# Transformation parameters
transform = tsfrm.Compose([
tsfrm.RandomHorizontalFlip(p=0.5),
tsfrm.RandomVerticalFlip(p=0.5),
tsfrm.RandomAffine(90,
translate=(0.15, 0.15),
scale=(0.75, 1.5),
resample=3,
fillcolor=0)
])
# Dataset definitions
dataset_train = OvaryDataset(im_dir='../datasets/ovarian/im/train/',
gt_dir='../datasets/ovarian/gt/train/',
clahe=False,
transform=transform,
ovary_inst=has_ovary)
dataset_val = OvaryDataset(im_dir='../datasets/ovarian/im/val/',
gt_dir='../datasets/ovarian/gt/val/',
clahe=False,
transform=False,
ovary_inst=has_ovary)
# Optmization
#optimizer = optim.Adam(model.parameters(), lr=0.001)
optimizer = optim.SGD(model.parameters(),
lr=0.005,
momentum=0.9,
weight_decay=0.0005)
]
gt_list = [
"../datasets/ovarian/gt/test", "../datasets/ovarian/gt/train",
"../datasets/ovarian/gt/val"
]
widths = []
heights = []
# Read all datasets
for (path_im, path_gt) in zip(im_list, gt_list):
# pre-set
dataset = OvaryDataset(im_dir=path_im,
gt_dir=path_gt,
clahe=False,
transform=False,
ovary_inst=ovary)
# Loader
data_loader = DataLoader(dataset,
batch_size=1,
shuffle=False,
collate_fn=dataset.collate_fn_yolo)
# iterate
for _, (fname, img, targets) in enumerate(data_loader):
# Load data
img_size = img.shape[-1]
# read each image in batch
opt = parser.parse_args()
print(opt)
# Classes names
class_names = ['background', 'follicle', 'ovary']
# Get data configuration
n_classes = opt.num_classes
has_ovary = True if n_classes > 2 else False
weights_path = opt.weights_path
# Dataset
dataset = OvaryDataset(im_dir='../datasets/ovarian/im/test/',
gt_dir='../datasets/ovarian/gt/test/',
clahe=False,
transform=False,
ovary_inst=has_ovary)
data_loader = DataLoader(dataset,
batch_size=opt.batch_size,
shuffle=False,
collate_fn=dataset.collate_fn_rcnn)
# Get device
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
# Initiate model
model = FasterRCNN(num_channels=opt.num_channels,
num_classes=n_classes,
pretrained=True).to(device)
if weights_path is not None:
elif net_type == 'unet_light':
model = UnetLight(n_channels=in_channels, n_classes=n_classes, bilinear=bilinear)
elif net_type == 'sp_unet':
model = SpatialPyramidUnet(n_channels=in_channels, n_classes=n_classes, bilinear=bilinear)
elif net_type == 'sp_unet2':
model = SpatialPyramidUnet2(n_channels=in_channels, n_classes=n_classes, bilinear=bilinear)
elif net_type == 'd_unet':
model = DilatedUnet2(n_channels=in_channels, n_classes=n_classes, bilinear=bilinear)
else:
model = Unet2(n_channels=in_channels, n_classes=n_classes, bilinear=bilinear)
# Load CUDA if exist
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
# Dataset definitions
dataset_test = OvaryDataset(im_dir='../datasets/ovarian/im/test/',
gt_dir='../datasets/ovarian/gt/test/')
# Test network model
print('Testing')
print('')
weights_path = folder_weights + train_name + '_weights.pth.tar'
# Output folder
out_folder = folder_preds + train_name + '/'
if not os.path.exists(out_folder):
os.makedirs(out_folder)
# Load inference
inference = Inference(model, device, weights_path,
batch_size=batch_size, folder=out_folder)
# Run inference
inference.predict(dataset_test)
]
data_table = [table_header]
# Bounding-box colors
cmap = plt.get_cmap("tab20b")
colormap = [cmap(i) for i in np.linspace(0, 1, 20)]
# colors = random.sample(colormap, len(dataset_names))
colors = [colormap[1], colormap[6], colormap[14]]
# Read datasets
for dname, fname in zip(dataset_names, dataset_folder):
# Set paths
path_im = dataset_path + '/im/' + fname + '/'
path_gt = path_im.replace('/im/', '/gt/')
# pre-set dataset
dataset = OvaryDataset(im_dir=path_im, gt_dir=path_gt, ovary_inst=True)
# define pytorch data loader
data_loader = DataLoader(dataset,
batch_size=1,
shuffle=False,
collate_fn=dataset.collate_fn_rcnn)
# Iterate dataset
for batch_idx, (names, imgs, targets) in enumerate(data_loader):
# Iterate inside batch
for i in range(len(names)):
filename = names[i]
full_path = os.path.join(path_im, filename)
# Create plot
img = np.array(
Image.open(full_path).convert('RGB')
) # Convert to RGB to save the image with original grayscale colormap
tsfrm.RandomVerticalFlip(p=0.5),
tsfrm.RandomAffine(90,
translate=(0.15, 0.15),
scale=(0.75, 1.5),
resample=3,
fillcolor=0)
])
# Dataset definitions
if dataset_name == 'ovarian':
im_dir = '../datasets/ovarian/im/'
gt_dir = '../datasets/ovarian/gt/'
dataset_train = OvaryDataset(im_dir=im_dir + 'train/',
gt_dir=gt_dir + 'train/',
imap=interaction,
clahe=clahe,
ovary_inst=train_with_targets,
out_tuple=train_with_targets,
transform=transform)
dataset_val = OvaryDataset(im_dir=im_dir + 'val/',
gt_dir=gt_dir + 'val/',
imap=interaction,
clahe=clahe)
dataset_test = OvaryDataset(im_dir=im_dir + 'test/',
gt_dir=gt_dir + 'test/',
imap=interaction,
clahe=clahe)
else:
im_dir = '../datasets/voc2012/JPEGImages/'
gt_dir = '../datasets/voc2012/SegmentationClass/'
list_dir = '../datasets/voc2012/'