshow(plt.imread(args.save_dir + "/real_and_recon.png"))
plt.show()
y_pred = np.argmax(y_pred, axis=1)
y_test = np.argmax(y_test, axis=1)
# Print classification report
print("Classification Report")
print(['exC', 'mddC', 'nC ', 'mC', 'mduC'])
print('')
print(" Test Accuracy : " + str(accuracy_score(y_test, y_pred)))
print("")
print('')
#y_test = np.argmax(y_test,axis=1)
#y_pred = np.argmax(y_pred, axis=1)
#print(classification_report(y_true,y_pred,digits=5))
print(
classification_report(
y_test, y_pred, target_names=['exC', 'mddC', 'nC ', 'mC', 'mduC']))
print('')
cnf_matrix = confusion_matrix(y_test, y_pred)
print(cnf_matrix)
print('')
def test(model, data, args):
x_test, y_test = data
y_pred, x_recon = model.predict(x_test, batch_size=100)
print('-'*30 + 'Begin: test' + '-'*30)
print('Test acc:', np.sum(np.argmax(y_pred, 1) == np.argmax(y_test, 1))/y_test.shape[0])
img = combine_images(np.concatenate([x_test[::10][:50],x_recon[::10][:50]]))
image = img * 255
Image.fromarray(image.astype(np.uint8)).save(args.save_dir + "/real_and_recon.png")
print()
print('Reconstructed images are saved to %s/real_and_recon.png' % args.save_dir)
print('-' * 30 + 'End: test' + '-' * 30)
plt.imshow(plt.imread(args.save_dir + "/real_and_recon.png"))
plt.show()
def test(model, data, args):
# Create an augmentation function and cache augmented samples
# to be displayed later
x_augmented = []
def test_generator_with_augmentation(x, batch_size, shift_range,
rotation_range):
test_datagen = ImageDataGenerator(width_shift_range=shift_range,
height_shift_range=shift_range,
rotation_range=rotation_range)
generator = test_datagen.flow(x, batch_size=batch_size, shuffle=False)
while 1:
x_batch = generator.next()
x_augmented.extend(x_batch)
yield (x_batch)
# Run predictions
test_batch_size = 100
x_true, y_true = data
generator = test_generator_with_augmentation(x_true, test_batch_size,
args.shift_fraction,
args.rotation_range)
y_pred, x_recon = model.predict_generator(generator=generator,
steps=len(x_true) //
test_batch_size)
# Print different metrics using the top score
y_true = np.argmax(y_true, 1)
y_pred = np.argmax(y_pred, 1)
print('Confusion matrix:\n', confusion_matrix(y_true, y_pred))
print('\nAccuracy: ', accuracy_score(y_true, y_pred))
print('Recall: ', recall_score(y_true, y_pred, average='weighted'))
print('Precision: ', precision_score(y_true, y_pred, average='weighted'))
print('F1-Score: ', f1_score(y_true, y_pred, average='weighted'))
img = utils.combine_images(np.concatenate([x_augmented[:50],
x_recon[:50]]))
image = img * 255
print('\nReconstructed images are saved to %s/real_and_recon.png' %
args.save_dir)
Image.fromarray(image.astype(np.uint8)).save(args.save_dir +
"/real_and_recon.png")
plt.imshow(plt.imread(args.save_dir + "/real_and_recon.png"))
plt.show()
def test(model, data):
x_test, y_test = data
y_pred, x_recon = model.predict(x_test, batch_size=100)
print('-'*50)
print('Test acc:', np.sum(np.argmax(y_pred, 1) == np.argmax(y_test, 1))/y_test.shape[0])
import matplotlib.pyplot as plt
from utils import combine_images
from PIL import Image
img = combine_images(np.concatenate([x_test[:50],x_recon[:50]]))
image = img * 255
Image.fromarray(image.astype(np.uint8)).save("real_and_recon.png")
print()
print('Reconstructed images are saved to ./real_and_recon.png')
print('-'*50)
plt.imshow(plt.imread("real_and_recon.png", ))
plt.show()
def test(model, data, args):
x_test, y_test = data
print('Testing the model...')
y_pred, x_recon = model.predict(x_test, batch_size=100)
print(
'Test Accuracy: ',
100.0 * np.sum(np.argmax(y_pred, 1) == np.argmax(y_test, 1)) /
(1.0 * y_test.shape[0]))
img = combine_images(np.concatenate([x_test[:50], x_recon[:50]]))
image = img * 255
Image.fromarray(image.astype(np.uint8)).save(args.save_dir +
"/real_and_recon.png")
print()
print('Reconstructed images are saved to %s/real_and_recon.png' %
args.save_dir)
plt.imshow(plt.imread(args.save_dir + "/real_and_recon.png"))
plt.show()
def test(model):
generator = captcha_generator('../recaptcha_capsnet_keras/recaptcha', 1,
args.batch_size, False)
x_test, y_test = next(generator)
y_pred, x_recon = model.predict(x_test, batch_size=100)
print('-' * 50)
print(
'Test acc:',
np.sum(np.argmax(y_pred, 1) == np.argmax(y_test, 1)) / y_test.shape[0])
print('Test loss:', margin_loss(y_test, y_pred))
import matplotlib.pyplot as plt
from utils import combine_images
from PIL import Image
img = combine_images(np.concatenate([x_test[:50], x_recon[:50]]))
image = img * 255
Image.fromarray(image.astype(np.uint8)).save("real_and_recon.png")
print()
print('Reconstructed images are saved to ./real_and_recon.png')
print('-' * 50)
def show_reconstruction(model, test_loader, n_images, args):
import matplotlib.pyplot as plt
from utils import combine_images, save_image # in utils.py save_image is the function that can change one-channel tensor to channels, but still black and white
from PIL import Image
import numpy as np
import cv2
import os.path
import glob
model.eval()
for x, _ in test_loader:
x = Variable(x[:min(n_images, x.size(0))].cuda(), volatile=True)
_, x_recon = model(x)
data = np.concatenate([x.cpu().data, x_recon.cpu().data])
img = combine_images(np.transpose(data, [0, 2, 3, 1]))
image = img * 255
Image.fromarray(image.astype(np.uint8)).save(args.save_dir +
"/real_and_recon.png")
print()
print('Reconstructed images are saved to %s/real_and_recon.png' %
args.save_dir)
print('-' * 70)
im_gray = cv2.imread(args.save_dir + "/real_and_recon.png",
cv2.IMREAD_GRAYSCALE)
im_color = cv2.applyColorMap(
im_gray, cv2.COLORMAP_HSV
) # cv2 applycolorap can change an one-channel to color one
plt.figure()
plt.imshow(im_color)
Image.fromarray(im_color.astype(
np.uint8)).save(args.save_dir + "/real_and_recon_color.png")
print()
print('Reconstructed images are saved to %s/real_and_recon_color.png' %
args.save_dir)
print('-' * 70)
plt.figure()
plt.imshow(plt.imread(
args.save_dir + "/real_and_recon.png",
)) #matplot imshow() WILL AUTO CHANGE ONE CHANNEL IMG TO COLOR ONE
plt.show()
break
def test(model, data):
x_test, y_test = data
one_hots = np.diag(np.ones(10))
x_test_new = np.array([x_test[0] for i in xrange(10)])
y_pred, x_recon = model.predict([x_test_new, one_hots], batch_size=10)
#print('-'*50)
#print('Test acc:', np.sum(np.argmax(y_pred, 1) == np.argmax(y_test, 1))*1.0/y_test.shape[0])
import matplotlib.pyplot as plt
from utils import combine_images
from PIL import Image
img = combine_images(np.concatenate([np.array([x_test_new[0]]), x_recon]))
image = img * 255
Image.fromarray(image.astype(
np.uint8)).save("real_and_recon_fmnist_rot.png")
print()
print('Reconstructed images are saved to ./real_and_recon_fmnist_rot.png')
print('-' * 50)
plt.imshow(plt.imread("real_and_recon_fmnist_rot.png", ))
plt.show()
def show_reconstruction(model, test_loader, n_images, args):
import matplotlib
matplotlib.use('agg')
from utils import combine_images
import numpy as np
import cv2
model.eval()
for x, _ in test_loader:
with torch.no_grad():
x = Variable(x[:min(n_images, x.size(0))].cuda())
_, x_recon = model(x)
data = np.concatenate([x.data.cpu(), x_recon.data.cpu()])
img = combine_images(np.transpose(data, [0, 2, 3, 1]))
image = (img * 255).astype(np.uint8)
cv2.imwrite(args.save_dir + "/output.png", image)
print()
print('Reconstructed images are saved to %s/output.png' %
args.save_dir)
print('-' * 70)
cv2.imshow('Image', image)
cv2.waitKey(0)
break
def manipulate_latent(model, data, args):
print('-'*30 + 'Begin: manipulate' + '-'*30)
x_test, y_test = data
index = np.argmax(y_test, 1) == args.digit
number = np.random.randint(low=0, high=sum(index) - 1)
x, y = x_test[index][number], y_test[index][number]
x, y = np.expand_dims(x, 0), np.expand_dims(y, 0)
noise = np.zeros([1, 10, 16])
x_recons = []
for dim in range(16):
for r in [-0.25, -0.2, -0.15, -0.1, -0.05, 0, 0.05, 0.1, 0.15, 0.2, 0.25]:
tmp = np.copy(noise)
tmp[:,:,dim] = r
x_recon = model.predict([x, y, tmp])
x_recons.append(x_recon)
x_recons = np.concatenate(x_recons)
img = combine_images(x_recons, height=16)
image = img*255
Image.fromarray(image.astype(np.uint8)).save(args.save_dir + '/manipulate-%d.png' % args.digit)
print('manipulated result saved to %s/manipulate-%d.png' % (args.save_dir, args.digit))
print('-' * 30 + 'End: manipulate' + '-' * 30)
def show_reconstruction(model, test_loader, n_images, args):
import matplotlib.pyplot as plt
from utils import combine_images
from PIL import Image
import numpy as np
model.eval()
for x, _ in test_loader:
x = Variable(x[:min(n_images, x.size(0))].cuda(), volatile=True)
_, x_recon = model(x)
data = np.concatenate([x.data.cpu(), x_recon.data.cpu()])
print(data.shape)
img = combine_images(np.transpose(data, [0, 2, 3, 1]))
print((img.shape))
image = img * 255
print((image.shape))
Image.fromarray(image.astype(np.uint8)).save(args.save_dir + "/real_and_recon.png")
print()
print('Reconstructed images are saved to %s/real_and_recon.png' % args.save_dir)
print('-' * 70)
plt.imshow(plt.imread(args.save_dir + "/real_and_recon.png", ))
plt.show()
# print('hello')
break
digits = np.where(y_test == 1)[1]
for i, num in enumerate(digits):
num = int(num)
if flags[num]:
continue
else:
flags[num] = 1
index[num] = i
if np.all(flags):
break
x_deform_test = np.array([affine(x) for x in x_test])
print(index)
print(x_test[index].shape)
input_img = np.concatenate([x_test[index], x_deform_test[index]])
input_img = combine_images(input_img, height=2, width=10)
input_img = input_img * 255
Image.fromarray(input_img.astype(np.uint8)).save(args.save_dir +
'/input.png')
model.load_weights(args.weights1)
_, x_recon = eval_model.predict(x_test, batch_size=100)
_, x_deform_recon = eval_model.predict(x_deform_test, batch_size=100)
recon_img = np.concatenate([x_recon[index], x_deform_recon[index]])
recon_img = combine_images(recon_img, height=2, width=10)
recon_img = recon_img * 255
Image.fromarray(recon_img.astype(np.uint8)).save(args.save_dir +
'/recon.png')
model.load_weights(args.weights2)
_, x_l1_recon = eval_model.predict(x_test, batch_size=100)
def train(config, netG, netD, solver_netG, solver_netD, train_iterator,
monitor):
if config["train"][
"feature_loss"] and config["train"]["feature_loss"]["lambda"] > 0:
print(
f'Applying VGG feature Loss, weight: {config["train"]["feature_loss"]["lambda"]}.'
)
with_feature_loss = True
else:
with_feature_loss = False
# Load image and boundary image to get Variable shapes
img, bod_map, bod_map_resize = train_iterator.next()
real_img = nn.Variable(img.shape)
real_bod_map = nn.Variable(bod_map.shape)
real_bod_map_resize = nn.Variable(bod_map_resize.shape)
################### Graph Construction ####################
# Generator
with nn.parameter_scope('netG_decoder'):
fake_img = netG(real_bod_map, test=False)
fake_img.persistent = True
fake_img_unlinked = fake_img.get_unlinked_variable()
# Discriminator
with nn.parameter_scope('netD_decoder'):
pred_fake = netD(F.concatenate(real_bod_map_resize,
fake_img_unlinked,
axis=1),
test=False)
pred_real = netD(F.concatenate(real_bod_map_resize, real_img, axis=1),
test=False)
real_target = F.constant(1, pred_fake.shape)
fake_target = F.constant(0, pred_real.shape)
################### Loss Definition ####################
# for Generator
gan_loss_G = gan_loss(pred_fake, real_target)
gan_loss_G.persistent = True
weight_L1 = config["train"]["weight_L1"]
L1_loss = recon_loss(fake_img_unlinked, real_img)
L1_loss.persistent = True
loss_netG = gan_loss_G + weight_L1 * L1_loss
if with_feature_loss:
feature_loss = vgg16_perceptual_loss(127.5 * (fake_img_unlinked + 1.),
127.5 * (real_img + 1.))
feature_loss.persistent = True
loss_netG += feature_loss * config["train"]["feature_loss"]["lambda"]
# for Discriminator
loss_netD = (gan_loss(pred_real, real_target) +
gan_loss(pred_fake, fake_target)) * 0.5
################### Setting Solvers ####################
# for Generator
with nn.parameter_scope('netG_decoder'):
solver_netG.set_parameters(nn.get_parameters())
# for Discrimintar
with nn.parameter_scope('netD_decoder'):
solver_netD.set_parameters(nn.get_parameters())
################### Create Monitors ####################
interval = config["monitor"]["interval"]
monitors_G_dict = {
'loss_netG': loss_netG,
'loss_gan': gan_loss_G,
'L1_loss': L1_loss
}
if with_feature_loss:
monitors_G_dict.update({'vgg_feature_loss': feature_loss})
monitors_G = MonitorManager(monitors_G_dict, monitor, interval=interval)
monitors_D_dict = {'loss_netD': loss_netD}
monitors_D = MonitorManager(monitors_D_dict, monitor, interval=interval)
monitor_time = nm.MonitorTimeElapsed('time_training',
monitor,
interval=interval)
monitor_vis = nm.MonitorImage('result',
monitor,
interval=1,
num_images=4,
normalize_method=lambda x: x)
# Dump training information
with open(os.path.join(monitor._save_path, "training_info.yaml"),
"w",
encoding="utf-8") as f:
f.write(yaml.dump(config))
# Training
epoch = config["train"]["epochs"]
i = 0
lr_decay_start_at = config["train"]["lr_decay_start_at"]
iter_per_epoch = train_iterator.size // config["train"]["batch_size"] + 1
for e in range(epoch):
logger.info(f'Epoch = {e} / {epoch}')
train_iterator._reset() # rewind the iterator
if e > lr_decay_start_at:
decay_coeff = 1.0 - max(0, e - lr_decay_start_at) / 50.
lr_decayed = config["train"]["lr"] * decay_coeff
print(f"learning rate decayed to {lr_decayed}")
solver_netG.set_learning_rate(lr_decayed)
solver_netD.set_learning_rate(lr_decayed)
for _ in range(iter_per_epoch):
img, bod_map, bod_map_resize = train_iterator.next()
# bod_map_noize = np.random.random_sample(bod_map.shape) * 0.01
# bod_map_resize_noize = np.random.random_sample(bod_map_resize.shape) * 0.01
real_img.d = img
real_bod_map.d = bod_map # + bod_map_noize
real_bod_map_resize.d = bod_map_resize # + bod_map_resize_noize
# Generate fake image
fake_img.forward(clear_no_need_grad=True)
# Update Discriminator
solver_netD.zero_grad()
solver_netG.zero_grad()
loss_netD.forward(clear_no_need_grad=True)
loss_netD.backward(clear_buffer=True)
solver_netD.update()
# Update Generator
solver_netD.zero_grad()
solver_netG.zero_grad()
fake_img_unlinked.grad.zero()
loss_netG.forward(clear_no_need_grad=True)
loss_netG.backward(clear_buffer=True)
fake_img.backward(grad=None)
solver_netG.update()
# Monitors
monitor_time.add(i)
monitors_G.add(i)
monitors_D.add(i)
i += 1
images_to_visualize = [real_bod_map_resize.d, fake_img.d, img]
visuals = combine_images(images_to_visualize)
monitor_vis.add(i, visuals)
if e % config["monitor"]["save_interval"] == 0 or e == epoch - 1:
# Save parameters of networks
netG_save_path = os.path.join(monitor._save_path,
f'netG_decoder_{e}.h5')
with nn.parameter_scope('netG_decoder'):
nn.save_parameters(netG_save_path)
netD_save_path = os.path.join(monitor._save_path,
f'netD_decoder_{e}.h5')
with nn.parameter_scope('netD_decoder'):
nn.save_parameters(netD_save_path)
face_bbox = extend_bbox_by_landmarks(detected_face, facial_landmarks)
face_attribute_mask = get_mask(face_attribute, facial_landmarks, img.shape)
face_bbox = extend_bbox(face_bbox, img.shape, 0.25, 0.2)
cropped_by_face_img = F.crop(PIL.Image.fromarray(
img), face_bbox[1], face_bbox[0], face_bbox[3] - face_bbox[1], face_bbox[2] - face_bbox[0])
with open(cfg['vae_cfg_path'], 'r') as f:
vae_cfg = yaml.safe_load(f)
model = VanillaVAE(**vae_cfg['model_params'])
state_dict = torch.load(cfg['vae_model_path'], map_location=torch.device('cpu'))['state_dict']
model.load_state_dict(state_dict)
model.eval()
inference_transform = get_preinference_transforms(cfg)
restored, _, mu, _ = model.forward(inference_transform(cropped_by_face_img).unsqueeze(0))
scale = cfg['face_attribute_change_scale']
generated_img = restore_normalized_img(model.decode(mu - scale * vae_latent_vector).detach()[0])
generated_img = F.resize(PIL.Image.fromarray(generated_img), cropped_by_face_img.size[::-1])
generated_img = np.array(generated_img)
modified_image = combine_images(img, generated_img, face_bbox, face_attribute_mask)
idx = img_path.index('.')
new_img_path = img_path[:idx] + '_' + face_attribute + '_scale_' + str(scale) + img_path[idx:]
cv2.imwrite(new_img_path, cv2.cvtColor(modified_image, cv2.COLOR_RGB2BGR))
from utils import combine_images, get_random_cards cards = get_random_cards() combine_images(cards)
noise = np.zeros([args.batch_size, args.noise_dim])
dimension = range(0, args.noise_dim, int(args.noise_dim / 10))
x_recons = []
for dim in list(dimension):
for r in [-1, -0.8, -0.6, -0.4, -0.2, 0, 0.2, 0.4, 0.6, 0.8, 1]:
tmp = np.copy(noise)
tmp[:, dim] = r
x_recon = sess.run(
[model.fake_img],
feed_dict={
model.noise_holder: tmp,
model.feat_holder: feat,
model.isTrain: False
})
x_recon = np.array(x_recon)
x_recon = np.reshape(x_recon, [
args.batch_size, args.img_width, args.img_height,
args.channel
])
x_recons.append(x_recon[0, :, :, :])
x_recons = np.array(x_recons)
x_recons = np.reshape(
x_recons,
[x_recons.shape[0], args.img_width, args.img_height, args.channel])
img = combine_images(x_recons, height=10)
image = (img / 2. + 0.5) * 255
Image.fromarray(image.astype(
np.uint8)).save(args.save_manipulate_dir +
'/manipulate-%d.png' % args.digit)