--- a
+++ b/src/ensemble/mean_ensemble.py
@@ -0,0 +1,97 @@
+import os
+import sys
+import torch
+import numpy as np
+
+from src.config import BratsConfiguration
+from src.dataset import brats_labels
+from src.dataset.utils import dataset
+from src.inference import crop_no_patch, return_to_size
+from src.models.io_model import load_model
+from src.models.unet3d import unet3d
+from src.models.vnet import asymm_vnet
+from src.test import predict
+
+
+def _load(network, model_path, checkpoint):
+    network.to(device)
+    checkpoint_path = os.path.join(model_path, checkpoint)
+    model, _, _, _ = load_model(network, checkpoint_path, device, None, False)
+    return model
+
+
+def load_model_1598550861(model_path):
+    check = "model_1598550861/checkpoint_epoch_215_val_loss_0.2378825504485875_dice_0.7621174487349105.pth"
+    network = asymm_vnet.VNet(non_linearity="prelu", in_channels=4,  classes=4, init_features_maps=32,
+                              kernel_size=5, padding=2)
+    return _load(network, model_path, check)
+
+
+def load_model_1598639885(model_path):
+    check = "model_1598639885/checkpoint_epoch_198_val_loss_0.19342842820572526_dice_0.8065715717942747.pth"
+    network = unet3d.ResidualUNet3D(in_channels=4, out_channels=4, final_sigmoid=False, f_maps=32,
+                                    layer_order="crg", num_levels=4, num_groups=4, conv_padding=1)
+    return _load(network, model_path, check)
+
+
+def load_model_1598640035(model_path):
+    check = "model_1598640035/checkpoint_epoch_142_val_loss_0.21437616135976087_dice_0.7856238380039416.pth"
+    network = unet3d.ResidualUNet3D(in_channels=4, out_channels=4, final_sigmoid=False,
+                                    f_maps=32, layer_order="crg", num_levels=4, num_groups=4, conv_padding=1)
+    return _load(network, model_path, check)
+
+
+def load_model_1598640005(model_path):
+    check = "model_1598640005/checkpoint_epoch_168_val_loss_0.20105469390137554_dice_0.7989453060986245.pth"
+    network = unet3d.UNet3D(in_channels=4, out_channels=4, final_sigmoid=False, f_maps=32,
+                            layer_order="crg", num_levels=4, num_groups=4, conv_padding=1)
+
+    return _load(network, model_path, check)
+
+
+if __name__ == "__main__":
+
+    config = BratsConfiguration(sys.argv[1])
+    model_config = config.get_model_config()
+    dataset_config = config.get_dataset_config()
+    basic_config = config.get_basic_config()
+
+    device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")
+    sampling = dataset_config.get("sampling_method").split(".")[-1]
+
+    models_gen_path = model_config.get("model_path")
+    task = "ensemble_segmentation"
+    compute_metrics = False
+
+    model_vnet = load_model_1598550861(models_gen_path)
+    model_2 = load_model_1598639885(models_gen_path)
+    model_3 = load_model_1598640035(models_gen_path)
+    model_4 = load_model_1598640005(models_gen_path)
+
+    data, _ = dataset.read_brats(dataset_config.get("val_csv"))
+
+    for idx in range(0, len(data)):
+        results = {}
+
+        images = data[idx].load_mri_volumes(normalize=True)
+
+        x_1, x_2, y_1, y_2, z_1, z_2, images, brain_mask, patch_size = crop_no_patch(data[idx].size, images,
+                                                                                     data[idx].get_brain_mask(),
+                                                                                     sampling)
+
+        _, prediction_four_channels_1 = predict.predict(model_vnet, images, device, False)
+        _, prediction_four_channels_2 = predict.predict(model_2, images, device, False)
+        _, prediction_four_channels_3 = predict.predict(model_3, images, device, False)
+        _, prediction_four_channels_4 = predict.predict(model_4, images, device, False)
+
+        prediction_four_channels_1 = prediction_four_channels_1.view((4, patch_size[0], patch_size[1], patch_size[2]))
+
+        pred_scores = torch.stack([prediction_four_channels_1, prediction_four_channels_2[0], prediction_four_channels_3[0], prediction_four_channels_4[0]]).cpu().numpy()
+        pred_scores_mean = np.mean(pred_scores, axis=0)
+        prediction_map = np.argmax(pred_scores_mean, axis=0)
+
+        prediction_map = brats_labels.convert_to_brats_labels(prediction_map)
+        prediction_map = return_to_size(prediction_map, sampling, x_1, x_2, y_1, y_2, z_1, z_2)
+
+        results["prediction"] = prediction_map
+        predict.save_predictions(data[idx], results, models_gen_path, task)