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)
def test_cnn(output_dir,
data_loader,
subset_name,
split,
criterion,
model_options,
gpu=False,
multiclass=False):
for selection in ["best_balanced_accuracy", "best_loss"]:
# load the best trained model during the training
model = create_model(model_options.model,
gpu,
dropout=model_options.dropout)
model, best_epoch = load_model(model,
os.path.join(output_dir,
'fold-%i' % split,
'models', selection),
gpu=gpu,
filename='model_best.pth.tar')
results_df, metrics = test(model, data_loader, gpu, criterion,
model_options.mode, multiclass)
print("%s level balanced accuracy is %f" %
(model_options.mode, metrics['balanced_accuracy']))
mode_level_to_tsvs(output_dir,
results_df,
metrics,
split,
selection,
model_options.mode,
dataset=subset_name)
# Soft voting
if model_options.mode in ["patch", "roi", "slice"]:
soft_voting_to_tsvs(
output_dir,
split,
selection=selection,
mode=model_options.mode,
dataset=subset_name,
selection_threshold=model_options.selection_threshold)
model_options.preprocessing,
transformations,
model_options,
cnn_index=cnn_index)
train_loader = DataLoader(data_train,
batch_size=options.batch_size,
shuffle=False,
num_workers=options.num_workers,
pin_memory=True)
valid_loader = DataLoader(data_valid,
batch_size=options.batch_size,
shuffle=False,
num_workers=options.num_workers,
pin_memory=True)
test_cnn(options.model_path, train_loader, "train", split,
criterion, cnn_index, model_options, options.gpu)
test_cnn(options.model_path, valid_loader, "validation", 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)
def inference_from_model(caps_dir,
tsv_path,
model_path=None,
json_file=None,
prefix=None,
no_labels=False,
gpu=True,
prepare_dl=False):
"""
Inference from previously trained model.
This functions uses a previously trained model to classify the input(s).
The model is stored in the variable model_path and it assumes the folder
structure given by the training stage. Particullary to have a prediction at
image level, it assumes that results of the validation set are stored in
the model_path folder in order to perform soft-voiting at the slice/patch
level and also for multicnn.
Args:
caps_dir: folder containing the tensor files (.pt version of MRI)
tsv_path: file with the name of the MRIs to process (single or multiple)
model_path: file with the model (pth format).
json_file: file containing the training parameters.
output_dir_arg: folder where results are stored. If None it uses current
structure.
no_labels: by default is false. In that case, output writes a file named
measurements.tsv
gpu: if true, it uses gpu.
prepare_dl: if true, uses extracted patches/slices otherwise extract them
on-the-fly.
Returns:
Files written in the output folder with prediction results and metrics. By
default the output folder is named cnn_classification and it is inside the
model_folder.
Raises:
"""
import argparse
parser = argparse.ArgumentParser()
parser.add_argument("model_path",
type=str,
help="Path to the trained model folder.")
options = parser.parse_args([model_path])
options = read_json(options, json_path=json_file)
num_cnn = compute_num_cnn(caps_dir, tsv_path, options, "classify")
print("Load model with these options:")
print(options)
# Overwrite options with user input
options.use_cpu = not gpu
options.prepare_dl = prepare_dl
# Define the path
currentDirectory = pathlib.Path(model_path)
# Search for 'fold-*' pattern
currentPattern = "fold-*"
best_model = {
'best_acc': 'best_balanced_accuracy',
'best_loss': 'best_loss'
}
# loop depending the number of folds found in the model folder
for fold_dir in currentDirectory.glob(currentPattern):
fold = int(str(fold_dir).split("-")[-1])
fold_path = join(model_path, fold_dir)
model_path = join(fold_path, 'models')
if options.mode_task == 'multicnn':
for cnn_dir in listdir(model_path):
if not exists(
join(model_path, cnn_dir, best_model['best_acc'],
'model_best.pth.tar')):
raise FileNotFoundError(
errno.ENOENT, strerror(errno.ENOENT),
join(model_path, cnn_dir, best_model['best_acc'],
'model_best.pth.tar'))
else:
full_model_path = join(model_path, best_model['best_acc'])
if not exists(join(full_model_path, 'model_best.pth.tar')):
raise FileNotFoundError(
errno.ENOENT, strerror(errno.ENOENT),
join(full_model_path, 'model_best.pth.tar'))
performance_dir = join(fold_path, 'cnn_classification',
best_model['best_acc'])
if not exists(performance_dir):
makedirs(performance_dir)
# It launch the corresponding function, depending on the mode.
infered_classes, metrics = inference_from_model_generic(
caps_dir, tsv_path, model_path, options, num_cnn=num_cnn)
# Prepare outputs
usr_prefix = str(prefix)
# Write output files at %mode level
print("Prediction results and metrics are written in the "
"following folder: %s" % performance_dir)
mode_level_to_tsvs(currentDirectory,
infered_classes,
metrics,
fold,
best_model['best_acc'],
options.mode,
dataset=usr_prefix)
# Soft voting
if hasattr(options, 'selection_threshold'):
selection_thresh = options.selection_threshold
else:
selection_thresh = 0.8
# Write files at the image level (for patch, roi and slice).
# It assumes the existance of validation files to perform soft-voting
if options.mode in ["patch", "roi", "slice"]:
soft_voting_to_tsvs(currentDirectory,
fold,
best_model["best_acc"],
options.mode,
usr_prefix,
num_cnn=num_cnn,
selection_threshold=selection_thresh)
def inference_from_model(caps_dir,
tsv_path,
model_path=None,
json_file=None,
prefix=None,
labels=True,
gpu=True,
num_workers=0,
batch_size=1,
prepare_dl=False,
selection_metrics=None,
diagnoses=None,
logger=None):
"""
Inference from previously trained model.
This functions uses a previously trained model to classify the input(s).
The model is stored in the variable model_path and it assumes the folder
structure given by the training stage. Particullary to have a prediction at
image level, it assumes that results of the validation set are stored in
the model_path folder in order to perform soft-voiting at the slice/patch
level and also for multicnn.
Args:
caps_dir: folder containing the tensor files (.pt version of MRI)
tsv_path: file with the name of the MRIs to process (single or multiple)
model_path: file with the model (pth format).
json_file: file containing the training parameters.
prefix: prefix of all classification outputs.
labels: by default is True. If False no metrics tsv files will be written.
measurements.tsv
gpu: if true, it uses gpu.
num_workers: num_workers used in DataLoader
batch_size: batch size of the DataLoader
prepare_dl: if true, uses extracted patches/slices otherwise extract them
on-the-fly.
selection_metrics: list of metrics to find best models to be evaluated.
diagnoses: list of diagnoses to be tested if tsv_path is a folder.
logger: Logger instance.
Returns:
Files written in the output folder with prediction results and metrics. By
default the output folder is named cnn_classification and it is inside the
model_folder.
Raises:
"""
import argparse
import logging
if logger is None:
logger = logging
parser = argparse.ArgumentParser()
parser.add_argument("model_path",
type=str,
help="Path to the trained model folder.")
options = parser.parse_args([model_path])
options = read_json(options, json_path=json_file)
logger.debug("Load model with these options:")
logger.debug(options)
# Overwrite options with user input
options.use_cpu = not gpu
options.nproc = num_workers
options.batch_size = batch_size
options.prepare_dl = prepare_dl
if diagnoses is not None:
options.diagnoses = diagnoses
options = translate_parameters(options)
if options.mode_task == "multicnn":
num_cnn = compute_num_cnn(caps_dir, tsv_path, options, "test")
else:
num_cnn = None
# Define the path
currentDirectory = pathlib.Path(model_path)
# Search for 'fold-*' pattern
currentPattern = "fold-*"
# loop depending the number of folds found in the model folder
for fold_dir in currentDirectory.glob(currentPattern):
fold = int(str(fold_dir).split("-")[-1])
fold_path = join(model_path, fold_dir)
model_path = join(fold_path, 'models')
for selection_metric in selection_metrics:
if options.mode_task == 'multicnn':
for cnn_dir in listdir(model_path):
if not exists(
join(model_path, cnn_dir, "best_%s" %
selection_metric, 'model_best.pth.tar')):
raise FileNotFoundError(
errno.ENOENT, strerror(errno.ENOENT),
join(model_path, cnn_dir,
"best_%s" % selection_metric,
'model_best.pth.tar'))
else:
full_model_path = join(model_path,
"best_%s" % selection_metric)
if not exists(join(full_model_path, 'model_best.pth.tar')):
raise FileNotFoundError(
errno.ENOENT, strerror(errno.ENOENT),
join(full_model_path, 'model_best.pth.tar'))
performance_dir = join(fold_path, 'cnn_classification',
'best_%s' % selection_metric)
makedirs(performance_dir, exist_ok=True)
# It launch the corresponding function, depending on the mode.
inference_from_model_generic(caps_dir,
tsv_path,
model_path,
options,
prefix,
currentDirectory,
fold,
"best_%s" % selection_metric,
labels=labels,
num_cnn=num_cnn,
logger=logger)
# Soft voting
if hasattr(options, 'selection_threshold'):
selection_thresh = options.selection_threshold
else:
selection_thresh = 0.8
# Write files at the image level (for patch, roi and slice).
# It assumes the existance of validation files to perform soft-voting
if options.mode in ["patch", "roi", "slice"]:
soft_voting_to_tsvs(currentDirectory,
fold,
"best_%s" % selection_metric,
options.mode,
prefix,
num_cnn=num_cnn,
selection_threshold=selection_thresh,
use_labels=labels,
logger=logger)
logger.info("Prediction results and metrics are written in the "
"following folder: %s" % performance_dir)
def test_cnn(output_dir,
data_loader,
subset_name,
split,
criterion,
model_options,
gpu=False,
train_begin_time=None):
metric_dict = {}
for selection in ["best_balanced_accuracy", "best_loss"]:
# load the best trained model during the training
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:
print('********** init model for test! **********')
model = create_model(model_options.model,
gpu=model_options.gpu,
dropout=model_options.dropout,
device_index=model_options.device)
model, best_epoch = load_model(model,
os.path.join(output_dir,
'fold-%i' % split,
'models', selection),
gpu=gpu,
filename='model_best.pth.tar',
device_index=model_options.device)
results_df, metrics = test(model,
data_loader,
gpu,
criterion,
model_options.mode,
device_index=model_options.device,
train_begin_time=train_begin_time,
model_options=model_options,
fi=split)
print("[%s]: %s level balanced accuracy is %f" %
(timeSince(train_begin_time), model_options.mode,
metrics['balanced_accuracy']))
print('[{}]: {}_{}_result_df:'.format(timeSince(train_begin_time),
subset_name, selection))
print(results_df)
print('[{}]: {}_{}_metrics:\n{}'.format(timeSince(train_begin_time),
subset_name, selection,
metrics))
wandb.log({
'{}_accuracy_{}_singel_model'.format(subset_name, selection):
metrics['accuracy'],
'{}_balanced_accuracy_{}_singel_model'.format(
subset_name, selection):
metrics['balanced_accuracy'],
'{}_sensitivity_{}_singel_model'.format(subset_name, selection):
metrics['sensitivity'],
'{}_specificity_{}_singel_model'.format(subset_name, selection):
metrics['specificity'],
'{}_ppv_{}_singel_model'.format(subset_name, selection):
metrics['ppv'],
'{}_npv_{}_singel_model'.format(subset_name, selection):
metrics['npv'],
'{}_total_loss_{}_singel_model'.format(subset_name, selection):
metrics['total_loss'],
})
mode_level_to_tsvs(output_dir,
results_df,
metrics,
split,
selection,
model_options.mode,
dataset=subset_name)
# Soft voting
if model_options.mode in ["patch", "roi", "slice"]:
soft_voting_to_tsvs(
output_dir,
split,
selection=selection,
mode=model_options.mode,
dataset=subset_name,
selection_threshold=model_options.selection_threshold)
# return metric dict
metric_temp_dict = {
'{}_accuracy_{}_singel_model'.format(subset_name, selection):
metrics['accuracy'],
'{}_balanced_accuracy_{}_singel_model'.format(
subset_name, selection):
metrics['balanced_accuracy'],
'{}_sensitivity_{}_singel_model'.format(subset_name, selection):
metrics['sensitivity'],
'{}_specificity_{}_singel_model'.format(subset_name, selection):
metrics['specificity'],
'{}_ppv_{}_singel_model'.format(subset_name, selection):
metrics['ppv'],
'{}_npv_{}_singel_model'.format(subset_name, selection):
metrics['npv'],
'{}_total_loss_{}_singel_model'.format(subset_name, selection):
metrics['total_loss'],
}
metric_dict.update(metric_temp_dict)
return metric_dict