def testing():
c = get_config()
c.do_load_checkpoint = True
#c.checkpoint_dir = c.base_dir + '/20190424-020641_unet_experiment' + '/checkpoint/checkpoint_current' # dice_cost train
# c.checkpoint_dir = c.base_dir + '/20190424-234657_unet_experiment' + '/checkpoint/checkpoint_last' # SDG
c.checkpoint_dir = c.base_dir + '/20190906-085449_unet_experiment' + '/checkpoint/checkpoint_current'
# c.checkpoint_file = "checkpoint_last.pth.tar"
# c.cross_vali_index = valiIndex
cross_vali_result_all_dir = os.path.join(c.base_dir, c.dataset_name
+ '_' + str(
c.batch_size) + c.cross_vali_result_all_dir + datetime.datetime.now().strftime("_%Y%m%d-%H%M%S"))
if not os.path.exists(cross_vali_result_all_dir):
os.makedirs(cross_vali_result_all_dir)
print('Created' + cross_vali_result_all_dir + '...')
c.base_dir = cross_vali_result_all_dir
c.cross_vali_result_all_dir = os.path.join(cross_vali_result_all_dir, "results")
os.makedirs(c.cross_vali_result_all_dir)
exp = UNetExperiment(config=c, name='unet_test', n_epochs=c.n_epochs,
seed=42, globs=globals())
exp.run_test(setup=True)
def training():
c = get_config()
dataset_name = 'Task04_Hippocampus'
# dataset_name = 'Task02_Heart'
# download_dataset(dest_path=c.data_root_dir, dataset=dataset_name, id=c.google_drive_id)
# c.do_load_checkpoint = True
# c.checkpoint_dir = c.base_dir + '/20190801-_unet_experiment' + '/checkpoint/checkpoint_current'
# c.checkpoint_file = "checkpoint_last.pth.tar"
if not exists(os.path.join(os.path.join(c.data_root_dir, dataset_name), 'preprocessed')):
print('Preprocessing data. [STARTED]')
preprocess_data(root_dir=os.path.join(c.data_root_dir, dataset_name))
create_splits(output_dir=c.split_dir, image_dir=c.data_dir)
print('Preprocessing data. [DONE]')
else:
print('The data has already been preprocessed. It will not be preprocessed again. Delete the folder to enforce it.')
exp = UNetExperiment(config=c, name='unet_experiment', n_epochs=c.n_epochs,
seed=42, append_rnd_to_name=c.append_rnd_string) # visdomlogger_kwargs={"auto_start": c.start_visdom}
exp.run()
exp.run_test(setup=False)
c = get_config()
#download_dataset(dest_path=c.data_root_dir, dataset=c.dataset_name, id=c.google_drive_id)
if not exists(
os.path.join(os.path.join(c.data_root_dir, c.dataset_name),
'preprocessed')):
print('Preprocessing data. [STARTED]')
#preprocess_data(root_dir=os.path.join(c.data_root_dir, c.dataset_name), y_shape=c.patch_size, z_shape=c.patch_size)
#create_splits(output_dir=c.split_dir, image_dir=c.data_dir)
print('Preprocessing data. [DONE]')
else:
print(
'The data has already been preprocessed. It will not be preprocessed again. Delete the folder to enforce it.'
)
exp = UNetExperiment(
config=c,
name=c.name,
n_epochs=c.n_epochs,
seed=42,
append_rnd_to_name=c.append_rnd_string,
globs=globals(),
# visdomlogger_kwargs={"auto_start": c.start_visdom},
loggers={"visdom": ("visdom", {
"auto_start": c.start_visdom
})})
exp.run()
exp.run_test(setup=False)
class LabelTensorToColor(object):
def __call__(self, label):
label = label.squeeze()
colored_label = torch.zeros(3, label.size(0), label.size(1)).byte()
for i, color in enumerate(class_color):
mask = label.eq(i)
for j in range(3):
colored_label[j].masked_fill_(mask, color[j])
return colored_label
color_class_converter = LabelTensorToColor()
# Load data
c = get_config()
exp = UNetExperiment(config=c, name=c.name, n_epochs=c.n_epochs,
seed=42, append_rnd_to_name=c.append_rnd_string, globs=globals())
exp.setup()
exp.test_data_loader.do_reshuffle = True
# Load checkpoint
checkpoint = torch.load('./output_experiment/20190604-154315_fine_tune_spleen_for_heart/checkpoint/checkpoint_last.pth.tar')
exp.model.load_state_dict(checkpoint['model'])
exp.model.eval()
batch_counter = 0
with torch.no_grad():
for data_batch in exp.test_data_loader:
# Get data_batches
mr_data = data_batch['data'][0].float().to(exp.device)
mr_target = data_batch['seg'][0].float().to(exp.device)
# the array with indices of training volumes to be used for training, validation
# and testing respectively.
# <YOUR CODE GOES HERE>
#__solution
# We are splitting 60:20:20 for train: val and test for the dataset
split = {
'train': keys[0:int(max(keys) * 0.6)],
'val': keys[int(max(keys) * 0.6):int(max(keys) * 0.8)],
'test': keys[int(max(keys) * 0.8):max(keys)]
}
# Set up and run experiment
# TASK: Class UNetExperiment has missing pieces. Go to the file and fill them in
#__solution in UNetExperiment
exp = UNetExperiment(c, split, data)
# You could free up memory by deleting the dataset
# as it has been copied into loaders
# del dataset
# run training
exp.run()
# prep and run testing
# TASK: Test method is not complete. Go to the method and complete it
results_json = exp.run_test()
results_json["config"] = vars(c)
)
exit()
shapesize = 64
if len(sys.argv) > 3:
shapesize = int(sys.argv[3])
# Get the header in order to preserve voxel dimensions to store the segmented image later on
print('Preprocessing data.')
data, header = preprocess_single_file(filename,
y_shape=shapesize,
z_shape=shapesize)
print('Setting up model and start segmentation.')
exp = UNetExperiment(config=c,
name=c.name,
n_epochs=c.n_epochs,
seed=42,
append_rnd_to_name=c.append_rnd_string,
globs=globals())
result = exp.segment_single_image(data)
print('Postprocessing data.')
result = postprocess_single_image(result)
pathname, fname = os.path.split(filename)
destination_filename = pathname + "/segmented_" + fname
print('Saving file to disk: {}'.format(destination_filename))
save_single_image(result, header, destination_filename)
if __name__ == "__main__":
c = get_config()
dataset_name = 'Task04_Hippocampus'
# dataset_name = 'Task01_BrainTumour'
download_dataset(dest_path=c.data_root_dir,
dataset=dataset_name,
id=c.google_drive_id)
if not exists(
os.path.join(os.path.join(c.data_root_dir, dataset_name),
'preprocessed')):
print('Preprocessing data. [STARTED]')
preprocess_data(root_dir=os.path.join(c.data_root_dir, dataset_name))
create_splits(output_dir=c.split_dir, image_dir=c.data_dir)
print('Preprocessing data. [DONE]')
else:
print(
'The data has already been preprocessed. It will not be preprocessed again. Delete the folder to enforce it.'
)
exp = UNetExperiment(config=c,
name='unet_experiment',
n_epochs=c.n_epochs,
seed=42,
append_rnd_to_name=c.append_rnd_string,
visdomlogger_kwargs={"auto_start": c.start_visdom})
exp.run()
exp.run_test(setup=False)
# a k-fold training and combining the results.
# lets split our dataset as 60, 20, 20
train, validate, test = np.split(
all_idx, [int(.6 * len(all_idx)),
int(.8 * len(all_idx))])
split = dict()
# create three keys in the dictionary: "train", "val" and "test". In each key, store
# the array with indices of training volumes to be used for training, validation
# and testing respectively.
split = {"train": train, "val": validate, "test": test}
# Set up and run experiment
exp = UNetExperiment(c, split, data)
# You could free up memory by deleting the dataset
# as it has been copied into loaders
# del dataset
# summary
exp.print_summary()
# run training
#exp.run()
exp.load_model_parameters(path='..\\out\\model.pth')
# prep and run testing
results_json = exp.run_test()
# limitations under the License.
import matplotlib
matplotlib.use('Agg')
from configs.Config_unet import get_config
from experiments.UNetExperiment import UNetExperiment
if __name__ == "__main__":
c = get_config()
exp = UNetExperiment(
config=c,
name=c.name,
n_epochs=c.n_epochs,
seed=42,
append_rnd_to_name=c.append_rnd_string,
globs=globals(),
# visdomlogger_kwargs={"auto_start": c.start_visdom},
loggers={
"visdom": ("visdom", {
"auto_start": c.start_visdom
}),
# "tb": ("tensorboard"),
# "slack": ("slack", {"token": "XXXXXXXX",
# "user_email": "x"})
})
exp.run()
exp.run_test(setup=False)
def run_on_dataset():
#models = os.listdir('./models/')
models = glob('./output_experiment/*/checkpoint/checkpoint_last.pth.tar')
metric_list = [
"Dice", "Hausdorff Distance", "Hausdorff Distance 95",
"Avg. Symmetric Surface Distance", "Avg. Surface Distance"
]
summary = pd.DataFrame(columns=['modelname'] + metric_list)
summary_mean = pd.DataFrame(columns=['modelname'] + metric_list)
for model_name in models:
# Load data
print('starting with ' + model_name)
exp = UNetExperiment(config=c,
name=c.name,
n_epochs=c.n_epochs,
seed=42,
append_rnd_to_name=c.append_rnd_string,
globs=globals())
exp.setup()
exp.test_data_loader.do_reshuffle = False
# Load checkpoint
checkpoint = torch.load(model_name)
exp.model.load_state_dict(checkpoint['model'])
# exp.model.eval() # done in UNetExperiment
exp.run_test(setup=False)
#TODO get metrics
# select interesting ones, add to pandas dataframe
json_scores = exp.scores
result_dict = json_scores[
"results"] #{"all": results, "mean": results_mean}
im_scores = result_dict['all']
model_res = {'modelname': model_name}
print('adding results to summary..')
for nr, im in enumerate(im_scores):
for cl in range(c.num_classes):
for m in metric_list:
model_res['class_' + str(cl) + '_imagenr_' + str(nr) +
'_' + m] = im[str(cl)][m]
summary = summary.append(model_res, ignore_index=True)
im_mean_scores = result_dict['mean']
model_res_mean = {'modelname': model_name}
for cl in range(c.num_classes):
for m in metric_list:
model_res_mean['class_' + str(cl) + '_' +
m] = im_mean_scores[str(cl)][m]
summary_mean = summary_mean.append(model_res_mean, ignore_index=True)
summary.to_csv('./output_experiment/summary_ims' + c.dataset_name +
'.csv')
summary_mean.to_csv('./output_experiment/summary_mean_' +
c.dataset_name + '.csv')
print('..summary updated')
