Add training part of PSPNet

chainercv/datasets/ade20k/ade20k_semantic_segmentation_dataset.py

@@ -38,7 +38,7 @@ class ADE20KSemanticSegmentationDataset(GetterDataset):
         :obj:`img`, ":math:`(3, H, W)`", :obj:`float32`, \
         "RGB, :math:`[0, 255]`"
         :obj:`label`, ":math:`(H, W)`", :obj:`int32`, \
-        ":math:`[0, \#class - 1]`"
+        ":math:`[-1, \#class - 1]`"
     """
 
     def __init__(self, data_dir='auto', split='train'):
@@ -72,5 +72,7 @@ def _get_image(self, i):
         return read_image(self.img_paths[i])
 
     def _get_label(self, i):
-        return read_image(
+        label = read_image(
             self.label_paths[i], dtype=np.int32, color=False)[0]
+        # [-1, n_class - 1]
+        return label - 1

chainercv/datasets/ade20k/ade20k_utils.py

@@ -17,7 +17,6 @@ def get_ade20k(root, url):
 
 
 ade20k_semantic_segmentation_label_names = (
-    'other_objects',
     'wall',
     'edifice',
     'sky',
@@ -171,7 +170,6 @@ def get_ade20k(root, url):
 )
 
 ade20k_semantic_segmentation_label_colors = (
-    (0, 0, 0),
     (120, 120, 120),
     (180, 120, 120),
     (6, 230, 230),

chainercv/experimental/links/model/pspnet/pspnet.py

@@ -305,8 +305,13 @@ class PSPNetResNet101(PSPNet):
     _models = {
         'cityscapes': {
             'param': {'n_class': 19, 'input_size': (713, 713)},
-            'url': 'https://github.com/yuyu2172/share-weights/releases/'
-            'download/0.0.6/pspnet_resnet101_cityscapes_convert_2018_05_22.npz'
+            'url': 'https://chainercv-models.preferred.jp/'
+            'pspnet_resnet101_cityscapes_trained_2018_12_19.npz',
+        },
+        'ade20k': {
+            'param': {'n_class': 150, 'input_size': (473, 473)},
+            'url': 'https://chainercv-models.preferred.jp/'
+            'pspnet_resnet101_ade20k_trained_2018_12_23.npz',
         },
     }
 
@@ -324,6 +329,16 @@ class PSPNetResNet50(PSPNet):
     _extractor_kwargs = {'n_layer': 50}
     _extractor_pick = ('res4', 'res5')
     _models = {
+        'cityscapes': {
+            'param': {'n_class': 19, 'input_size': (713, 713)},
+            'url': 'https://chainercv-models.preferred.jp/'
+            'pspnet_resnet50_cityscapes_trained_2018_12_19.npz',
+        },
+        'ade20k': {
+            'param': {'n_class': 150, 'input_size': (473, 473)},
+            'url': 'https://chainercv-models.preferred.jp/'
+            'pspnet_resnet50_ade20k_trained_2018_12_23.npz',
+        },
     }
 
 

examples/pspnet/README.md

@@ -1,32 +1,68 @@
 # Examples of Pyramid Scene Parsing Network (PSPNet)
 
-## Performance
+## Demo
+This demo downloads a pretrained model automatically if a pretrained model path is not given.
+```
+$ python demo.py [--gpu <gpu>] [--pretrained-model <model_path>] [--input-size <size>] <image>.jpg
+```
+
+## Weight Covnersion
+
+Convert `*.caffemodel` to `*.npz`. Some layers are renamed to fit ChainerCV.
+```
+$ python caffe2npz.py <source>.caffemodel <target>.npz
+```
+
+The converted weight can be downloaded from [here](https://chainercv-models.preferred.jp/pspnet_resnet101_cityscapes_converted_2018_05_22.npz).
+
+The performance on the Cityscapes dataset is as follows with single scale inference.
+Scores are measured by mean Intersection over Union (mIoU).
 
 | Model | Reference | ChainerCV (weight conversion) |
 |:-:|:-:|:-:|
 | Cityscapes (single scale) | 79.70 % [1] | 79.03 % |
 
-Scores are measured by mean Intersection over Union (mIoU).
+## Training model
 
+The model can be trained with a script `train_mutli.py`.
+
+### Cityscapes
+
+The following table shows the performance of the models trained with our scripts.
+
+| Base model | Training Data |  Reference | ChainerCV |
+|:-:|:-:|:-:|:-:|
+| Dilated ResNet50 | fine only (3K) | 76.9 % [2] |  73.99 % |
+| Dilated ResNet101 | fine only (3K) |  77.9 % [2] | 76.01 % |
+
+Here are the commands used to train the models included in the table.
 
-## Demo
-This demo downloads Cityscapes pretrained model automatically if a pretrained model path is not given.
 ```
-$ python demo.py [--gpu <gpu>] [--pretrained-model <model_path>] [--input-size <size>] <image>.jpg
+$ mpiexec -n 8 python3 train_multi.py --dataset cityscapes --model pspnet_resnet50 --iteration 90000
+$ mpiexec -n 8 python3 train_multi.py --dataset cityscapes --model pspnet_resnet101 --iteration 90000
 ```
 
+### ADE20K
+
+The following table shows the performance of the models trained with our scripts.
+
+| Base model |  Reference | ChainerCV |
+|:-:|:-:|:-:|
+| Dilated ResNet50 | 41.68 % [1] |  34.97 % |
+| Dilated ResNet101 |  | 36.55 % |
+
+Here are the commands used to train the models included in the table.
 
-## Convert Caffe model
-Convert `*.caffemodel` to `*.npz`. Some layers are renamed to fit ChainerCV.
 ```
-$ python caffe2npz.py <source>.caffemodel <target>.npz
+$ mpiexec -n 8 python3 train_multi.py --dataset ade20k --model pspnet_resnet50 --iteration 150000
+$ mpiexec -n 8 python3 train_multi.py --dataset ade20k --model pspnet_resnet101 --iteration 150000
 ```
 
-
 ## Evaluation
-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).
 
 
 ## References
 1. Hengshuang Zhao et al. "Pyramid Scene Parsing Network" CVPR 2017.
-2. [chainer-pspnet by mitmul](https://github.com/mitmul/chainer-pspnet)
+2. https://github.com/holyseven/PSPNet-TF-Reproduce (Validation scores for Cityscapes are lacking in the original paper)
+3. [chainer-pspnet by mitmul](https://github.com/mitmul/chainer-pspnet)