Merge pull request #749 from yuyu2172/simplify-camvid

Unify evaluation for semantic segmentation

chainercv/links/model/segnet/segnet_basic.py

@@ -52,7 +52,7 @@ class SegNetBasic(chainer.Chain):
         'camvid': {
             'param': {'n_class': 11},
             'url': 'https://chainercv-models.preferred.jp/'
-            'segnet_camvid_trained_2017_05_28.npz'
+            'segnet_camvid_trained_2018_12_05.npz'
         }
     }
 

examples/segnet/README.md

@@ -53,7 +53,7 @@ $ python demo.py [--gpu <gpu>] [--pretrained-model <model_path>] 0001TP_008550.p
 | ChainerCV      | 82.7 %          | **67.1 %**     | **49.4 %** |
 | Official       | **82.8 %**      | 62.3%          | 46.3 %     |
 
-The evaluation can be conducted using [`chainercv/examples/semantic_segmentation/eval_cityscapes.py`](https://github.com/chainer/chainercv/blob/master/examples/semantic_segmentation).
+The evaluation can be conducted using [`chainercv/examples/semantic_segmentation/eval_semantic_segmentation.py`](https://github.com/chainer/chainercv/blob/master/examples/semantic_segmentation).
 
 
 # Reference

examples/segnet/train.py

@@ -1,8 +1,11 @@
 import argparse
+from collections import defaultdict
+import os
 
 import chainer
 import numpy as np
 
+from chainer.dataset import concat_examples
 from chainer.datasets import TransformDataset
 from chainer import iterators
 from chainer import optimizers
@@ -16,6 +19,31 @@
 from chainercv.links import SegNetBasic
 
 
+def recalculate_bn_statistics(model, batchsize):
+    train = CamVidDataset(split='train')
+    it = chainer.iterators.SerialIterator(
+        train, batchsize, repeat=False, shuffle=False)
+    bn_avg_mean = defaultdict(np.float32)
+    bn_avg_var = defaultdict(np.float32)
+
+    n_iter = 0
+    for batch in it:
+        imgs, _ = concat_examples(batch)
+        model(model.xp.array(imgs))
+        for name, link in model.namedlinks():
+            if name.endswith('_bn'):
+                bn_avg_mean[name] += link.avg_mean
+                bn_avg_var[name] += link.avg_var
+        n_iter += 1
+
+    for name, link in model.namedlinks():
+        if name.endswith('_bn'):
+            link.avg_mean = bn_avg_mean[name] / n_iter
+            link.avg_var = bn_avg_var[name] / n_iter
+
+    return model
+
+
 def transform(in_data):
     img, label = in_data
     if np.random.rand() > 0.5:
@@ -81,9 +109,6 @@ def main():
             ['validation/main/miou'], x_key='iteration',
             file_name='miou.png'))
 
-    trainer.extend(extensions.snapshot_object(
-        model.predictor, filename='model_iteration-{.updater.iteration}'),
-        trigger=end_trigger)
     trainer.extend(extensions.PrintReport(
         ['epoch', 'iteration', 'elapsed_time', 'lr',
          'main/loss', 'validation/main/miou',
@@ -100,6 +125,10 @@ def main():
 
     trainer.run()
 
+    chainer.serializers.save_npz(
+        os.path.join(args.out, 'snapshot_model.npz'),
+        recalculate_bn_statistics(model.predictor, 24))
+
 
 if __name__ == '__main__':
     main()

examples/semantic_segmentation/README.md

@@ -18,15 +18,11 @@ The scores are mIoU.
 | PSPNet with ResNet101 (single scale) | 79.70 % [1] | 79.03 % |
 
 ```
-$ python eval_cityscapes.py [--model pspnet_resnet101] [--gpu <gpu>] [--pretrained-model <model_path>]
+$ python eval_semantic_segmentation.py --gpu <GPU> --dataset cityscapes --model pspnet_resnet101
+# with multiple GPUs
+$ mpiexec -n <#gpu> python eval_semantic_segmentation_multi.py --dataset cityscapes --model pspnet_resnet101
 ```
 
-You can conduct evaluation with multiple GPUs by `eval_cityscapes_multi.py`.
-Note that this script requires ChainerMN.
-
-```
-$ mpiexec -n <#gpu> python eval_cityscapes_multi.py [--model pspnet_resnet101] [--pretrained-model <model_path>]
-```
 
 ### CamVid
 
@@ -35,7 +31,7 @@ $ mpiexec -n <#gpu> python eval_cityscapes_multi.py [--model pspnet_resnet101] [
 | SegNet | 46.3 % [2] | 49.4 % |
 
 ```
-$ python eval_camvid.py [--gpu <gpu>] [--pretrained-model <model_path>] [--batchsize <batchsize>]
+$ python eval_semantic_segmentation.py --gpu <GPU> --dataset camvid --model segnet
 ```
 
 

examples/semantic_segmentation/eval_semantic_segmentation.py

@@ -0,0 +1,101 @@
+import argparse
+
+import chainer
+from chainer import iterators
+
+from chainercv.datasets import ade20k_semantic_segmentation_label_names
+from chainercv.datasets import ADE20KSemanticSegmentationDataset
+from chainercv.datasets import camvid_label_names
+from chainercv.datasets import CamVidDataset
+from chainercv.datasets import cityscapes_semantic_segmentation_label_names
+from chainercv.datasets import CityscapesSemanticSegmentationDataset
+
+from chainercv.evaluations import eval_semantic_segmentation
+from chainercv.experimental.links import PSPNetResNet101
+from chainercv.experimental.links import PSPNetResNet50
+from chainercv.links import SegNetBasic
+from chainercv.utils import apply_to_iterator
+from chainercv.utils import ProgressHook
+
+
+def get_dataset_and_model(dataset_name, model_name, pretrained_model,
+                          input_size):
+    if dataset_name == 'cityscapes':
+        dataset = CityscapesSemanticSegmentationDataset(
+            split='val', label_resolution='fine')
+        label_names = cityscapes_semantic_segmentation_label_names
+    elif dataset_name == 'ade20k':
+        dataset = ADE20KSemanticSegmentationDataset(split='val')
+        label_names = ade20k_semantic_segmentation_label_names
+    elif dataset_name == 'camvid':
+        dataset = CamVidDataset(split='test')
+        label_names = camvid_label_names
+
+    n_class = len(label_names)
+
+    if pretrained_model:
+        pretrained_model = pretrained_model
+    else:
+        pretrained_model = dataset_name
+    if model_name == 'pspnet_resnet101':
+        model = PSPNetResNet101(
+            n_class=n_class,
+            pretrained_model=pretrained_model,
+            input_size=input_size
+        )
+    elif model_name == 'pspnet_resnet50':
+        model = PSPNetResNet50(
+            n_class=n_class,
+            pretrained_model=pretrained_model,
+            input_size=input_size
+        )
+    elif model_name == 'segnet':
+        model = SegNetBasic(
+            n_class=n_class, pretrained_model=pretrained_model)
+    return dataset, label_names, model
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        '--dataset', choices=('cityscapes', 'ade20k', 'camvid'))
+    parser.add_argument(
+        '--model', choices=(
+            'pspnet_resnet101', 'segnet'))
+    parser.add_argument('--gpu', type=int, default=-1)
+    parser.add_argument('--pretrained-model')
+    parser.add_argument('--input-size', type=int, default=None)
+    args = parser.parse_args()
+
+    dataset, label_names, model = get_dataset_and_model(
+        args.dataset, args.model, args.pretrained_model,
+        (args.input_size, args.input_size))
+
+    if args.gpu >= 0:
+        chainer.cuda.get_device_from_id(args.gpu).use()
+        model.to_gpu()
+
+    it = iterators.SerialIterator(
+        dataset, 1, repeat=False, shuffle=False)
+
+    in_values, out_values, rest_values = apply_to_iterator(
+        model.predict, it, hook=ProgressHook(len(dataset)))
+    # Delete an iterator of images to save memory usage.
+    del in_values
+    pred_labels, = out_values
+    gt_labels, = rest_values
+
+    result = eval_semantic_segmentation(pred_labels, gt_labels)
+
+    for iu, label_name in zip(result['iou'], label_names):
+        print('{:>23} : {:.4f}'.format(label_name, iu))
+    print('=' * 34)
+    print('{:>23} : {:.4f}'.format('mean IoU', result['miou']))
+    print('{:>23} : {:.4f}'.format(
+        'Class average accuracy', result['mean_class_accuracy']))
+    print('{:>23} : {:.4f}'.format(
+        'Global average accuracy', result['pixel_accuracy']))
+
+
+if __name__ == '__main__':
+    main()