def main(options):
# Read json
model_options = argparse.Namespace()
json_path = path.join(options.model_path, "commandline_cnn.json")
model_options = read_json(model_options, json_path=json_path)
num_cnn = compute_num_cnn(options.input_dir,
options.tsv_path,
model_options,
data="test")
# Load test data
if options.diagnoses is None:
options.diagnoses = model_options.diagnoses
test_df = load_data_test(options.tsv_path, options.diagnoses)
transformations = get_transforms(model_options.mode,
model_options.minmaxnormalization)
criterion = torch.nn.CrossEntropyLoss()
# Loop on all folds trained
best_model_dir = os.path.join(options.model_path, 'best_model_dir')
folds_dir = os.listdir(best_model_dir)
# Loop on folds
for fold_dir in folds_dir:
split = int(fold_dir[-1])
print("Fold %i" % split)
for cnn_index in range(num_cnn):
dataset = return_dataset(model_options.mode,
options.input_dir,
test_df,
options.preprocessing,
transformations,
options,
cnn_index=cnn_index)
test_loader = DataLoader(dataset,
batch_size=options.batch_size,
shuffle=False,
num_workers=options.num_workers,
pin_memory=True)
test_cnn(options.model_path, test_loader, options.dataset, split,
criterion, cnn_index, model_options, options.gpu)
for selection in ['best_acc', 'best_loss']:
soft_voting_to_tsvs(
options.model_path,
split,
selection,
mode=options.mode,
dataset=options.dataset,
num_cnn=num_cnn,
selection_threshold=model_options.selection_threshold)
if __name__ == "__main__":
ret = parser.parse_known_args()
options = ret[0]
if ret[1]:
print("unknown arguments: %s" % parser.parse_known_args()[1])
# Read json
model_options = argparse.Namespace()
json_path = path.join(options.model_path, "commandline_cnn.json")
model_options = read_json(model_options, json_path=json_path)
num_cnn = compute_num_cnn(model_options.input_dir,
model_options.tsv_path,
model_options,
data="train")
transformations = get_transforms(model_options.mode,
model_options.minmaxnormalization)
criterion = nn.CrossEntropyLoss()
# Loop on all folds trained
best_model_dir = os.path.join(options.model_path, 'best_model_dir')
folds_dir = os.listdir(best_model_dir)
for fold_dir in folds_dir:
split = int(fold_dir[-1])
print("Fold %i" % split)
# Data management
training_df, valid_df = load_data(model_options.tsv_path,
model_options.diagnoses, split,
model_options.n_splits,
model_options.baseline)
def group_backprop(options):
main_logger = return_logger(options.verbose, "main process")
options = translate_parameters(options)
fold_list = [
fold for fold in os.listdir(options.model_path) if fold[:5:] == "fold-"
]
if len(fold_list) == 0:
raise ValueError("No folds were found at path %s" % options.model_path)
for fold in fold_list:
main_logger.info(fold)
for selection in options.selection:
results_path = path.join(options.model_path, fold, 'gradients',
selection, options.name)
model_options = argparse.Namespace()
model_options = read_json(
model_options, path.join(options.model_path,
'commandline.json'))
model_options = translate_parameters(model_options)
model_options.gpu = options.gpu
if options.tsv_path is None:
options.tsv_path = model_options.tsv_path
if options.input_dir is None:
options.input_dir = model_options.input_dir
if options.target_diagnosis is None:
options.target_diagnosis = options.diagnosis
criterion = get_criterion(model_options.loss)
# Data management (remove data not well predicted by the CNN)
training_df = load_data_test(options.tsv_path, [options.diagnosis],
baseline=options.baseline)
training_df.reset_index(drop=True, inplace=True)
# Model creation
_, all_transforms = get_transforms(
model_options.mode,
minmaxnormalization=model_options.minmaxnormalization)
data_example = return_dataset(model_options.mode,
options.input_dir,
training_df,
model_options.preprocessing,
train_transformations=None,
all_transformations=all_transforms,
params=options)
model = create_model(model_options, data_example.size)
model_dir = os.path.join(options.model_path, fold, 'models',
selection)
model, best_epoch = load_model(model,
model_dir,
gpu=options.gpu,
filename='model_best.pth.tar')
options.output_dir = results_path
commandline_to_json(options, logger=main_logger)
# Keep only subjects who were correctly / wrongly predicted by the network
training_df = sort_predicted(model,
training_df,
options.input_dir,
model_options,
criterion,
options.keep_true,
batch_size=options.batch_size,
num_workers=options.num_workers,
gpu=options.gpu)
if len(training_df) > 0:
# Save the tsv files used for the saliency maps
training_df.to_csv(path.join('data.tsv'),
sep='\t',
index=False)
data_train = return_dataset(model_options.mode,
options.input_dir,
training_df,
model_options.preprocessing,
train_transformations=None,
all_transformations=all_transforms,
params=options)
train_loader = DataLoader(data_train,
batch_size=options.batch_size,
shuffle=True,
num_workers=options.num_workers,
pin_memory=True)
interpreter = VanillaBackProp(model, gpu=options.gpu)
cum_map = 0
for data in train_loader:
if options.gpu:
input_batch = data['image'].cuda()
else:
input_batch = data['image']
maps = interpreter.generate_gradients(
input_batch,
data_train.diagnosis_code[options.target_diagnosis])
cum_map += maps.sum(axis=0)
mean_map = cum_map / len(data_train)
if len(data_train.size) == 4:
if options.nifti_template_path is not None:
image_nii = nib.load(options.nifti_template_path)
affine = image_nii.affine
else:
affine = np.eye(4)
mean_map_nii = nib.Nifti1Image(mean_map[0], affine)
nib.save(mean_map_nii, path.join(results_path,
"map.nii.gz"))
np.save(path.join(results_path, "map.npy"), mean_map[0])
else:
jpg_path = path.join(results_path, "map.jpg")
plt.imshow(mean_map[0],
cmap="coolwarm",
vmin=-options.vmax,
vmax=options.vmax)
plt.colorbar()
plt.savefig(jpg_path)
plt.close()
numpy_path = path.join(results_path, "map.npy")
np.save(numpy_path, mean_map[0])
else:
main_logger.warn("There are no subjects for the given options")
def inference_from_model_generic(caps_dir,
tsv_path,
model_path,
model_options,
num_cnn=None,
selection="best_balanced_accuracy"):
'''
Inference using an image/subject CNN model
'''
from os.path import join
gpu = not model_options.use_cpu
# Recreate the model with the network described in the json file
# Initialize the model
model = create_model(model_options.model,
gpu,
dropout=model_options.dropout)
transformations = get_transforms(model_options.mode,
model_options.minmaxnormalization)
# Define loss and optimizer
criterion = nn.CrossEntropyLoss()
if model_options.mode_task == 'multicnn':
predictions_df = pd.DataFrame()
metrics_df = pd.DataFrame()
for n in range(num_cnn):
dataset = return_dataset(model_options.mode,
caps_dir,
tsv_path,
model_options.preprocessing,
transformations,
model_options,
cnn_index=n)
test_loader = DataLoader(dataset,
batch_size=model_options.batch_size,
shuffle=False,
num_workers=model_options.nproc,
pin_memory=True)
# load the best trained model during the training
model, best_epoch = load_model(model,
join(model_path, 'cnn-%i' % n,
selection),
gpu,
filename='model_best.pth.tar')
cnn_df, cnn_metrics = test(model,
test_loader,
gpu,
criterion,
mode=model_options.mode)
predictions_df = pd.concat([predictions_df, cnn_df])
metrics_df = pd.concat(
[metrics_df, pd.DataFrame(cnn_metrics, index=[0])])
predictions_df.reset_index(drop=True, inplace=True)
metrics_df.reset_index(drop=True, inplace=True)
else:
# Load model from path
best_model, best_epoch = load_model(model,
join(model_path, selection),
gpu,
filename='model_best.pth.tar')
# Read/localize the data
data_to_test = return_dataset(model_options.mode, caps_dir, tsv_path,
model_options.preprocessing,
transformations, model_options)
# Load the data
test_loader = DataLoader(data_to_test,
batch_size=model_options.batch_size,
shuffle=False,
num_workers=model_options.nproc,
pin_memory=True)
# Run the model on the data
predictions_df, metrics = test(best_model,
test_loader,
gpu,
criterion,
mode=model_options.mode)
metrics_df = pd.DataFrame(metrics, index=[0])
return predictions_df, metrics_df
def inference_from_model_generic(caps_dir,
tsv_path,
model_path,
model_options,
prefix,
output_dir,
fold,
selection,
labels=True,
num_cnn=None,
logger=None):
from os.path import join
import logging
if logger is None:
logger = logging
gpu = not model_options.use_cpu
_, all_transforms = get_transforms(model_options.mode,
model_options.minmaxnormalization)
test_df = load_data_test(tsv_path, model_options.diagnoses)
# Define loss and optimizer
criterion = get_criterion(model_options.loss)
if model_options.mode_task == 'multicnn':
for n in range(num_cnn):
test_dataset = return_dataset(model_options.mode,
caps_dir,
test_df,
model_options.preprocessing,
train_transformations=None,
all_transformations=all_transforms,
params=model_options,
cnn_index=n,
labels=labels)
test_loader = DataLoader(test_dataset,
batch_size=model_options.batch_size,
shuffle=False,
num_workers=model_options.nproc,
pin_memory=True)
# load the best trained model during the training
model = create_model(model_options, test_dataset.size)
model, best_epoch = load_model(model,
join(model_path, 'cnn-%i' % n,
selection),
gpu,
filename='model_best.pth.tar')
cnn_df, cnn_metrics = test(model,
test_loader,
gpu,
criterion,
mode=model_options.mode,
use_labels=labels)
if labels:
logger.info(
"%s balanced accuracy is %f for %s %i and model selected on %s"
% (prefix, cnn_metrics["balanced_accuracy"],
model_options.mode, n, selection))
mode_level_to_tsvs(output_dir,
cnn_df,
cnn_metrics,
fold,
selection,
model_options.mode,
dataset=prefix,
cnn_index=n)
else:
# Read/localize the data
test_dataset = return_dataset(model_options.mode,
caps_dir,
test_df,
model_options.preprocessing,
train_transformations=None,
all_transformations=all_transforms,
params=model_options,
labels=labels)
# Load the data
test_loader = DataLoader(test_dataset,
batch_size=model_options.batch_size,
shuffle=False,
num_workers=model_options.nproc,
pin_memory=True)
# Load model from path
model = create_model(model_options, test_dataset.size)
best_model, best_epoch = load_model(model,
join(model_path, selection),
gpu,
filename='model_best.pth.tar')
# Run the model on the data
predictions_df, metrics = test(best_model,
test_loader,
gpu,
criterion,
mode=model_options.mode,
use_labels=labels)
if labels:
logger.info(
"%s level %s balanced accuracy is %f for model selected on %s"
% (model_options.mode, prefix, metrics["balanced_accuracy"],
selection))
mode_level_to_tsvs(output_dir,
predictions_df,
metrics,
fold,
selection,
model_options.mode,
dataset=prefix)
def inference_from_model_generic(caps_dir,
tsv_path,
model_path,
model_options,
num_cnn=None,
selection="best_balanced_accuracy"):
'''
Inference using an image/subject CNN model
'''
from os.path import join
gpu = not model_options.use_cpu
# Recreate the model with the network described in the json file
# Initialize the model
if model_options.model == 'UNet3D':
print('********** init UNet3D model for test! **********')
model = create_model(model_options.model,
gpu=model_options.gpu,
dropout=model_options.dropout,
device_index=model_options.device,
in_channels=model_options.in_channels,
out_channels=model_options.out_channels,
f_maps=model_options.f_maps,
layer_order=model_options.layer_order,
num_groups=model_options.num_groups,
num_levels=model_options.num_levels)
elif model_options.model == 'ResidualUNet3D':
print('********** init ResidualUNet3D model for test! **********')
model = create_model(model_options.model,
gpu=model_options.gpu,
dropout=model_options.dropout,
device_index=model_options.device,
in_channels=model_options.in_channels,
out_channels=model_options.out_channels,
f_maps=model_options.f_maps,
layer_order=model_options.layer_order,
num_groups=model_options.num_groups,
num_levels=model_options.num_levels)
elif model_options.model == 'UNet3D_add_more_fc':
print('********** init UNet3D_add_more_fc model for test! **********')
model = create_model(model_options.model,
gpu=model_options.gpu,
dropout=model_options.dropout,
device_index=model_options.device,
in_channels=model_options.in_channels,
out_channels=model_options.out_channels,
f_maps=model_options.f_maps,
layer_order=model_options.layer_order,
num_groups=model_options.num_groups,
num_levels=model_options.num_levels)
elif model_options.model == 'ResidualUNet3D_add_more_fc':
print(
'********** init ResidualUNet3D_add_more_fc model for test! **********'
)
model = create_model(model_options.model,
gpu=model_options.gpu,
dropout=model_options.dropout,
device_index=model_options.device,
in_channels=model_options.in_channels,
out_channels=model_options.out_channels,
f_maps=model_options.f_maps,
layer_order=model_options.layer_order,
num_groups=model_options.num_groups,
num_levels=model_options.num_levels)
elif model_options.model == 'VoxCNN':
print('********** init VoxCNN model for test! **********')
model = create_model(model_options.model,
gpu=model_options.gpu,
device_index=model_options.device)
elif model_options.model == 'ConvNet3D':
print('********** init ConvNet3D model for test! **********')
model = create_model(model_options.model,
gpu=model_options.gpu,
device_index=model_options.device)
elif 'gcn' in model_options.model:
print('********** init {}-{} model for test! **********'.format(
model_options.model, model_options.gnn_type))
model = create_model(
model_options.model,
gpu=model_options.gpu,
device_index=model_options.device,
gnn_type=model_options.gnn_type,
gnn_dropout=model_options.gnn_dropout,
gnn_dropout_adj=model_options.gnn_dropout_adj,
gnn_non_linear=model_options.gnn_non_linear,
gnn_undirected=model_options.gnn_undirected,
gnn_self_loop=model_options.gnn_self_loop,
gnn_threshold=model_options.gnn_threshold,
)
elif model_options.model == 'ROI_GCN':
print('********** init ROI_GCN model for test! **********')
model = create_model(
model_options.model,
gpu=model_options.gpu,
device_index=model_options.device,
gnn_type=model_options.gnn_type,
gnn_dropout=model_options.gnn_dropout,
gnn_dropout_adj=model_options.gnn_dropout_adj,
gnn_non_linear=model_options.gnn_non_linear,
gnn_undirected=model_options.gnn_undirected,
gnn_self_loop=model_options.gnn_self_loop,
gnn_threshold=model_options.gnn_threshold,
nodel_vetor_layer=model_options.nodel_vetor_layer,
classify_layer=model_options.classify_layer,
num_node_features=model_options.num_node_features,
num_class=model_options.num_class,
roi_size=model_options.roi_size,
num_nodes=model_options.num_nodes,
gnn_pooling_layers=model_options.gnn_pooling_layers,
global_sort_pool_k=model_options.global_sort_pool_k,
layers=model_options.layers,
shortcut_type=model_options.shortcut_type,
use_nl=model_options.use_nl,
dropout=model_options.dropout,
device=model_options.device)
elif model_options.model == 'SwinTransformer3d':
print('********** init SwinTransformer3d model for test! **********')
model = create_model(
model_options.model,
gpu=model_options.gpu,
dropout=model_options.dropout,
device_index=model_options.device,
sw_patch_size=model_options.sw_patch_size,
window_size=model_options.window_size,
mlp_ratio=model_options.mlp_ratio,
drop_rate=model_options.drop_rate,
attn_drop_rate=model_options.attn_drop_rate,
drop_path_rate=model_options.drop_path_rate,
qk_scale=model_options.qk_scale,
embed_dim=model_options.embed_dim,
depths=model_options.depths,
num_heads=model_options.num_heads,
qkv_bias=model_options.qkv_bias,
ape=model_options.ape,
patch_norm=model_options.patch_norm,
)
else:
model = create_model(
model_options.model,
gpu=model_options.gpu,
dropout=model_options.dropout,
device_index=model_options.device,
)
transformations = get_transforms(model_options.mode,
model_options.minmaxnormalization)
# Define loss and optimizer
criterion = nn.CrossEntropyLoss()
if model_options.mode_task == 'multicnn':
predictions_df = pd.DataFrame()
metrics_df = pd.DataFrame()
for n in range(num_cnn):
dataset = return_dataset(model_options.mode,
caps_dir,
tsv_path,
model_options.preprocessing,
transformations,
model_options,
cnn_index=n)
print('test data size:{}'.format(len(dataset)))
test_loader = DataLoader(dataset,
batch_size=model_options.batch_size,
shuffle=False,
num_workers=model_options.nproc,
pin_memory=True,
drop_last=model_options.drop_last)
# load the best trained model during the training
model, best_epoch = load_model(model,
join(model_path, 'cnn-%i' % n,
selection),
gpu,
filename='model_best.pth.tar',
device_index=model_options.device)
cnn_df, cnn_metrics = test(model,
test_loader,
gpu,
criterion,
mode=model_options.mode,
device_index=model_options.device)
predictions_df = pd.concat([predictions_df, cnn_df])
metrics_df = pd.concat(
[metrics_df, pd.DataFrame(cnn_metrics, index=[0])])
predictions_df.reset_index(drop=True, inplace=True)
metrics_df.reset_index(drop=True, inplace=True)
else:
# Load model from path
best_model, best_epoch = load_model(model,
join(model_path, selection),
gpu,
filename='model_best.pth.tar',
device_index=model_options.device)
# Read/localize the data
data_to_test = return_dataset(model_options.mode, caps_dir, tsv_path,
model_options.preprocessing,
transformations, model_options)
print('test data size:{}'.format(len(data_to_test)))
# Load the data
test_loader = DataLoader(data_to_test,
batch_size=model_options.batch_size,
shuffle=False,
num_workers=model_options.nproc,
pin_memory=True,
drop_last=model_options.drop_last)
# Run the model on the data
predictions_df, metrics = test(best_model,
test_loader,
gpu,
criterion,
mode=model_options.mode,
device_index=model_options.device,
model_options=model_options)
metrics_df = pd.DataFrame(metrics, index=[0])
return predictions_df, metrics_df
def individual_backprop(options):
main_logger = return_logger(options.verbose, "main process")
options = translate_parameters(options)
fold_list = [
fold for fold in os.listdir(options.model_path) if fold[:5:] == "fold-"
]
if len(fold_list) == 0:
raise ValueError("No folds were found at path %s" % options.model_path)
model_options = argparse.Namespace()
model_options = read_json(
model_options, path.join(options.model_path, 'commandline.json'))
model_options = translate_parameters(model_options)
model_options.gpu = options.gpu
if model_options.network_type == "multicnn":
raise NotImplementedError(
"The interpretation of multi-CNN is not implemented.")
if options.tsv_path is None and options.input_dir is None:
options.multi_cohort = model_options.multi_cohort
if options.tsv_path is None:
options.tsv_path = model_options.tsv_path
if options.input_dir is None:
options.input_dir = model_options.input_dir
if options.target_diagnosis is None:
options.target_diagnosis = options.diagnosis
for fold in fold_list:
main_logger.info(fold)
for selection in options.selection:
results_path = path.join(options.model_path, fold, 'gradients',
selection, options.name)
criterion = get_criterion(model_options.loss)
# Data management (remove data not well predicted by the CNN)
training_df = load_data_test(options.tsv_path, [options.diagnosis],
baseline=options.baseline,
multi_cohort=options.multi_cohort)
training_df.reset_index(drop=True, inplace=True)
# Model creation
_, all_transforms = get_transforms(
model_options.mode,
minmaxnormalization=model_options.minmaxnormalization)
with warnings.catch_warnings():
warnings.simplefilter("ignore")
data_example = return_dataset(
model_options.mode,
options.input_dir,
training_df,
model_options.preprocessing,
train_transformations=None,
all_transformations=all_transforms,
prepare_dl=options.prepare_dl,
multi_cohort=options.multi_cohort,
params=model_options)
model = create_model(model_options, data_example.size)
model_dir = os.path.join(options.model_path, fold, 'models',
selection)
model, best_epoch = load_model(model,
model_dir,
gpu=options.gpu,
filename='model_best.pth.tar')
options.output_dir = results_path
commandline_to_json(options, logger=main_logger)
# Keep only subjects who were correctly / wrongly predicted by the network
training_df = sort_predicted(model,
training_df,
options.input_dir,
model_options,
criterion,
options.keep_true,
batch_size=options.batch_size,
num_workers=options.num_workers,
gpu=options.gpu)
if len(training_df) > 0:
# Save the tsv files used for the saliency maps
training_df.to_csv(path.join('data.tsv'),
sep='\t',
index=False)
with warnings.catch_warnings():
warnings.simplefilter("ignore")
data_train = return_dataset(
model_options.mode,
options.input_dir,
training_df,
model_options.preprocessing,
train_transformations=None,
all_transformations=all_transforms,
prepare_dl=options.prepare_dl,
multi_cohort=options.multi_cohort,
params=model_options)
train_loader = DataLoader(data_train,
batch_size=options.batch_size,
shuffle=True,
num_workers=options.num_workers,
pin_memory=True)
interpreter = VanillaBackProp(model, gpu=options.gpu)
for data in train_loader:
if options.gpu:
input_batch = data['image'].cuda()
else:
input_batch = data['image']
map_np = interpreter.generate_gradients(
input_batch,
data_train.diagnosis_code[options.target_diagnosis])
for i in range(options.batch_size):
single_path = path.join(results_path,
data['participant_id'][i],
data['session_id'][i])
os.makedirs(single_path, exist_ok=True)
if len(data_train.size) == 4:
if options.nifti_template_path is not None:
image_nii = nib.load(
options.nifti_template_path)
affine = image_nii.affine
else:
affine = np.eye(4)
map_nii = nib.Nifti1Image(map_np[i, 0, :, :, :],
affine)
nib.save(map_nii,
path.join(single_path, "map.nii.gz"))
else:
jpg_path = path.join(single_path, "map.jpg")
plt.imshow(map_np[i, 0, :, :],
cmap="coolwarm",
vmin=-options.vmax,
vmax=options.vmax)
plt.colorbar()
plt.savefig(jpg_path)
plt.close()
np.save(path.join(single_path, "map.npy"), map_np[i])
