def test(img_path, data_size='single'):
device = torch.device('cuda') if torch.cuda.is_available else torch.device(
'cpu')
# Image input
im = Image.open(img_path)
im = np.array(im, dtype=np.float32) / 255
image = np.transpose(im, (2, 0, 1))
data = torch.from_numpy(image).unsqueeze(0)
data = Variable(data).to(device)
model = SegNet(opt, data.shape[1])
if opt.model_path:
model.load_state_dict(torch.load(opt.model_path))
model = model.to(device)
model.train()
feats, output = model(data)
output = output[0].permute(1, 2, 0).contiguous().view(-1, opt.nClass)
feats = feats[0].permute(1, 2, 0).contiguous().view(-1, opt.nChannel)
_, pred_clusters = torch.max(output, 1)
pred_clusters = pred_clusters.data.cpu().numpy()
# Post processing
labels = np.unique(pred_clusters)
counts = {}
for i in pred_clusters:
counts[i] = counts.get(i, 0) + 1
sorts = sorted(counts.items(), key=lambda x: x[1])
cache = {}
cache[sorts[-1][0]] = 0
n = 1
for (num, _) in sorts[:-1]:
cache[num] = n
n += 1
label_colors = [[10, 10, 10], [0, 0, 255], [0, 255, 0], [255, 0, 0],
[255, 255, 0], [0, 255, 255], [255, 0, 255]]
im_target_rgb = np.array([label_colors[cache[c]] for c in pred_clusters])
im_target_rgb = im_target_rgb.reshape(im.shape).astype(np.uint8)
# change path
path = ".".join(img_path.split('/')[1].split('.')[:2])
#path = img_path.split('/')[1].split('.')[0]
if data_size == 'single':
cv2.imwrite("outputs_single/{}_out.png".format(path), im_target_rgb)
elif data_size == 'all':
cv2.imwrite("outputs_all/{}_out.png".format(path), im_target_rgb)
def test_autoencoder(epoch_plus, text, index):
use_gpu = torch.cuda.is_available()
ngpu = torch.cuda.device_count()
device = torch.device("cuda:0" if (torch.cuda.is_available() and ngpu > 0) else "cpu")
model = SegNet(3)
if ngpu > 1:
model = nn.DataParallel(model)
if use_gpu:
model = model.to(device, non_blocking=True)
text = text.to(device, non_blocking=True)
if epoch_plus > 0:
model.load_state_dict(torch.load('./autoencoder_models_2/autoencoder_{}.pth'.format(epoch_plus)))
model.eval()
if use_gpu:
text.to(device, non_blocking=True)
predicted = model(text)
predicted[predicted > 1.0] = 1.0
save_path1 = './results/text'
save_path2 = './results/masked'
if not os.path.exists(save_path1):
os.mkdir(save_path1)
if not os.path.exists(save_path2):
os.mkdir(save_path2)
binary_predicted = predicted.clone()
binary_mask = predicted.clone()
binary_predicted[binary_predicted > 0.0] = 1.0
binary_mask[binary_mask > 0.1] = 1.0
masked = text + binary_mask
masked[masked > 1.0] = 1.0
trans = torchvision.transforms.ToPILImage()
predicted = predicted.squeeze().cpu()
masked = masked.squeeze().cpu()
image = trans(predicted)
image2 = trans(masked)
image.save(os.path.join(save_path1, 'text_{}.png'.format(index)))
image2.save(os.path.join(save_path2, 'masked_{}.png'.format(index)))
del text
del predicted
del masked
del binary_predicted
def predict_image(dir):
use_gpu = torch.cuda.is_available()
ngpu = torch.cuda.device_count()
device = torch.device("cuda:0" if (
torch.cuda.is_available() and ngpu > 0) else "cpu")
image_to_tensor = torchvision.transforms.ToTensor()
tensor_to_image = torchvision.transforms.ToPILImage()
save_path = Path(dir).parent
image = Image.open(dir).convert('RGB')
image = image_to_tensor(image)
c, w, h = image.shape
image = torch.reshape(image, (1, c, w, h))
model = SegNet(3)
if use_gpu:
model = model.to(device, non_blocking=True)
image = image.to(device, non_blocking=True)
model.load_state_dict(torch.load('./models/model.pth',
map_location=device))
model.eval()
predicted = model(image)
predicted[predicted > 1.0] = 1.0
binary_predicted = predicted.clone()
binary_mask = predicted.clone()
binary_predicted[binary_predicted > 0.0] = 1.0
binary_mask[binary_mask > 0.1] = 1.0
masked = image + binary_mask
masked[masked > 1.0] = 1.0
predicted = predicted.squeeze().cpu()
masked = masked.squeeze().cpu()
image = tensor_to_image(predicted)
image2 = tensor_to_image(masked)
image.save(os.path.join(save_path, 'tmp_text.png'))
image2.save(os.path.join(save_path, 'tmp_masked.png'))
def train(data_size='all'):
device = torch.device('cuda') if torch.cuda.is_available else torch.device(
'cpu')
# Image input
model = SegNet(opt, 3)
model = model.to(device)
model.train()
criterion = torch.nn.CrossEntropyLoss()
criterion_d = DiscriminativeLoss()
optimizer = SGD(model.parameters(), lr=opt.lr, momentum=opt.momentum)
if data_size == 'all':
dataloader = get_dataloader(opt.paths, opt, device)
model = batch_step(opt, optimizer, model, dataloader, criterion,
criterion_d, device)
torch.save(model.state_dict(), 'model_all.pth')
else:
im = Image.open(opt.img_path)
im = np.array(im, dtype=np.float32) / 255
image = np.transpose(im, (2, 0, 1))
data = torch.from_numpy(image).unsqueeze(0)
data = Variable(data).to(device)
labels = segmentation.slic(im,
compactness=opt.compactness,
n_segments=opt.num_superpixels)
labels = labels.reshape(-1)
label_nums = np.unique(labels)
label_indices = [
np.where(labels == label_nums[i])[0]
for i in range(len(label_nums))
]
model = one_step(opt, optimizer, model, data, label_indices, criterion,
criterion_d, device)
torch.save(model.state_dict(), 'model_single.pth')
v_data = Cityscapes(
'./data',
target_type=['semantic'],
split='val',
transform=in_t,
target_transform=out_t,
)
train_data = DataLoader(tr_data, batch_size=BATCH_SIZE, shuffle=True)
val_data = DataLoader(v_data, batch_size=BATCH_SIZE, shuffle=False)
# define model
model = SegNet(IN_CHANNELS, CLASSES)
if gpu:
model.to(torch.device("cuda:0"))
if monitor:
wandb.watch(model)
# optimizer and loss definition
criterion = torch.nn.CrossEntropyLoss().cuda()
optimizer = torch.optim.SGD(model.parameters(), lr=LR, momentum=0.9)
# training
for epoch in range(EPOCHS):
tr_epoch_loss = 0
tr_epoch_iou = 0
tr_batch_count = 0
val_batch_count = 0
def train_autoencoder(epoch_plus):
writer = SummaryWriter(log_dir='./runs_autoencoder_2')
num_epochs = 400 - epoch_plus
lr = 0.001
bta1 = 0.9
bta2 = 0.999
weight_decay = 0.001
# model = autoencoder(nchannels=3, width=172, height=600)
model = SegNet(3)
if ngpu > 1:
model = nn.DataParallel(model)
if use_gpu:
model = model.to(device, non_blocking=True)
if epoch_plus > 0:
model.load_state_dict(
torch.load('./autoencoder_models_2/autoencoder_{}.pth'.format(
epoch_plus)))
criterion = nn.MSELoss(reduction='sum')
optimizer = torch.optim.Adam(model.parameters(),
lr=lr,
betas=(bta1, bta2),
weight_decay=weight_decay)
for epoch in range(num_epochs):
degree = randint(-180, 180)
transforms = torchvision.transforms.Compose([
torchvision.transforms.CenterCrop((172, 200)),
torchvision.transforms.Resize((172, 200)),
torchvision.transforms.RandomRotation((degree, degree)),
torchvision.transforms.ToTensor()
])
dataloader = get_dataloader(data_dir,
train=True,
transform=transforms,
batch_size=batch_size)
model.train()
epoch_losses = AverageMeter()
with tqdm(total=(1000 - 1000 % batch_size)) as _tqdm:
_tqdm.set_description('epoch: {}/{}'.format(
epoch + 1 + epoch_plus, num_epochs + epoch_plus))
for data in dataloader:
gt, text = data
if use_gpu:
gt, text = gt.to(device, non_blocking=True), text.to(
device, non_blocking=True)
predicted = model(text)
# loss = criterion_bce(predicted, gt) + criterion_dice(predicted, gt)
loss = criterion(
predicted, gt - text
) # predicts extracted text in white, all others in black
epoch_losses.update(loss.item(), len(gt))
optimizer.zero_grad()
loss.backward()
optimizer.step()
_tqdm.set_postfix(loss='{:.6f}'.format(epoch_losses.avg))
_tqdm.update(len(gt))
save_path = './autoencoder_models_2'
if not os.path.exists(save_path):
os.mkdir(save_path)
gt_text = gt - text
predicted_mask = text + predicted
torch.save(
model.state_dict(),
os.path.join(save_path,
'autoencoder_{}.pth'.format(epoch + 1 + epoch_plus)))
writer.add_scalar('Loss', epoch_losses.avg, epoch + 1 + epoch_plus)
writer.add_image('text/text_image_{}'.format(epoch + 1 + epoch_plus),
text[0].squeeze(), epoch + 1 + epoch_plus)
writer.add_image('gt/gt_image_{}'.format(epoch + 1 + epoch_plus),
gt[0].squeeze(), epoch + 1 + epoch_plus)
writer.add_image('gt_text/gt_image_{}'.format(epoch + 1 + epoch_plus),
gt_text[0].squeeze(), epoch + 1 + epoch_plus)
writer.add_image(
'predicted/predicted_image_{}'.format(epoch + 1 + epoch_plus),
predicted_mask[0].squeeze(), epoch + 1 + epoch_plus)
writer.add_image(
'predicted_text/predicted_image_{}'.format(epoch + 1 + epoch_plus),
predicted[0].squeeze(), epoch + 1 + epoch_plus)
writer.close()