def visualize_segmenter(self,
dataset,
export_path,
use_gpu=False,
with_ground_truth=False,
n_inst=None):
chk_mkdir(export_path)
self.s.train(False)
for batch_idx, (X_batch, y_batch,
name) in enumerate(DataLoader(dataset, batch_size=1)):
if n_inst is None:
n_inst = len(dataset)
else:
assert n_inst >= 0, 'n_inst must be nonnegative'
if use_gpu:
X_batch = Variable(X_batch.cuda())
y_out = self.s(X_batch).cpu().data.numpy()
X_im = X_batch[0].data.cpu().numpy().transpose((1, 2, 0))
else:
X_batch = Variable(X_batch)
y_out = self.s(X_batch).data.numpy()
X_im = X_batch[0].data.numpy().transpose((1, 2, 0))
if with_ground_truth:
visualize_segmentation(X=X_im,
y_mask=y_batch[0, 1, :, :].numpy(),
y_pred=y_out[0, 1, :, :],
export_path=os.path.join(
export_path, name[0] + '.png'))
else:
io.imsave(os.path.join(export_path, name[0] + '.png'),
y_out[0, 1, :, :])
if batch_idx > n_inst:
break
def visualize(self, dataset, n_inst=20, folder_name='comparison'):
self.model.train(False)
export_path = os.path.join(self.model_results_path, folder_name)
chk_mkdir(export_path)
for batch_idx, (X_batch, name) in enumerate(DataLoader(dataset, batch_size=1)):
if n_inst is None:
n_inst = len(dataset)
else:
assert n_inst >= 0, 'n_inst must be nonnegative'
if self.use_gpu:
X_batch = Variable(X_batch.cuda())
X_out = self.model(X_batch).cpu().data.numpy()
X_im = X_batch[0].data.cpu().numpy().transpose((1, 2, 0))
else:
X_batch = Variable(X_batch)
X_out = self.model(X_batch).data.numpy()
X_im = X_batch[0].data.numpy().transpose((1, 2, 0))
if sum([abs(X_out.shape[k] - X_batch.shape[k]) for k in range(len(X_out.shape))]) != 0:
print(X_out.shape)
visualize_segmentation(
X=X_im,
y_mask=1-X_im,
y_pred=1-X_out[0].transpose((1, 2, 0)),
export_path=os.path.join(export_path, name[0] + '.png')
)
if batch_idx > n_inst:
break
def visualize(self, dataset, folder_name='visualize', n_inst=None):
self.model.train(False)
export_path = os.path.join(self.model_results_path, folder_name)
chk_mkdir(export_path)
for batch_idx, (X_batch, y_mask_batch, y_rcnn_batch, name) in enumerate(DataLoader(dataset, batch_size=1)):
if n_inst is None:
n_inst = len(dataset)
else:
assert n_inst >= 0, 'n_inst must be nonnegative'
if self.use_gpu:
X_batch, y_rcnn_batch = Variable(X_batch.cuda()), Variable(y_rcnn_batch.cuda())
y_out = self.model(torch.cat([X_batch, y_rcnn_batch], dim=1)).cpu().data.numpy()
y_rcnn_batch = y_rcnn_batch.data.cpu().numpy()
X_im = X_batch[0].data.cpu().numpy().transpose((1, 2, 0))
else:
X_batch, y_rcnn_batch = Variable(X_batch), Variable(y_rcnn_batch)
y_out = self.model(torch.cat([X_batch, y_rcnn_batch], dim=1)).data.numpy()
y_rcnn_batch = y_rcnn_batch.data.numpy()
X_im = X_batch[0].data.numpy().transpose((1, 2, 0))
visualize_segmentation(
X=X_im,
y_mask=1 - y_out[0, 1:].transpose((1, 2, 0)),
y_pred=1 - y_out[0, 1:].transpose((1, 2, 0)),
export_path=os.path.join(export_path, name[0] + '.png')
)
if batch_idx > n_inst:
break
def visualize(self, dataset, n_inst=20, folder_name='comparison'):
self.model.train(False)
export_path = os.path.join(self.model_results_path, folder_name)
chk_mkdir(export_path)
for batch_idx, (X_batch, y_batch, *rest) in enumerate(DataLoader(dataset, batch_size=1)):
name = rest[-1]
if n_inst is None:
n_inst = len(dataset)
else:
assert n_inst >= 0, 'n_inst must be nonnegative'
if self.use_gpu:
X_batch = Variable(X_batch.cuda())
y_out = self.model(X_batch).cpu().data.numpy()
X_im = X_batch[0].data.cpu().numpy().transpose((1, 2, 0))
else:
X_batch = Variable(X_batch)
y_out = self.model(X_batch).data.numpy()
X_im = X_batch[0].data.numpy().transpose((1, 2, 0))
visualize_segmentation(
X=X_im,
y_mask=y_batch[0, 0, :, :].numpy(),
y_pred=y_out[0, 0],
export_path=os.path.join(export_path, name[0] + '.png')
)
if batch_idx > n_inst:
break
def visualize(self, dataset, folder_name):
self.model.train(False)
export_path = os.path.join(self.model_results_path, folder_name)
chk_mkdir(export_path)
for batch_idx, (X_image, y_prob, y_label, name) in enumerate(DataLoader(dataset, batch_size=1)):
if self.use_gpu:
X_image = Variable(X_image.cuda())
X_pred = self.model(X_image).data.cpu().numpy()
else:
X_image = Variable(X_image)
X_pred = self.model(X_image).data.numpy()
def predict(self, dataset, folder_name='prediction'):
self.model.train(False)
export_path = os.path.join(self.model_results_path, folder_name)
chk_mkdir(export_path)
for batch_idx, (X_batch, name) in enumerate(DataLoader(dataset, batch_size=1)):
if self.use_gpu:
X_batch = Variable(X_batch.cuda())
X_out = self.model(X_batch).cpu().data.numpy()
else:
X_batch = Variable(X_batch)
X_out = self.model(X_batch).data.numpy()
io.imsave(os.path.join(export_path, name[0] + '.png'), X_out[0].transpose((1, 2, 0)))
def __init__(
self, g, g_optim, g_loss, t, t_optim,
model_name, results_root_path, g_scheduler=None, t_scheduler=None
):
self.model_name = model_name
self.model_results_path = os.path.join(results_root_path, model_name)
self.use_gpu = torch.cuda.is_available()
chk_mkdir(self.model_results_path)
if self.use_gpu:
self.g.cuda()
self.g_loss.cuda()
self.t.cuda()
def __init__(
self, model, optimizer,
model_name, results_root_path,
scheduler=None
):
self.model = model
self.optimizer = optimizer
self.scheduler = scheduler
self.model_name = model_name
self.model_results_path = os.path.join(results_root_path, model_name)
self.use_gpu = torch.cuda.is_available()
if self.use_gpu:
self.model.cuda()
chk_mkdir(self.model_results_path)
def predict(self, dataset, folder_name='prediction'):
self.model.train(False)
export_path = os.path.join(self.model_results_path, folder_name)
chk_mkdir(export_path)
for batch_idx, (X_batch, y_mask_batch, y_rcnn_batch, name) in enumerate(DataLoader(dataset, batch_size=1)):
if self.use_gpu:
X_batch, y_rcnn_batch = Variable(X_batch.cuda()), Variable(y_rcnn_batch.cuda())
y_out = self.model(torch.cat([X_batch, y_rcnn_batch], dim=1)).cpu().data.numpy()
y_rcnn_batch = y_rcnn_batch.data.cpu().numpy()
X_im = X_batch[0].data.cpu().numpy().transpose((1, 2, 0))
else:
X_batch, y_rcnn_batch = Variable(X_batch), Variable(y_rcnn_batch)
y_out = self.model(torch.cat([X_batch, y_rcnn_batch], dim=1)).data.numpy()
y_rcnn_batch = y_rcnn_batch.data.numpy()
X_im = X_batch[0].data.numpy().transpose((1, 2, 0))
io.imsave(os.path.join(export_path, name[0] + '.png'), y_out[0].transpose((1, 2, 0)))
def __init__(
self, g, g_optim, d, d_optim, d_loss, model_name, results_root_path,
g_scheduler=None, d_scheduler=None,
):
self.g = g
self.d = d
self.g_optim = g_optim
self.d_optim = d_optim
self.d_loss = d_loss
self.g_scheduler = g_scheduler
self.d_scheduler = d_scheduler
self.model_name = model_name
self.model_results_path = os.path.join(results_root_path, model_name)
self.use_gpu = torch.cuda.is_available()
chk_mkdir(self.model_results_path)
if self.use_gpu:
self.g.cuda()
self.d.cuda()
self.d_loss.cuda()
def __init__(
self, model, loss, optimizer,
model_name, results_root_path,
scheduler=None, validation_loss=None
):
self.model = model
self.loss = loss
self.optimizer = optimizer
self.scheduler = scheduler
self.model_name = model_name
self.model_results_path = os.path.join(results_root_path, model_name)
self.use_gpu = torch.cuda.is_available()
if validation_loss is None:
self.validation_loss = loss
else:
self.validation_loss = validation_loss
if self.use_gpu:
self.model.cuda()
self.loss.cuda()
#self.validation_loss.cuda()
chk_mkdir(self.model_results_path)
y_out[0, 1, :, :])
if batch_idx > n_inst:
break
use_gpu = torch.cuda.is_available()
# paths
model_name = 'UNet_GANv2_test'
result_root_path = '/media/tdanka/B8703F33703EF828/tdanka'
all_datasets = '/home/tdanka/Data/new'
results_path = os.path.join(result_root_path, model_name)
s_comparison_path = os.path.join(results_path, 's_comparison')
# creating folders
chk_mkdir(result_root_path, results_path)
# datasets
train_dataset = TrainFromFolder(os.path.join(
all_datasets,
'/home/tdanka/Data/new/stage1_train_tissue_weighted_patch=256/loc.csv'),
transform=T.ToTensor(),
remove_alpha=True,
one_hot_mask=2,
weighted=True)
x = next(iter(train_dataset))
train_original_dataset = TrainFromFolder(os.path.join(
all_datasets, 'stage1_train_merged/loc.csv'),
transform=T.ToTensor(),
remove_alpha=True,
one_hot_mask=2)
test_dataset = TestFromFolder(os.path.join(all_datasets,
