Thanks for your answer. Sorry, I wasn’t clear. I presented a part of my script which doesn’t include the joblib call, on purpose. Here is how I tried to call it:
def fit_and_evaluate_decoder(
z_maps, mask_img, conditions_label, run_label,
screening_percentile=5, C=1.0, n_leave_out=1, n_jobs_cv=1
):
"""Fit and evaluate a decoder with feature screening and linear SVM.
Parameters
----------
z_maps : array
List or array of z-maps per condition/run.
mask_img : Niimg-like
Brain mask.
conditions_label : array
Condition labels (y).
run_label : array
Run IDs (used as groups in CV).
screening_percentile : int
Percentile of most informative voxels selected.
C : float
Regularization strength for SVC.
n_leave_out : int
Number of groups to leave out during cross-validation.
n_jobs_cv : int
Number of parallel jobs for CV (-1 = all cores).
Returns
-------
mean_accuracy : float
chance_level : float
accuracy_scores : list of float
cm : ndarray
best_model : sklearn.pipeline.Pipeline
Last model from CV (not retrained on full set).
"""
print(f"\nFitting scikit-learn decoder with L{n_leave_out}O, screening={screening_percentile}%, C={C}...")
z_maps = np.array(z_maps)
conditions_label = np.array(conditions_label)
run_label = np.array(run_label)
z_maps_reshaped = z_maps.reshape(-1, z_maps.shape[-1])
masked_data = apply_mask(z_maps_reshaped, mask_img)
logo = LeavePGroupsOut(n_groups=n_leave_out)
cv_splits = list(logo.split(masked_data, conditions_label, groups=run_label))
def fit_and_predict(train_idx, test_idx):
model = make_pipeline(
StandardScaler(),
SelectPercentile(f_classif, percentile=screening_percentile),
SVC(kernel='linear', C=C)
)
model.fit(masked_data[train_idx], conditions_label[train_idx])
y_pred = model.predict(masked_data[test_idx])
y_true = conditions_label[test_idx]
score = accuracy_score(y_true, y_pred)
return y_true, y_pred, score, model
with parallel_backend("loky"):
results = Parallel(n_jobs=n_jobs_cv)(
delayed(fit_and_predict)(train_idx, test_idx)
for train_idx, test_idx in cv_splits
)
all_y_true, all_y_pred, accuracy_scores = [], [], []
best_model = None
for y_true, y_pred, score, model in results:
all_y_true.extend(y_true)
all_y_pred.extend(y_pred)
accuracy_scores.append(score)
best_model = model # last one
mean_accuracy = np.mean(accuracy_scores)
std_accuracy = np.std(accuracy_scores)
chance_level = 1.0 / len(np.unique(conditions_label))
cm = confusion_matrix(all_y_true, all_y_pred)
print(f" - Accuracy: {mean_accuracy:.4f} ± {std_accuracy:.4f}")
print(f" - Chance level: {chance_level:.4f}")
print(f" - Confusion Matrix:\n{cm}")
return mean_accuracy, chance_level, accuracy_scores, cm, best_model
(…)
def main():
visual_events_all = {}
conditions_all = {}
bold_images_all = {}
contrasts_all = {}
decoding_results = {}
weight_maps_all = {}
roi_masks = {}
for roi_name in roi_names:
try:
mask_img = find_nii_or_niigz(os.path.join(roi_dir, roi_name))
roi_masks[roi_name] = mask_img
except FileNotFoundError as e:
print(f"ROI not found: {roi_name}\n {e}")
for subject_id in subject_list:
print(f"\n{'-'*50}\nSubject: {subject_id}")
bids_func_dir = os.path.join(bids_dir, f"sub-{subject_id}", "func")
fmriprep_func_dir = os.path.join(fmriprep_dir, f"sub-{subject_id}", "func")
if not os.path.isdir(bids_func_dir):
print(f"Skipping sub-{subject_id}, directory not found: {bids_func_dir}")
continue
visual_events, conditions = load_visual_events_by_run(bids_func_dir, subject_id, task_name)
bold_images = load_bold_by_run(fmriprep_func_dir, subject_id, task_name)
visual_events_all[subject_id] = visual_events
conditions_all[subject_id] = conditions
bold_images_all[subject_id] = bold_images
contrasts_all[subject_id] = {}
for roi_name, mask_img in roi_masks.items():
print(f"\nComputing contrasts for sub-{subject_id}, ROI={roi_name}...")
try:
z_maps, cond_labels, run_labels = compute_contrasts_by_run(
subject_id, roi_name, mask_img, bold_images, visual_events, unique_runs
)
contrasts_all[subject_id][roi_name] = {
'z_maps': z_maps,
'condition_labels': cond_labels,
'run_labels': run_labels,
}
accuracy, chance, accuracy_scores, cm, model = fit_and_evaluate_decoder(
z_maps, mask_img, cond_labels, run_labels,
screening_percentile=screening_percentile,
C=C, n_leave_out=n_leave_out, n_jobs_cv=n_jobs_decoder
)
decoding_results[(subject_id, roi_name)] = {
'accuracy': accuracy,
'chance_level': chance,
'accuracy_scores': accuracy_scores,
'confusion_matrix': cm,
'conditions': np.unique(cond_labels).tolist()
}
if save_weight_maps:
try:
output_prefix = os.path.join(output_dir, f"sub-{subject_id}_roi-{roi_name}")
save_decoder_weights(model, mask_img, output_prefix, class_labels=np.unique(cond_labels).tolist())
except Exception as e:
print(f"Failed to save weights for sub-{subject_id}, ROI={roi_name}\n {e}")
del z_maps, cond_labels, run_labels, model, cm
gc.collect()
except Exception as e:
print(f"Failed for sub-{subject_id}, ROI={roi_name}\n {e}")
del visual_events, conditions, bold_images
gc.collect()
return (
visual_events_all,
conditions_all,
bold_images_all,
contrasts_all,
decoding_results,
weight_maps_all
)
And here is some print I obtain. From other scripts, this specific part seems very verbose but fine, and a mark that n_jobs is called:
Exception ignored in: <function ResourceTracker.del at 0x7fa1dd5c39c0>
Traceback (most recent call last):
File “/home/touhara/anaconda3/envs/nilearn/lib/python3.13/multiprocessing/resource_tracker.py”, line 82, in del
File “/home/touhara/anaconda3/envs/nilearn/lib/python3.13/multiprocessing/resource_tracker.py”, line 91, in _stop
File “/home/touhara/anaconda3/envs/nilearn/lib/python3.13/multiprocessing/resource_tracker.py”, line 116, in _stop_locked
ChildProcessError: [Errno 10] No child processes
But my Activity Monitor shows that only 1 core is used 100% at a time.
Thank you.
Matthieu