--- a
+++ b/tools/ensemble_inferencer.py
@@ -0,0 +1,76 @@
+import argparse
+import os
+import shutil
+import warnings
+import numpy as np
+from PIL import Image
+
+import mmcv
+import torch
+from mmcv.parallel import MMDataParallel, MMDistributedDataParallel
+from mmcv.runner import (get_dist_info, init_dist, load_checkpoint,
+                         wrap_fp16_model)
+from mmcv.utils import DictAction
+
+from mmseg.apis import multi_gpu_test, single_gpu_test
+from mmseg.datasets import build_dataloader, build_dataset
+from mmseg.models import build_segmentor
+
+configs = [
+    "../work_dirs/remote/swb384_22k_1x_16bs_all/remote_swb.py",
+    "../work_dirs/remote/dl3pr101_1x_16bs_all/remote_dl3pr101.py"
+]
+ckpts = [
+    "../work_dirs/remote/swb384_22k_1x_16bs_all/latest.pth",
+    "../work_dirs/remote/dl3pr101_1x_16bs_all/latest.pth"
+]
+
+cfg = mmcv.Config.fromfile(configs[0])
+torch.backends.cudnn.benchmark = True
+cfg.model.pretrained = None
+cfg.data.test.test_mode = True
+dataset = build_dataset(cfg.data.test)
+data_loader = build_dataloader(
+    dataset,
+    samples_per_gpu=64,
+    workers_per_gpu=4, #cfg.data.workers_per_gpu,
+    dist=False,
+    shuffle=False)
+
+models = []
+for config, ckpt in zip(configs, ckpts):
+    cfg = mmcv.Config.fromfile(config)
+    torch.backends.cudnn.benchmark = True
+    cfg.model.pretrained = None
+    cfg.data.test.test_mode = True
+    cfg.model.train_cfg = None
+    model = build_segmentor(cfg.model, test_cfg=cfg.get('test_cfg'))
+    fp16_cfg = cfg.get('fp16', None)
+    wrap_fp16_model(model)
+    checkpoint = load_checkpoint(model, ckpt, map_location='cpu')
+    model.CLASSES = dataset.CLASSES
+    model.PALETTE = dataset.PALETTE
+    torch.cuda.empty_cache()
+    eval_kwargs = {"imgfile_prefix": "../work_imgs"}
+    tmpdir = eval_kwargs['imgfile_prefix']
+    mmcv.mkdir_or_exist(tmpdir)
+    model = MMDataParallel(model, device_ids=[0])
+    model.eval()
+    models.append(model)
+
+results = []
+dataset = data_loader.dataset
+prog_bar = mmcv.ProgressBar(len(dataset))
+loader_indices = data_loader.batch_sampler
+
+for batch_indices, data in zip(loader_indices, data_loader):
+    result = []
+    for model in models:
+        with torch.no_grad():
+            result.append(model(return_loss=False, **data))
+    result = [np.stack(_, 0).sum(0).argmax(0) for _ in zip(*result)]
+    for res, batch_index in zip(result, batch_indices):
+        img_info = dataset.img_infos[batch_index]
+        file_name = os.path.join(tmpdir, img_info['ann']['seg_map'])
+        Image.fromarray(res.astype(np.uint8)).save(file_name)
+        prog_bar.update()
\ No newline at end of file