def predict(conv=args.conv):
assert os.path.isdir(
constants.CHECKPOINT_DIR), 'Error: Model is not availabel'
if conv:
checkpoint = torch.load(
os.path.join(constants.CHECKPOINT_DIR, 'convergence.t7'))
model.load_state_dict(checkpoint['model'])
else:
checkpoint = torch.load(
os.path.join(constants.CHECKPOINT_DIR, 'best_acc_model.t7'))
model.load_state_dict(checkpoint['model'])
torch.set_grad_enabled(False)
net.eval()
test_correct = 0
for batch_id, (images, labels) in enumerate(test_loader):
images, labels = images.to(device), labels.to(device)
outputs = models(images)
_, predicted = outputs.max(1)
test_correct += predicted.eq(labels).sum().item()
print('Accuracy on test data: {}%'.format((test_correct / total) * 100))
x: 网络输出
target表示label（M，）
"""
smoothed_label = x.data.clone()
num_classes = x.size(1)
smoothed_label.fill_(self.smoothing / (num_classes - 1)) # otherwise的公式
#target.data.unsqueeze(1)表示索引,confidence表示填充的数字
smoothed_label.scatter_(1, target.data.unsqueeze(1), self.confidence)
return smoothed_label
if __name__=="__main__":
# Example of label smoothing.
import sys
sys.path.append('/home/gfx/Projects/Tinymind')
import torchvision.models as models
from networks.network import *
from config import config
backbone = models.resnet18(pretrained=True)
models = ResNet18(backbone, 100)
# print models
data = torch.randn(8, 1, 128, 128)
x = models(data)
print(x)
labelsmooth = LabelSmoothing(smoothing= 0.1)
smoothed_label = labelsmooth(x, Variable(torch.LongTensor([2, 0, 10, 90, 1, 2, 4, 6])))
print smoothed_label
spoof_data_list.append(temp_dic)
face_model = None
if run_parameters['model'] == 'vggface':
weight_pth = os.path.join('models', args.exp_name,
before_model_name + '_face_model.pth')
print(weight_pth)
vgg_face = vgg_face_dag(weights_path=weight_pth,
return_layer=run_parameters['return_layer'])
for p in vgg_face.parameters():
p.requires_grad = False
face_model = vgg_face
if run_parameters['model'] == 'vgg16':
vgg16 = torchvision.models().vgg16(pretrained=True)
face_model = vgg16
face_model.eval()
if run_parameters['multi_gpu']:
face_model = nn.DataParallel(face_model)
face_model.cuda()
inm = nn.InstanceNorm1d(1, affine=False)
print("No of training samples: {0}, {1}".format(len(live_data_list),
len(spoof_data_list)))
# exit()
data_cell = {}
models.train()
ACC = 0
batchHelper_train.resetIndex()
while batchHelper_train._epochs_completed == 0:
input_image, labelImage = batchHelper_train.next_batch(
batch_size, True)
# for i in range(int(totalSample/batch_size)):
loss = 0
optimizer.zero_grad()
inputImageDataset = torch.from_numpy(input_image)
inputImageDataset = inputImageDataset.to(device=device,
dtype=torch.float)
# print(torch.sum(inputImageDataset))
target_output = torch.from_numpy(labelImage.astype(int))
target_output = target_output.to(device=device, dtype=torch.long)
output_pred = models(inputImageDataset)
loss = criterion(output_pred, target_output)
loss.backward()
optimizer.step()
topv, topi = output_pred.topk(1)
check_target = (topi.reshape(-1) == target_output)
ACC += check_target.float().sum()
if batchHelper_train._index_in_epoch % batch_size == 0:
print('Train Epoch: {} [{}/{} ({:.0f}%)]\tLoss: {:.9f} current ACC {} '\
.format(y, batchHelper_train._index_in_epoch, totalSample_train,\
(batchHelper_train._index_in_epoch * 100.) / totalSample_train, \
loss.item(),(check_target.float().sum()/batch_size)*100))
############# validation #####################
batchHelper_val.resetIndex()
def test_visualisation(model_path, models, test_images_paths, test_dataset, mode ,device):
'''
Test and save sample images with bounding boxes and the labels
Inputs:
- model_path : string( path to the model )
- models : weights (model weights)
- test_images_paths : string ( model path )
- test_dataset : Dataset (testing dataset)
- device : device (cuda)
Outputs:
saved images
'''
import random
label_dict= { 1: 'Car', 2 : 'Van' , 3 : 'Truck', 0 : 'Background'}
directory = os.path.join('results', 'sample_bbox')
def save_image(image,directory, filename):
if not os.path.exists(directory):
os.makedirs(directory)
img_file = os.path.join(directory, filename)
image.save(img_file)
print(f'Loading model to score images. Scores saved {model_path}') # Testing the model
mean_test_accuracy = []
model_file = torch.load(model_path)
models.load_state_dict(model_file)
print('Model loaded')
models.to(device)
models.eval()
with torch.no_grad():
for path in test_images_paths:
with tqdm(total = len(test_dataset)) as bar:
for image in test_dataset.images:
img = os.path.join(path, image)
l_img, pic_image = load_image(img)
output = models(l_img)
im_show = l_img.permute(2,0,3,1)
im_show = im_show.squeeze(1)
labels = output[0]['labels']
scores = output[0]['scores']
mean_score = scores.mean()
if torch.isnan(mean_score).any():
continue
# print(f'This is the mean score : {mean_score}')
rect = output[0]['boxes']
if rect.nelement() != 0:
i = 0
int_rects = rect.int().cpu().numpy()
labels = labels.int().cpu().numpy()
scores = scores.float().cpu().numpy()
for int_rect, label, score in zip(int_rects, labels, scores):
# print(label_dict[label], score)
if score >= 0.5:
r = random.randint(20,255)
g = random.randint(20,255)
b = random.randint(20,255)
rgb = (r,g,b)
x0,y0 ,x1,y1 = int_rect
img1 = ImageDraw.Draw( pic_image )
font = ImageFont.truetype("bevan.ttf", 20)
# img1.text([x0,y0,x1,y1+10], label, fill=(255,255,0))
img1.text((0,0+i),f'{label_dict[label]} {score} ', rgb,font=font)
img1.rectangle([x0,y0 ,x1,y1], outline = rgb, width = 3) # Draw the text on the
i += 20
else:
continue
save_image(pic_image, os.path.join(directory, str(mode)), f'{image[:-4]}_samplebbox.png')
mean_score = mean_score.float().cpu().numpy()
mean_test_accuracy.append(mean_score)
bar.update()
print('FINISHED TESTING')
def train(models, writer, device):
vgg19 = torchvision.models.vgg19(pretrained=True)
for param in vgg19.parameters():
param.required_grad = False
vgg19.to(device)
real_labels = torch.ones((config.batch_size, 1), device=device)
fake_labels = torch.zeros((config.batch_size, 1), device=device)
for idx in range(config.load_iter, config.train_iters):
train_phone, train_dslr = load_train_data(
config.dataset_dir, config.phone, config.batch_size,
(config.channels, config.height, config.width))
train_phone = torch.as_tensor(train_phone, device=device)
train_dslr = torch.as_tensor(train_dslr, device=device)
# --------------------------------------------------------------------------------------------------------------
# Train discriminators
# --------------------------------------------------------------------------------------------------------------
enhanced, phone_rec = models(train_phone)
# 1) Adversarial color loss
dslr_blur = gaussian_blur(train_dslr, config.kernel_size, config.sigma,
config.channels, device)
dslr_blur_logits = models.dis_c(dslr_blur)
enhanced_blur = gaussian_blur(enhanced.detach(), config.kernel_size,
config.sigma, config.channels, device)
enhanced_blur_logits = models.dis_c(enhanced_blur)
dis_loss_color = models.bce_criterion(dslr_blur_logits, real_labels) \
+ models.bce_criterion(enhanced_blur_logits, fake_labels)
# 2) Adversarial texture loss
dslr_gray = rgb_to_gray(train_dslr, device)
dslr_gray_logits = models.dis_t(dslr_gray)
enhanced_gray = rgb_to_gray(enhanced.detach(), device)
enhanced_gray_logits = models.dis_t(enhanced_gray)
dis_loss_texture = models.bce_criterion(dslr_gray_logits, real_labels) \
+ models.bce_criterion(enhanced_gray_logits, fake_labels)
# Sum of losses
dis_loss = dis_loss_color + dis_loss_texture
models.dis_optimizer.zero_grad()
dis_loss.backward()
models.dis_optimizer.step()
# --------------------------------------------------------------------------------------------------------------
# Train generators
# --------------------------------------------------------------------------------------------------------------
# 1) Content consistency loss
phone_vgg = get_content(vgg19, train_phone, config.content_id, device)
phone_rec_vgg = get_content(vgg19, phone_rec, config.content_id,
device)
gen_loss_content = models.mse_criterion(phone_vgg, phone_rec_vgg)
# 2) Adversarial color loss
enhanced_blur = gaussian_blur(enhanced, config.kernel_size,
config.sigma, config.channels, device)
enhanced_blur_logits = models.dis_c(enhanced_blur)
gen_loss_color = models.bce_criterion(enhanced_blur_logits,
real_labels)
# 3) Adversarial texture loss
enhanced_gray = rgb_to_gray(enhanced, device)
enhanced_gray_logits = models.dis_t(enhanced_gray)
gen_loss_texture = models.bce_criterion(enhanced_gray_logits,
real_labels)
# 4) TV loss
y_tv = models.mse_criterion(enhanced[:, :, 1:, :],
enhanced[:, :, :-1, :])
x_tv = models.mse_criterion(enhanced[:, :, :, 1:],
enhanced[:, :, :, :-1])
gen_loss_tv = y_tv + x_tv
# Sum of losses
gen_loss = config.w_content * gen_loss_content + config.w_color * gen_loss_color \
+ config.w_texture * gen_loss_texture + config.w_tv * gen_loss_tv
models.gen_optimizer.zero_grad()
gen_loss.backward()
models.gen_optimizer.step()
if (idx + 1) % config.print_step == 0:
print(
"Iteration: {}/{}, gen_loss: {:.4f}, dis_loss: {:.4f}".format(
idx + 1, config.train_iters, gen_loss.item(),
dis_loss.item()))
print(
"gen_loss_content: {:.4f}, gen_loss_color: {:.4f}, gen_loss_texture: {:.4f}, gen_loss_tv: {:.4f}"
.format(gen_loss_content.item(), gen_loss_color.item(),
gen_loss_texture.item(), gen_loss_tv.item()))
print("dis_loss_color: {:.4f}, dis_loss_texture: {:.4f}".format(
dis_loss_color.item(), dis_loss_texture.item()))
if (idx + 1) % config.tensorboard_step == 0:
scalar_dict = {
'gen_loss': gen_loss.item(),
'dis_loss': dis_loss.item(),
'gen_loss_content': gen_loss_content.item(),
'gen_loss_color': gen_loss_color.item(),
'gen_loss_texture': gen_loss_texture.item(),
'gen_loss_tv': gen_loss_tv.item(),
'dis_loss_color': dis_loss_color.item(),
'dis_loss_texture': dis_loss_texture.item()
}
img_dict = {
'phone': train_phone,
'dslr': train_dslr,
'enhanced': enhanced,
'phone_rec': phone_rec,
'dslr_blur': dslr_blur,
'enhanced_blur': enhanced_blur,
'dslr_gray': dslr_gray,
'enhanced_gray': enhanced_gray
}
for tag, scalar_value in scalar_dict.items():
writer.add_scalar(tag, scalar_value, idx + 1)
for tag, img_tensor in img_dict.items():
writer.add_images(tag, img_tensor, idx + 1)
if (idx + 1) % config.checkpoint_step == 0:
checkpoint_path = os.path.join(config.checkpoint_dir, config.phone)
torch.save(
models.gen_g.state_dict(),
os.path.join(checkpoint_path,
'{:05d}-gen_g.ckpt'.format(idx + 1)))
torch.save(
models.gen_f.state_dict(),
os.path.join(checkpoint_path,
'{:05d}-gen_f.ckpt'.format(idx + 1)))
torch.save(
models.dis_c.state_dict(),
os.path.join(checkpoint_path,
'{:05d}-dis_c.ckpt'.format(idx + 1)))
torch.save(
models.dis_t.state_dict(),
os.path.join(checkpoint_path,
'{:05d}-dis_t.ckpt'.format(idx + 1)))
