def load_data(batch_size, training_set_feat, training_set_labels,
validation_set_feat, validation_set_labels):
train_dataset = ImgDataset(training_set_feat, training_set_labels)
validation_dataset = ImgDataset(validation_set_feat, validation_set_labels)
train_loader = torch.utils.data.DataLoader(train_dataset,
batch_size=batch_size,
shuffle=True)
val_loader = torch.utils.data.DataLoader(validation_dataset,
batch_size=batch_size,
shuffle=True)
return train_loader, val_loader
def load_dataset(args):
img_transform = transforms.Compose([
transforms.Resize(256),
transforms.RandomCrop(224),
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
])
train_caption, validation_caption, test_caption, train_id, validation_id, val_img_id, test_img_id, img_idx, vocab_table = read_text_data(
args)
train_dataset = TrainDataset(args.img_dir, vocab_table, img_idx,
train_caption, train_id, img_transform)
val_img_dataset = ImgDataset(args.img_dir,
val_img_id,
is_val=True,
transform=img_transform,
img_idx=img_idx)
test_img_dataset = ImgDataset(args.img_dir,
test_img_id,
transform=img_transform)
val_caption_dataset = CaptionDataset(vocab_table,
validation_caption,
is_val=True,
label=validation_id)
test_caption_dataset = CaptionDataset(vocab_table, test_caption)
train_loader = DataLoader(train_dataset,
batch_size=args.batch_size,
num_workers=4,
shuffle=True,
collate_fn=train_collate_fn)
val_img_loader = DataLoader(val_img_dataset,
batch_size=args.batch_size,
num_workers=4)
test_img_loader = DataLoader(test_img_dataset,
batch_size=args.batch_size,
num_workers=4)
val_caption_loader = DataLoader(val_caption_dataset,
batch_size=args.batch_size,
num_workers=4,
collate_fn=test_collate_fn)
test_caption_loader = DataLoader(test_caption_dataset,
batch_size=args.batch_size,
num_workers=4,
collate_fn=test_collate_fn)
return train_loader, val_img_loader, test_img_loader, val_caption_loader, test_caption_loader, vocab_table
def train_val_data(dir_path1, dir_path2):
x1, y1 = read_imgs(dir_path1, True)
x2, y2 = read_imgs(dir_path2, True)
x = np.concatenate((x1, x2), axis=0)
y = np.concatenate((y1, y2), axis=0)
return ImgDataset(x, y, train_transform)
def main(args):
segmen_A = Unet(3, 34).to(args.device)
if args.model_path is not None:
segmen_path = os.path.join(args.model_path, 'semsg.pt')
with open(segmen_path, 'rb') as f:
state_dict = torch.load(f)
segmen_A.load_state_dict(state_dict)
else:
raise Exception('please specify model path!')
segmen_A = nn.DataParallel(segmen_A)
transforms_ = [
transforms.ToTensor(),
transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
]
testloader = DataLoader(ImgDataset(args.image_path,
transforms_=transforms_,
mode='eval_unet'),
batch_size=args.batchSize,
shuffle=False,
num_workers=0)
segmen_A.eval()
with torch.no_grad():
total_iou = []
for i, batch in enumerate(testloader):
name, toTest, labels = batch
#segmentation
pred_label = segmen_A(toTest)
for idx in range(args.batchSize):
pred = pred_label[idx].cpu().numpy()
label = labels.cpu().numpy()[idx]
img = np.zeros(
(label.shape[0], label.shape[1], 3)).astype('uint8')
original_img = np.zeros(
(label.shape[0], label.shape[1], 3)).astype('uint8')
prediction = np.zeros(
(label.shape[0], label.shape[1])).astype('uint8')
for c in range(len(palette)):
indices = np.argmax(pred, axis=0) == c
prediction[indices] = c
img[indices] = palette[c]
original_img[label == c] = palette[c]
original_img = Image.fromarray(original_img.astype('uint8'))
original_img.save(
os.path.join(args.out_dir, 'original_' +
name[idx].replace('jpg', 'png')))
img = Image.fromarray(img.astype('uint8'))
img.save(
os.path.join(args.out_dir, name[idx].replace('jpg',
'png')))
total_iou.append(IOU(prediction, label, 34))
print(sum(total_iou) / len(total_iou))
def main(args):
G = Generator(args.in_channel, args.out_channel).to(args.device)
G_reverse = Generator(args.in_channel, args.out_channel).to(args.device)
if args.model_path is not None:
AB_path = os.path.join(args.model_path,'ab.pt')
BA_path = os.path.join(args.model_path,'ba.pt')
if args.direction == 'AB':
with open(AB_path, 'rb') as f:
state_dict = torch.load(f)
G.load_state_dict(state_dict)
with open(BA_path, 'rb') as f:
state_dict = torch.load(f)
G_reverse.load_state_dict(state_dict)
elif args.direction == 'BA':
with open(BA_path, 'rb') as f:
state_dict = torch.load(f)
G.load_state_dict(state_dict)
with open(AB_path, 'rb') as f:
state_dict = torch.load(f)
G_reverse.load_state_dict(state_dict)
else:
raise Exception('direction has to be BA OR AB!')
else:
raise Exception('please specify model path!')
G = nn.DataParallel(G)
transforms_ = [ transforms.ToTensor(),
transforms.Normalize((0.5,0.5,0.5), (0.5,0.5,0.5)) ]
testloader = DataLoader(ImgDataset(args.image_path, transforms_=transforms_, mode='test'),
batch_size=args.batchSize, shuffle=False, num_workers=0)
G.eval()
with torch.no_grad():
for i, batch in enumerate(testloader):
name, toTest = batch
transformed_ = G(toTest)
# recovered = G_reverse(transformed_)
for idx in range(len(name)):
# utils.save_image(torch.cat((toTest[idx].to(args.device), recovered[idx], transformed_[idx]),axis=1), os.path.join(args.out_dir, args.direction+'_'+name[idx].split('/')[-1]), normalize=True, range=(-1, 1))
utils.save_image(transformed_[idx], os.path.join(args.out_dir, name[idx].split('/')[-1]), normalize=True, range=(-1, 1))
def load_dataset():
transform = transforms.Compose([
transforms.Resize((args.input_size, args.input_size)),
transforms.ToTensor(),
transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
# transforms.Normalize((0.5,), (0.5,)) # convert (0,1) to (-1,1)
])
# train_set = datasets.MNIST(
# args.dataset_path, train=True, download=True, transform=transform)
# train_set = datasets.CIFAR10(
# args.dataset_path, train=True, download=True, transform=transform)
class_label_dct = {
'冬瓜排骨汤': 0,
'土豆丝': 1,
'椒盐虾': 2,
'番茄炒蛋': 3,
'糖醋里脊': 4,
'红烧肉': 5,
'莴笋肉片': 6,
'辣子鸡': 7,
'香菇青菜': 8,
'鱼香茄子': 9
}
train_set = ImgDataset(root=args.dataset_path,
type_='train',
transforms=transform,
class_label_dct=class_label_dct,
num_per_class=200)
train_loader = DataLoader(train_set,
batch_size=args.batch_size,
shuffle=True)
return train_loader
import torch
import matplotlib.pyplot as plt
from dataset import ImgDataset
from model import Model
"""Plot Saliency Map"""
cuda = False
img_indices = [83, 4218, 4707, 8598] # 选择数据集中的几张图片
# 加载模型
if cuda:
model = torch.load('../hw3_CNN/model.pth')
else:
model = torch.load('../hw3_CNN/model.pth', map_location='cpu')
# 选择数据集
training_dataset = ImgDataset('../hw3_CNN/data/training', cuda)
images, labels = training_dataset.get_batch(img_indices)
# 计算Loss并求导
model.eval()
images.requires_grad_() # 追踪loss关于images的梯度
y_pred = model(images)
loss_func = torch.nn.CrossEntropyLoss()
loss = loss_func(y_pred, labels)
loss.backward()
# 取出loss关于images的梯度
if cuda:
saliencies = images.grad.abs().detach().cpu()
else:
saliencies = images.grad.abs().detach()
saliencies = torch.stack([(item - item.min()) / (item.max() - item.min())
transforms.ToTensor(
), #將圖片轉成 Tensor，並把數值normalize到[0,1](data normalization)
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])
#testing 時不需做 data augmentation
test_transform = transforms.Compose([
transforms.ToPILImage(),
transforms.CenterCrop(224),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])
# Parameters initialize
batch_size = 32
train_set = ImgDataset(train_X, train_Y, train_transform)
val_set = ImgDataset(val_X, val_Y, test_transform)
test_set = ImgDataset(test_x, transform=train_transform)
train_loader = DataLoader(train_set, batch_size=batch_size, shuffle=True)
val_loader = DataLoader(val_set, batch_size=batch_size, shuffle=False)
test_loader = DataLoader(test_set, batch_size=batch_size, shuffle=True)
teacher_net = []
teacher_net1 = VGG16().cuda()
teacher_net1.load_state_dict(torch.load('model/vgg16.model'))
teacher_net2 = VGG13().cuda()
teacher_net2.load_state_dict(torch.load('model/vgg13.model'))
# teacher_net3 = VGG19().cuda()
# teacher_net3.load_state_dict(torch.load('teacher_model/vgg19.model'))
teacher_net.append(teacher_net1)
teacher_net.append(teacher_net2)
from dataset import ImgDataset
from gan import GAN
import os.path as op
path = op.join('..', 'datasets', 'small')
data = ImgDataset(path)
gan = GAN(data)
# reading testing set
print("Reading data")
test_x = readfile(os.path.join(workspace_dir, "testing"), False)
print("Size of Testing data = {}".format(len(test_x)))
# testing configuration
model_best = Classifier().cuda()
model_best.load_state_dict(torch.load(model_filename))
loss = nn.CrossEntropyLoss()
optimizer = torch.optim.Adam(model_best.parameters(), lr=0.001)
# testing dataset
batch_size = 48
test_set = ImgDataset(test_x, transform=test_transform)
test_loader = DataLoader(test_set, batch_size=batch_size, shuffle=False)
# predict
model_best.eval()
prediction = []
with torch.no_grad():
for i, data in enumerate(test_loader):
test_pred = model_best(data.cuda())
test_label = np.argmax(test_pred.cpu().data.numpy(), axis=1)
for y in test_label:
prediction.append(y)
# Write the result to csv file
with open(output_filename, 'w') as f:
f.write('Id,Category\n')
# reading training set, validation set
workspace_dir = sys.argv[1] #'/home/shannon/Downloads/food-11'
print("Reading data")
train_x, train_y = readfile(os.path.join(workspace_dir, "training"), True)
print("Size of training data = {}".format(len(train_x)))
val_x, val_y = readfile(os.path.join(workspace_dir, "validation"), True)
print("Size of validation data = {}".format(len(val_x)))
# train in training & validating set (final step)
#train_x = np.concatenate((train_x, val_x), axis=0)
#train_y = np.concatenate((train_y, val_y), axis=0)
# create train and valid dataset
batch_size = 48
train_set = ImgDataset(train_x, train_y, train_transform)
val_set = ImgDataset(val_x, val_y, test_transform)
train_loader = DataLoader(train_set, batch_size=batch_size, shuffle=True)
val_loader = DataLoader(val_set, batch_size=batch_size, shuffle=False)
# Keep the loss and accuracy at every iteration for plotting
train_loss_list = []
dev_loss_list = []
train_acc_list = []
dev_acc_list = []
# training configuration
model = Classifier().cuda()
#model_filename = "./model/vgg16_lite_drop_bth48_lr0.002_ep200_deg60_img168_112/lr0001_train_n_val/model_0.9848396501457726_ep150"
#model.load_state_dict(torch.load(model_filename))
loss = nn.CrossEntropyLoss(
print("Reading data")
train_x, train_y = readfile(os.path.join(workspace_dir, "Train"), True)
print("Size of training data = {}".format(len(train_x)))
print("label Size of training data = {}".format(len(train_y)))
val_x, val_y = readfile(os.path.join(workspace_dir, "Dev"), True)
print("Size of validation data = {}".format(len(val_x)))
print(train_x.shape)
#ensure every time has same dataset
torch.manual_seed(567)
torch.cuda.manual_seed(567)
np.random.seed(567)
#data augmentation & load
batch_size = 64
train_set = ImgDataset(train_x, train_y, train_transform)
val_set = ImgDataset(val_x, val_y, test_transform)
train_loader = DataLoader(train_set, batch_size=batch_size, shuffle=True)
val_loader = DataLoader(val_set, batch_size=batch_size, shuffle=False)
model = Classifier().cuda()
loss = nn.CrossEntropyLoss(
) # 因為是 classification task，所以 loss 使用 CrossEntropyLoss
optimizer = torch.optim.Adam(model.parameters(),
lr=0.001) # optimizer 使用 Adam
#num_epoch = 60
num_epoch = 1
for epoch in range(num_epoch):
epoch_start_time = time.time()
train_acc = 0.0
train_loss = 0.0
print('[*] Loading pickles...')
with open(f'preprocessed/{dataset_type}train_x.pkl', 'rb') as f:
train_x = pickle.load(f)
with open(f'preprocessed/{dataset_type}train_y.pkl', 'rb') as f:
train_y = pickle.load(f)
with open(f'preprocessed/valid_x.pkl', 'rb') as f:
valid_x = pickle.load(f)
with open(f'preprocessed/valid_y.pkl', 'rb') as f:
valid_y = pickle.load(f)
with open(f'preprocessed/{dataset_type}train_val_x.pkl', 'rb') as f:
train_val_x = pickle.load(f)
with open(f'preprocessed/{dataset_type}train_val_y.pkl', 'rb') as f:
train_val_y = pickle.load(f)
train_dataset = ImgDataset(train_x, train_y, transform=train_transform)
valid_dataset = ImgDataset(valid_x, valid_y, transform=test_transform)
train_val_dataset = ImgDataset(train_val_x,
train_val_y,
transform=train_transform)
model = build_model()
device = torch.device(
f'cuda:{args.cuda}' if torch.cuda.is_available() else 'cpu')
if args.resume:
model.load_state_dict(torch.load(f'{arch}/model.ckpt'))
model = model.to(device)
if not device == 'cpu':
cudnn.benchmark = True
trainer = Trainer(arch, model, args.batch_size, args.lr, args.accum_steps,
def val_data(dir_path):
x, y = read_imgs(dir_path, True)
return ImgDataset(x, y, test_transform)
total += 1
with env.begin(write = True) as txn:
txn.put('length'.encode('utf-8'), str(total).encode('utf-8'))
if __name__ == "__main__":
parser = argparse.ArgumentParser(description = 'Preprocess images for model training')
parser.add_argument('--out', type = str, help = 'filename of the result lmdb dataset')
parser.add_argument('--size', type = str, default = '128,256,512,1024', \
help = 'resolutions of images for the dataset')
parser.add_argument('--n_worker', type = int, default = 8, \
help = 'number of workers for preparing dataset')
parser.add_argument('--resample', type = str, default = 'lanczos', \
help = 'resampling methods for resizing images')
parser.add_argument('path', type = str, help = 'path to the image dataset')
args = parser.parse_args()
sizes = []
for s in args.size.split(','):
try:
size = int(s.strip())
sizes += [size]
except ValueError as s:
pass
print('Make dataset of image sizes:' + ','.join('%d'%s for s in sizes))
try:
from torchvision import datasets
imgset = datasets.ImageFolder(args.path)
except:
from dataset import ImgDataset
imgset = ImgDataset(args.path)
with lmdb.open(args.out, map_size = 1024 ** 4, readahead = False) as env:
prepare(env, imgset, args.n_worker, sizes = sizes, resample = args.resample)
def main(args):
writer = SummaryWriter(os.path.join(args.out_dir, 'logs'))
current_time = datetime.now().strftime("%d-%m-%Y_%H-%M-%S")
os.makedirs(
os.path.join(args.out_dir, 'models',
args.model_name + '_' + current_time))
os.makedirs(
os.path.join(args.out_dir, 'logs',
args.model_name + '_' + current_time))
G_AB = Generator(args.in_channel, args.out_channel).to(args.device)
G_BA = Generator(args.in_channel, args.out_channel).to(args.device)
D_A = Discriminator(args.in_channel).to(args.device)
D_B = Discriminator(args.out_channel).to(args.device)
segmen_B = Unet(3, 34).to(args.device)
if args.model_path is not None:
AB_path = os.join.path(args.model_path, 'ab.pt')
BA_path = os.join.path(args.model_path, 'ba.pt')
DA_path = os.join.path(args.model_path, 'da.pt')
DB_path = os.join.path(args.model_path, 'db.pt')
segmen_path = os.join.path(args.model_path, 'semsg.pt')
with open(AB_path, 'rb') as f:
state_dict = torch.load(f)
G_AB.load_state_dict(state_dict)
with open(BA_path, 'rb') as f:
state_dict = torch.load(f)
G_BA.load_state_dict(state_dict)
with open(DA_path, 'rb') as f:
state_dict = torch.load(f)
D_A.load_state_dict(state_dict)
with open(DB_path, 'rb') as f:
state_dict = torch.load(f)
D_B.load_state_dict(state_dict)
with open(segmen_path, 'rb') as f:
state_dict = torch.load(f)
segmen_B.load_state_dict(state_dict)
else:
G_AB.apply(weights_init_normal)
G_BA.apply(weights_init_normal)
D_A.apply(weights_init_normal)
D_B.apply(weights_init_normal)
G_AB = nn.DataParallel(G_AB)
G_BA = nn.DataParallel(G_BA)
D_A = nn.DataParallel(D_A)
D_B = nn.DataParallel(D_B)
segmen_B = nn.DataParallel(segmen_B)
criterion_GAN = torch.nn.MSELoss()
criterion_cycle = torch.nn.L1Loss()
criterion_identity = torch.nn.L1Loss()
criterion_segmen = torch.nn.BCELoss()
optimizer_G = torch.optim.Adam(itertools.chain(G_AB.parameters(),
G_BA.parameters()),
lr=args.lr,
betas=(0.5, 0.999))
optimizer_D_A = torch.optim.Adam(D_A.parameters(),
lr=args.lr,
betas=(0.5, 0.999))
optimizer_D_B = torch.optim.Adam(D_B.parameters(),
lr=args.lr,
betas=(0.5, 0.999))
optimizer_segmen_B = torch.optim.Adam(segmen_B.parameters(),
lr=args.lr,
betas=(0.5, 0.999))
lr_scheduler_G = torch.optim.lr_scheduler.LambdaLR(
optimizer_G,
lr_lambda=LambdaLR(args.n_epochs, args.epoch, args.decay_epoch).step)
lr_scheduler_D_A = torch.optim.lr_scheduler.LambdaLR(
optimizer_D_A,
lr_lambda=LambdaLR(args.n_epochs, args.epoch, args.decay_epoch).step)
lr_scheduler_D_B = torch.optim.lr_scheduler.LambdaLR(
optimizer_D_B,
lr_lambda=LambdaLR(args.n_epochs, args.epoch, args.decay_epoch).step)
fake_A_buffer = ReplayBuffer()
fake_B_buffer = ReplayBuffer()
transforms_ = [
transforms.ToTensor(),
transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
]
dataloader = DataLoader(ImgDataset(args.dataset_path,
transforms_=transforms_,
unaligned=True,
device=args.device),
batch_size=args.batchSize,
shuffle=True,
num_workers=0)
logger = Logger(args.n_epochs, len(dataloader))
target_real = Variable(torch.Tensor(args.batchSize,
1).fill_(1.)).to(args.device).detach()
target_fake = Variable(torch.Tensor(args.batchSize,
1).fill_(0.)).to(args.device).detach()
G_AB.train()
G_BA.train()
D_A.train()
D_B.train()
segmen_B.train()
for epoch in range(args.epoch, args.n_epochs):
for i, batch in enumerate(dataloader):
real_A = batch['A'].clone()
real_B = batch['B'].clone()
B_label = batch['B_label'].clone()
fake_b = G_AB(real_A)
fake_a = G_BA(real_B)
same_b = G_AB(real_B)
same_a = G_BA(real_A)
recovered_A = G_BA(fake_b)
recovered_B = G_AB(fake_a)
pred_Blabel = segmen_B(real_B)
pred_fakeAlabel = segmen_B(fake_a)
optimizer_segmen_B.zero_grad()
#segmen loss, do we assume that it also learns how to segment images after doing domain transfer?
loss_segmen_B = criterion_segmen(
pred_Blabel, B_label) + criterion_segmen(
segmen_B(fake_a.detach()), B_label)
loss_segmen_B.backward()
optimizer_segmen_B.step()
optimizer_G.zero_grad()
#gan loss
pred_fakeb = D_B(fake_b)
loss_gan_AB = criterion_GAN(pred_fakeb, target_real)
pred_fakea = D_A(fake_a)
loss_gan_BA = criterion_GAN(pred_fakea, target_real)
#identity loss
loss_identity_B = criterion_identity(same_b, real_B) * 5
loss_identity_A = criterion_identity(same_a, real_A) * 5
#cycle consistency loss
loss_cycle_ABA = criterion_cycle(recovered_A, real_A) * 10
loss_cycle_BAB = criterion_cycle(recovered_B, real_B) * 10
#cycle segmen diff loss
loss_segmen_diff = criterion_segmen(segmen_B(recovered_B),
pred_Blabel.detach())
loss_G = loss_gan_AB + loss_gan_BA + loss_identity_B + loss_identity_A + loss_cycle_ABA + loss_cycle_BAB + loss_segmen_diff
loss_G.backward()
optimizer_G.step()
##discriminator a
optimizer_D_A.zero_grad()
pred_realA = D_A(real_A)
loss_D_A_real = criterion_GAN(pred_realA, target_real)
fake_A = fake_A_buffer.push_and_pop(fake_a)
pred_fakeA = D_A(fake_A.detach())
loss_D_A_fake = criterion_GAN(pred_fakeA, target_fake)
loss_D_A = (loss_D_A_real + loss_D_A_fake) * 0.5
loss_D_A.backward()
optimizer_D_A.step()
#discriminator b
optimizer_D_B.zero_grad()
pred_realB = D_B(real_B)
loss_D_B_real = criterion_GAN(pred_realB, target_real)
fake_B = fake_B_buffer.push_and_pop(fake_b)
pred_fakeB = D_B(fake_B.detach())
loss_D_B_fake = criterion_GAN(pred_fakeB, target_fake)
loss_D_B = (loss_D_B_real + loss_D_B_fake) * 0.5
loss_D_B.backward()
optimizer_D_B.step()
logger.log(
{
'loss_segmen_B': loss_segmen_B,
'loss_G': loss_G,
'loss_G_identity': (loss_identity_A + loss_identity_B),
'loss_G_GAN': (loss_gan_AB + loss_gan_BA),
'loss_G_cycle': (loss_cycle_ABA + loss_cycle_BAB),
'loss_D': (loss_D_A + loss_D_B)
},
images={
'real_A': real_A,
'real_B': real_B,
'fake_A': fake_a,
'fake_B': fake_b,
'reconstructed_A': recovered_A,
'reconstructed_B': recovered_B
},
out_dir=os.path.join(
args.out_dir, 'logs',
args.model_name + '_' + current_time + '/' + str(epoch)),
writer=writer)
if (epoch + 1) % args.save_per_epochs == 0:
os.makedirs(
os.path.join(args.out_dir, 'models',
args.model_name + '_' + current_time, str(epoch)))
torch.save(
G_AB.module.state_dict(),
os.path.join(args.out_dir,
'models', args.model_name + '_' + current_time,
str(epoch), 'ab.pt'))
torch.save(
G_BA.module.state_dict(),
os.path.join(args.out_dir,
'models', args.model_name + '_' + current_time,
str(epoch), 'ba.pt'))
torch.save(
D_A.module.state_dict(),
os.path.join(args.out_dir,
'models', args.model_name + '_' + current_time,
str(epoch), 'da.pt'))
torch.save(
D_B.module.state_dict(),
os.path.join(args.out_dir,
'models', args.model_name + '_' + current_time,
str(epoch), 'db.pt'))
torch.save(
segmen_B.module.state_dict(),
os.path.join(args.out_dir,
'models', args.model_name + '_' + current_time,
str(epoch), 'semsg.pt'))
lr_scheduler_G.step()
lr_scheduler_D_A.step()
lr_scheduler_D_B.step()
# 轉灰階: 將輸入3維壓成1維。
transforms.Grayscale(),
# 縮放: 因為source data是32x32，我們將target data的28x28放大成32x32。
transforms.Resize((32, 32)),
# 水平翻轉 (Augmentation)
transforms.RandomHorizontalFlip(),
# 旋轉15度內 (Augmentation)，旋轉後空的地方補0
transforms.RandomRotation(15, fill=(0, )),
# 最後轉成Tensor供model使用。
transforms.ToTensor(),
])
#source_dataset = ImageFolder('real_or_drawing/train_data', transform=source_transform)
#target_dataset = ImageFolder('real_or_drawing/test_data', transform=target_transform)
source_dataset = ImgDataset(source_transform,
target_transform,
'./real_or_drawing',
train=True)
print('dataset done')
target_dataset = ImgDataset(source_transform,
target_transform,
'./real_or_drawing',
train=False)
print('dataset done 2')
source_dataloader = DataLoader(source_dataset, batch_size=32, shuffle=True)
target_dataloader = DataLoader(target_dataset, batch_size=32, shuffle=True)
test_dataloader = DataLoader(target_dataset, batch_size=128, shuffle=False)
feature_extractor = FeatureExtractor().cuda()
label_predictor = LabelPredictor().cuda()
domain_classifier = DomainClassifier().cuda()
y[i] = int(file.split("_")[0])
# if label=true: train&valid
if label:
return x, y
# if label=false: test
else:
return x
# reading testing set
print("Reading data")
val_x, val_y = readfile(os.path.join(workspace_dir, "validation"), True)
print("Size of Testing data = {}".format(len(val_x)))
# testing dataset
batch_size = 48
val_set = ImgDataset(val_x, val_y, test_transform)
val_loader = DataLoader(val_set, batch_size=batch_size, shuffle=False)
# testing configuration
model_best = Classifier().cuda()
model_best.load_state_dict(torch.load(model_filename))
# predict
model_best.eval()
val_y_hat = []
with torch.no_grad():
for i, data in enumerate(val_loader):
test_pred = model_best(data[0].cuda())
test_label = np.argmax(test_pred.cpu().data.numpy(), axis=1)
for y in test_label:
val_y_hat.append(y)
def train(df, img_dir, pretrained_file=None):
pdb.set_trace()
# set up dataset for training
df = df.sample(frac=1., random_state=42)
train_number = int(len(df) * 0.8)
train_df = df.iloc[:train_number, :]
valid_df = df.iloc[train_number:, :]
del df
cols = ["label_" + str(idx) for idx in range(1, 5)]
train_dataset = ImgDataset(train_df["imageId"].values,
img_dir,
mask_list=train_df[cols])
train_data_loader = DataLoader(
train_dataset,
batch_size=8,
shuffle=True,
)
valid_dataset = ImgDataset(valid_df["imageId"].values,
img_dir,
mask_list=train_df[cols])
valid_data_loader = DataLoader(
valid_dataset,
batch_size=4,
shuffle=True,
)
# set up model parameters
model = res34Unet(num_classes=4)
print(model)
pdb.set_trace()
if pretrained_file is not None:
skip = ['block.5.weight', 'block.5.bias']
load_pretained_weights(model,
pretained_file,
skip=skip,
first_layer=["block.0.0.weight"])
summary(model, torch.zeros(2, 1, 224, 224))
LR = 3e-4
optimizer = optim.Adam(model.parameters(), lr=LR)
scheduler = optim.lr_scheduler.ReduceLROnPlateau(optimizer,
mode="min",
patience=3,
verbose=True)
# set up train parameters
DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu")
SAVER = "./model_data/model.bin"
EPOCHS = 50
CNT = 0
BEST_VALID_LOSS = float("inf")
PATIENCE = 5
train_meter = Meter(phase="train")
valid_meter = Meter(phase="valid")
model.to(DEVICE)
for epoch in range(EPOCHS):
st_time = time.time()
train_loss = train_epoch(model,
train_data_loader,
optimizer,
DEVICE,
train_meter,
schedule=None)
current_time = time.time()
train_meter.epoch_log(epoch, train_loss, current_time - st_time)
valid_loss = valid_epoch(model, valid_data_loader, DEVICE, valid_meter)
valid_meter.epoch_log(epoch, valid_loss, time.time() - current_time)
scheduler.step(valid_loss)
if valid_loss < BEST_VALID_LOSS:
CNT = 0
BEST_VALID_LOSS = valid_loss
torch.save(model.state_dict(), SAVER)
else:
CNT += 1
if CNT >= PATIENCE:
print("Early stopping ... ")
break
def main(args):
current_time = datetime.now().strftime("%d-%m-%Y_%H-%M-%S")
os.makedirs(
os.path.join(args.out_dir, 'models',
args.model_name + '_' + current_time))
os.makedirs(
os.path.join(args.out_dir, 'logs',
args.model_name + '_' + current_time))
segmen_B = Unet(3, 34).to(args.device)
if args.model_path is not None:
segmen_path = os.join.path(args.model_path, 'semsg.pt')
with open(segmen_path, 'rb') as f:
state_dict = torch.load(f)
segmen_B.load_state_dict(state_dict)
segmen_B = nn.DataParallel(segmen_B)
criterion_segmen = torch.nn.BCELoss()
optimizer_segmen_B = torch.optim.Adam(segmen_B.parameters(),
lr=args.lr,
betas=(0.5, 0.999))
transforms_ = [
transforms.ToTensor(),
transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
]
dataloader = DataLoader(ImgDataset(args.dataset_path,
transforms_=transforms_,
mode='unetTrain',
unaligned=False,
device=args.device),
batch_size=args.batchSize,
shuffle=True,
num_workers=0)
logger = Logger(args.n_epochs, len(dataloader))
segmen_B.train()
for epoch in range(args.epoch, args.n_epochs):
for i, batch in enumerate(dataloader):
real_B = batch['B'].clone()
B_label = batch['B_label'].clone()
optimizer_segmen_B.zero_grad()
#segmen loss
pred_Blabel = segmen_B(real_B)
loss_segmen_B = criterion_segmen(pred_Blabel, B_label)
loss_segmen_B.backward()
optimizer_segmen_B.step()
logger.log({'loss_segmen': loss_segmen_B},
out_dir=os.path.join(
args.out_dir, 'logs', args.model_name + '_' +
current_time + '/' + str(epoch)))
if (epoch + 1) % args.save_per_epochs == 0:
os.makedirs(
os.path.join(args.out_dir, 'models',
args.model_name + '_' + current_time, str(epoch)))
torch.save(
segmen_B.module.state_dict(),
os.path.join(args.out_dir,
'models', args.model_name + '_' + current_time,
str(epoch), 'semsg.pt'))
def train_data(dir_path):
x, y = read_imgs(dir_path, True)
return ImgDataset(x, y, train_transform)
def test_data(dir_path):
x = read_imgs(dir_path, False)
return ImgDataset(x, None, test_transform)
if __name__ == '__main__':
parser = argparse.ArgumentParser(description = 'Preprocess for Faces')
parser.add_argument('path', type = str, help = 'path to image/images folder')
parser.add_argument('--lmk', default = 'Exec', \
help = 'Landmarks Detection Method')
parser.add_argument('--bfm', default = '/data/BaselFaceModel.mat', \
help = 'Morphable Face Model')
parser.add_argument('--mask', default = '', \
help = 'Skin Segmentation Model')
parser.add_argument('--disp', action = 'store_true', \
help = 'Show Landmarks')
parser.add_argument('--output', default = '', \
help = 'Output folder for processed images')
args = parser.parse_args()
from dataset import ImgDataset
data = ImgDataset(args.path)
if 'exe' in args.lmk.lower():
detector = LandmarksDetectorExec()
base_lmk = [0]*68
elif 'dlib' in args.lmk.lower():
detector = LandmarksDetectorDlib()
base_lmk = [0]*68
elif 'torch' in args.lmk.lower():
detector = LandmarksDetectorPytorch()
base_lmk = [0]*68
elif os.path.exists(args.lmk) and args.lmk[-4:].lower() == '.txt':
detector = LandmarksReader(args.lmk)
base_lmk = [0]*(args.lmk.shape[1]//2)
if 'torch' in args.mask.lower():
mask = SkinSegmentationPytorch()
else:
transforms.ToTensor(
), #將圖片轉成 Tensor，並把數值normalize到[0,1](data normalization)
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])
#testing 時不需做 data augmentation
test_transform = transforms.Compose([
transforms.ToPILImage(),
transforms.CenterCrop(224),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])
# Parameters initialize
batch_size = 32
train_set = ImgDataset(train_x, train_y, train_transform)
val_set = ImgDataset(val_x, val_y, test_transform)
train_loader = DataLoader(train_set, batch_size=batch_size, shuffle=True)
val_loader = DataLoader(val_set, batch_size=batch_size, shuffle=False)
model = EfficientNet.from_pretrained('efficientnet-b3',
num_classes=42).to(device)
#model = models.resnet152(pretrained=True).to(device)
#model.fc = nn.Linear(2048, 42).to(device)
#model.classifier = nn.Linear(1664, 42).to(device)
loss = nn.CrossEntropyLoss(
) # 因為是 classification task，所以 loss 使用 CrossEntropyLoss
optimizer = torch.optim.SGD(model.parameters(),
lr=0.0125,
momentum=0.9,
def main():
global args
args = parser.parse_args()
# create Light CNN for face recognition
if args.model == 'LightCNN-29':
model = LightCNN_29Layers(num_classes=args.num_classes)
elif args.model == 'LightCNN-29v2':
model = LightCNN_29Layers_v2(num_classes=args.num_classes)
else:
print('Error model type\n')
model = model.cuda()
print(model)
if not os.path.exists(args.save_path):
os.makedirs(args.save_path)
# large lr for last fc parameters
params = []
for name, value in model.named_parameters():
if 'bias' in name:
if 'fc2' in name:
params += [{
'params': value,
'lr': 20 * args.lr,
'weight_decay': 0
}]
else:
params += [{
'params': value,
'lr': 2 * args.lr,
'weight_decay': 0
}]
else:
if 'fc2' in name:
params += [{'params': value, 'lr': 10 * args.lr}]
else:
params += [{'params': value, 'lr': 1 * args.lr}]
optimizer = torch.optim.SGD(params,
args.lr,
momentum=args.momentum,
weight_decay=args.weight_decay)
# optionally resume from a checkpoint
state_dict = torch.load(args.model_path)
model.load_state_dict(state_dict)
cudnn.benchmark = True
# load image
train_loader = torch.utils.data.DataLoader(ImgDataset(
args.dataroot, False, args.crop, args.preload),
batch_size=args.batch_size,
shuffle=True,
num_workers=2,
pin_memory=True)
val_loader = torch.utils.data.DataLoader(ImgDataset(
args.dataroot, True, args.crop, args.preload),
batch_size=args.batch_size,
shuffle=False,
num_workers=0,
pin_memory=True)
# define loss function and optimizer
criterion = nn.CrossEntropyLoss()
criterion.cuda()
validate(val_loader, model)
for epoch in range(args.start_epoch, args.epochs):
adjust_learning_rate(optimizer, epoch)
# train for one epoch
train(train_loader, model, criterion, optimizer, epoch)
# evaluate on validation set
validate(val_loader, model)
if not os.path.exists(args.save_path):
os.makedirs(args.save_path)
if epoch % 5 == 0:
save_checkpoint(
model.state_dict(),
join(args.save_path,
'lightCNN_' + str(epoch + 1) + '_checkpoint.pth'))
save_checkpoint(model.state_dict(),
join(args.save_path, 'lightCNN_latest_checkpoint.pth'))
transforms.ToTensor(
), #將圖片轉成 Tensor，並把數值normalize到[0,1](data normalization)
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])
#testing 時不需做 data augmentation
test_transform = transforms.Compose([
transforms.ToPILImage(),
transforms.CenterCrop(224),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])
# Parameters initialize
batch_size = 32
train_set = ImgDataset(train_X, train_Y, train_transform)
val_set = ImgDataset(val_X, val_Y, test_transform)
train_loader = DataLoader(train_set, batch_size=batch_size, shuffle=True)
val_loader = DataLoader(val_set, batch_size=batch_size, shuffle=False)
model = Classifier().to(device)
#import torchvision.models as models
#model = models.densenet169(pretrained=False).to(device)
#model = models.resnet152(pretrained=False).to(device)
'''
model.classifier = nn.Sequential(
nn.Dropout(0.3),
nn.Linear(1664, 256),
nn.ReLU(inplace=True),
nn.Linear(256, 11)).to(device)
'''
print("raw maximized layer_activations imgs:",x.shape,"select:",0,"in",x.shape[0])
filter_visualization = x.detach().cpu().squeeze()[0]
hook_handle.remove() # reminber to rm it, or it exsits forever in every time forwarding
return filter_activations, filter_visualization
if __name__ == "__main__":
workspace_dir = sys.argv[1] #'/home/shannon/Downloads/food-11'
model_filename = sys.argv[2]
output_dir = sys.argv[3]
cnnids = [7,14,14,14,24]
filterids = [0,0,1,2,0]
print("Reading data")
train_x, train_y = readfile(os.path.join(workspace_dir, "training"), True)
print("Size of training data = {}".format(len(train_x)))
train_set = ImgDataset(train_x, train_y, test_transform)
print("Loading model")
model = Classifier().cuda()
model.load_state_dict(torch.load(model_filename))
# showing filters from assigned indices image
img_indices = [800,1602,2001,3201,4001,4800,5600,7000,7400,8003,8801]
images, labels = train_set.getbatch(img_indices)
for i, (cnnid,filterid) in enumerate(zip(cnnids,filterids)):
filter_activations, filter_visualization = filter_explaination(images, model, cnnid=cnnid, filterid=filterid, iteration=100, lr=0.1)
print(images.shape)
print(filter_activations.shape)
print(filter_visualization.shape)
else:
return x
if __name__ == '__main__':
#model_dir = sys.argv[1]
workspace_dir = sys.argv[1]
out_dir = sys.argv[2]
model = Classifier().cuda()
checkpoint = torch.load('VGG_150.pt')
model.load_state_dict(checkpoint)
train_paths, train_labels = readfile(workspace_dir, True)
# 這邊在 initialize dataset 時只丟「路徑」和「class」，之後要從 dataset 取資料時
# dataset 的 __getitem__ method 才會動態的去 load 每個路徑對應的圖片
train_set = ImgDataset(train_paths, train_labels, mode='eval')
# 指定想要一起 visualize 的圖片 indices
img_indices = [83, 4218, 4707, 8598]
img_indices2 = [993+200+709, 993+429+250+709]
images, labels = train_set.getbatch(img_indices)
images2, labels2 = train_set.getbatch2(img_indices)
images3, labels3 = train_set.getbatch(img_indices2)
saliencies = compute_saliency_maps(images, labels, model)
# 使用 matplotlib 畫出來
fig, axs = plt.subplots(2, len(img_indices), figsize=(15, 8))
for row, target in enumerate([images, saliencies]):
for column, img in enumerate(target):
img = cv2.cvtColor(img.permute(1, 2, 0).numpy(), cv2.COLOR_BGR2RGB)
def get_data(dir_path, ids):
x, y = read_imgs(dir_path, ids, True)
return ImgDataset(x, y, test_transform)
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5), inplace=True),
])
dataset = MultiResolutionDataset(args.path, transform, args.size)
loader = data.DataLoader(
dataset,
batch_size=args.batch,
sampler=data_sampler(dataset,
shuffle=True,
distributed=args.distributed),
drop_last=True,
)
content_sample_dataset = ImgDataset(args.content_sample_path, args.size,
'sample')
style_sample_dataset = ImgDataset(args.style_sample_path, args.size,
'sample')
content_loader_sample = data.DataLoader(
content_sample_dataset,
batch_size=args.n_sample,
sampler=data_sampler(content_sample_dataset,
shuffle=True,
distributed=args.distributed),
drop_last=True,
)
style_loader_sample = data.DataLoader(
style_sample_dataset,
batch_size=args.n_sample,
sampler=data_sampler(style_sample_dataset,
shuffle=True,
