Add VGG16

chainercv/links/__init__.py

@@ -9,3 +9,4 @@
 from chainercv.links.model.segnet.segnet_basic import SegNetBasic  # NOQA
 from chainercv.links.model.ssd import SSD300  # NOQA
 from chainercv.links.model.ssd import SSD512  # NOQA
+from chainercv.links.model.vgg import VGG16  # NOQA

chainercv/links/model/faster_rcnn/__init__.py

@@ -1,7 +1,6 @@
 from chainercv.links.model.faster_rcnn.faster_rcnn import FasterRCNN  # NOQA
 from chainercv.links.model.faster_rcnn.faster_rcnn_train_chain import FasterRCNNTrainChain  # NOQA
 from chainercv.links.model.faster_rcnn.faster_rcnn_vgg import FasterRCNNVGG16  # NOQA
-from chainercv.links.model.faster_rcnn.faster_rcnn_vgg import VGG16FeatureExtractor  # NOQA
 from chainercv.links.model.faster_rcnn.faster_rcnn_vgg import VGG16RoIHead  # NOQA
 from chainercv.links.model.faster_rcnn.region_proposal_network import RegionProposalNetwork  # NOQA
 from chainercv.links.model.faster_rcnn.utils.anchor_target_creator import AnchorTargetCreator  # NOQA

chainercv/links/model/faster_rcnn/faster_rcnn_vgg.py

@@ -1,14 +1,13 @@
-import collections
 import numpy as np
 
 import chainer
 import chainer.functions as F
 import chainer.links as L
-from chainer.links import VGG16Layers
 
 from chainercv.links.model.faster_rcnn.faster_rcnn import FasterRCNN
 from chainercv.links.model.faster_rcnn.region_proposal_network import \
     RegionProposalNetwork
+from chainercv.links.model.vgg.vgg16 import VGG16
 from chainercv.utils import download_model
 
 
@@ -74,7 +73,8 @@ class FasterRCNNVGG16(FasterRCNN):
         'voc07': {
             'n_fg_class': 20,
             'url': 'https://github.com/yuyu2172/share-weights/releases/'
-            'download/0.0.3/faster_rcnn_vgg16_voc07_2017_06_06.npz'
+            'download/0.0.4/'
+            'faster_rcnn_vgg16_voc07_trained_2017_08_06_trial_4.npz'
         }
     }
     feat_stride = 16
@@ -103,7 +103,10 @@ def __init__(self,
         if vgg_initialW is None and pretrained_model:
             vgg_initialW = chainer.initializers.constant.Zero()
 
-        extractor = VGG16FeatureExtractor(initialW=vgg_initialW)
+        extractor = VGG16(initialW=vgg_initialW)
+        extractor.feature_names = 'conv5_3'
+        # Delete all layers after conv5_3.
+        extractor.remove_unused()
         rpn = RegionProposalNetwork(
             512, 512,
             ratios=ratios,
@@ -139,12 +142,8 @@ def __init__(self,
             chainer.serializers.load_npz(pretrained_model, self)
 
     def _copy_imagenet_pretrained_vgg16(self):
-        pretrained_model = VGG16Layers()
+        pretrained_model = VGG16(pretrained_model='imagenet')
         self.extractor.conv1_1.copyparams(pretrained_model.conv1_1)
-        # The pretrained weights are trained to accept BGR images.
-        # Convert weights so that they accept RGB images.
-        self.extractor.conv1_1.W.data[:] =\
-            self.extractor.conv1_1.W.data[:, ::-1]
         self.extractor.conv1_2.copyparams(pretrained_model.conv1_2)
         self.extractor.conv2_1.copyparams(pretrained_model.conv2_1)
         self.extractor.conv2_2.copyparams(pretrained_model.conv2_2)
@@ -225,75 +224,8 @@ def __call__(self, x, rois, roi_indices):
         return roi_cls_locs, roi_scores
 
 
-class VGG16FeatureExtractor(chainer.Chain):
-    """Truncated VGG-16 that extracts a conv5_3 feature map.
-
-    Args:
-        initialW (callable): Initializer for the weights.
-
-    """
-
-    def __init__(self, initialW=None):
-        super(VGG16FeatureExtractor, self).__init__()
-        with self.init_scope():
-            self.conv1_1 = L.Convolution2D(3, 64, 3, 1, 1, initialW=initialW)
-            self.conv1_2 = L.Convolution2D(64, 64, 3, 1, 1, initialW=initialW)
-            self.conv2_1 = L.Convolution2D(64, 128, 3, 1, 1, initialW=initialW)
-            self.conv2_2 = L.Convolution2D(
-                128, 128, 3, 1, 1, initialW=initialW)
-            self.conv3_1 = L.Convolution2D(
-                128, 256, 3, 1, 1, initialW=initialW)
-            self.conv3_2 = L.Convolution2D(
-                256, 256, 3, 1, 1, initialW=initialW)
-            self.conv3_3 = L.Convolution2D(
-                256, 256, 3, 1, 1, initialW=initialW)
-            self.conv4_1 = L.Convolution2D(
-                256, 512, 3, 1, 1, initialW=initialW)
-            self.conv4_2 = L.Convolution2D(
-                512, 512, 3, 1, 1, initialW=initialW)
-            self.conv4_3 = L.Convolution2D(
-                512, 512, 3, 1, 1, initialW=initialW)
-            self.conv5_1 = L.Convolution2D(
-                512, 512, 3, 1, 1, initialW=initialW)
-            self.conv5_2 = L.Convolution2D(
-                512, 512, 3, 1, 1, initialW=initialW)
-            self.conv5_3 = L.Convolution2D(
-                512, 512, 3, 1, 1, initialW=initialW)
-
-        self.functions = collections.OrderedDict([
-            ('conv1_1', [self.conv1_1, F.relu]),
-            ('conv1_2', [self.conv1_2, F.relu]),
-            ('pool1', [_max_pooling_2d]),
-            ('conv2_1', [self.conv2_1, F.relu]),
-            ('conv2_2', [self.conv2_2, F.relu]),
-            ('pool2', [_max_pooling_2d]),
-            ('conv3_1', [self.conv3_1, F.relu]),
-            ('conv3_2', [self.conv3_2, F.relu]),
-            ('conv3_3', [self.conv3_3, F.relu]),
-            ('pool3', [_max_pooling_2d]),
-            ('conv4_1', [self.conv4_1, F.relu]),
-            ('conv4_2', [self.conv4_2, F.relu]),
-            ('conv4_3', [self.conv4_3, F.relu]),
-            ('pool4', [_max_pooling_2d]),
-            ('conv5_1', [self.conv5_1, F.relu]),
-            ('conv5_2', [self.conv5_2, F.relu]),
-            ('conv5_3', [self.conv5_3, F.relu]),
-        ])
-
-    def __call__(self, x):
-        h = x
-        for key, funcs in self.functions.items():
-            for func in funcs:
-                h = func(h)
-        return h
-
-
 def _roi_pooling_2d_yx(x, indices_and_rois, outh, outw, spatial_scale):
     xy_indices_and_rois = indices_and_rois[:, [0, 2, 1, 4, 3]]
     pool = F.roi_pooling_2d(
         x, xy_indices_and_rois, outh, outw, spatial_scale)
     return pool
-
-
-def _max_pooling_2d(x):
-    return F.max_pooling_2d(x, ksize=2)

chainercv/links/model/vgg/__init__.py

@@ -0,0 +1 @@
+from chainercv.links.model.vgg.vgg16 import VGG16  # NOQA

chainercv/links/model/vgg/vgg16.py

@@ -0,0 +1,156 @@
+from __future__ import division
+
+import numpy as np
+
+import chainer
+from chainer.functions import dropout
+from chainer.functions import max_pooling_2d
+from chainer.functions import relu
+from chainer.functions import softmax
+from chainer.initializers import constant
+from chainer.initializers import normal
+
+from chainer.links import Linear
+
+from chainercv.utils import download_model
+
+from chainercv.links.connection.conv_2d_activ import Conv2DActiv
+from chainercv.links.model.sequential_feature_extractor import \
+    SequentialFeatureExtractor
+
+
+# RGB order
+_imagenet_mean = np.array(
+    [123.68, 116.779, 103.939], dtype=np.float32)[:, np.newaxis, np.newaxis]
+
+
+class VGG16(SequentialFeatureExtractor):
+
+    """VGG16 Network for classification and feature extraction.
+
+    This is a feature extraction model.
+    The network can choose to output features from set of all
+    intermediate features.
+    The value of :obj:`VGG16.feature_names` selects the features that are going
+    to be collected by :meth:`__call__`.
+    :obj:`self.all_feature_names` is the list of the names of features
+    that can be collected.
+
+    Examples:
+
+        >>> model = VGG16()
+        # By default, VGG16.__call__ returns a probability score.
+        >>> prob = model(imgs)
+
+        >>> model.feature_names = 'conv5_3'
+        # This is feature conv5_3.
+        >>> feat5_3 = model(imgs)
+
+        >>> model.feature_names = ['conv5_3', 'fc6']
+        >>> # These are features conv5_3 and fc6.
+        >>> feat5_3, feat6 = model(imgs)
+
+    .. seealso::
+        :class:`chainercv.links.model.SequentialFeatureExtractor`
+
+    When :obj:`pretrained_model` is the path of a pre-trained chainer model
+    serialized as a :obj:`.npz` file in the constructor, this chain model
+    automatically initializes all the parameters with it.
+    When a string in the prespecified set is provided, a pretrained model is
+    loaded from weights distributed on the Internet.
+    The list of pretrained models supported are as follows:
+
+    * :obj:`imagenet`: Loads weights trained with ImageNet and distributed \
+        at `Model Zoo \
+        <https://github.com/BVLC/caffe/wiki/Model-Zoo>`_.
+
+    Args:
+        pretrained_model (str): The destination of the pre-trained
+            chainer model serialized as a :obj:`.npz` file.
+            If this is one of the strings described
+            above, it automatically loads weights stored under a directory
+            :obj:`$CHAINER_DATASET_ROOT/pfnet/chainercv/models/`,
+            where :obj:`$CHAINER_DATASET_ROOT` is set as
+            :obj:`$HOME/.chainer/dataset` unless you specify another value
+            by modifying the environment variable.
+        n_class (int): The number of classes.
+        mean (numpy.ndarray): A mean value. If :obj:`None` and
+            a supported pretrained model is used,
+            the mean value used to train the pretrained model will be used.
+        initialW (callable): Initializer for the weights.
+        initial_bias (callable): Initializer for the biases.
+
+    """
+
+    _models = {
+        'imagenet': {
+            'n_class': 1000,
+            'url': 'https://github.com/yuyu2172/share-weights/releases/'
+            'download/0.0.4/vgg16_imagenet_convert_2017_07_18.npz',
+            'mean': _imagenet_mean
+        }
+    }
+
+    def __init__(self,
+                 pretrained_model=None, n_class=None, mean=None,
+                 initialW=None, initial_bias=None):
+        if n_class is None:
+            if pretrained_model in self._models:
+                n_class = self._models[pretrained_model]['n_class']
+            else:
+                n_class = 1000
+
+        if mean is None:
+            if pretrained_model in self._models:
+                mean = self._models[pretrained_model]['mean']
+            else:
+                mean = _imagenet_mean
+        self.mean = mean
+
+        if initialW is None:
+            # Employ default initializers used in the original paper.
+            initialW = normal.Normal(0.01)
+        if pretrained_model:
+            # As a sampling process is time-consuming,
+            # we employ a zero initializer for faster computation.
+            initialW = constant.Zero()
+        kwargs = {'initialW': initialW, 'initial_bias': initial_bias}
+
+        super(VGG16, self).__init__()
+        with self.init_scope():
+            self.conv1_1 = Conv2DActiv(None, 64, 3, 1, 1, **kwargs)
+            self.conv1_2 = Conv2DActiv(None, 64, 3, 1, 1, **kwargs)
+            self.pool1 = _max_pooling_2d
+            self.conv2_1 = Conv2DActiv(None, 128, 3, 1, 1, **kwargs)
+            self.conv2_2 = Conv2DActiv(None, 128, 3, 1, 1, **kwargs)
+            self.pool2 = _max_pooling_2d
+            self.conv3_1 = Conv2DActiv(None, 256, 3, 1, 1, **kwargs)
+            self.conv3_2 = Conv2DActiv(None, 256, 3, 1, 1, **kwargs)
+            self.conv3_3 = Conv2DActiv(None, 256, 3, 1, 1, **kwargs)
+            self.pool3 = _max_pooling_2d
+            self.conv4_1 = Conv2DActiv(None, 512, 3, 1, 1, **kwargs)
+            self.conv4_2 = Conv2DActiv(None, 512, 3, 1, 1, **kwargs)
+            self.conv4_3 = Conv2DActiv(None, 512, 3, 1, 1, **kwargs)
+            self.pool4 = _max_pooling_2d
+            self.conv5_1 = Conv2DActiv(None, 512, 3, 1, 1, **kwargs)
+            self.conv5_2 = Conv2DActiv(None, 512, 3, 1, 1, **kwargs)
+            self.conv5_3 = Conv2DActiv(None, 512, 3, 1, 1, **kwargs)
+            self.pool5 = _max_pooling_2d
+            self.fc6 = Linear(None, 4096, **kwargs)
+            self.fc6_relu = relu
+            self.fc6_dropout = dropout
+            self.fc7 = Linear(None, 4096, **kwargs)
+            self.fc7_relu = relu
+            self.fc7_dropout = dropout
+            self.fc8 = Linear(None, n_class, **kwargs)
+            self.prob = softmax
+
+        if pretrained_model in self._models:
+            path = download_model(self._models[pretrained_model]['url'])
+            chainer.serializers.load_npz(path, self)
+        elif pretrained_model:
+            chainer.serializers.load_npz(pretrained_model, self)
+
+
+def _max_pooling_2d(x):
+    return max_pooling_2d(x, ksize=2)

docs/source/reference/links.rst

@@ -1,10 +1,28 @@
 Links
 =====
 
-.. module:: chainercv.links.model.faster_rcnn
+
+Model
+-----
+
+
+Feature Extraction
+~~~~~~~~~~~~~~~~~~
+Feature extraction models can be used to extract feature(s) given images.
+
+.. toctree::
+
+   links/vgg
+
+
+.. autoclass:: chainercv.links.SequentialFeatureExtractor
+   :members:
+
+.. autoclass:: chainercv.links.FeaturePredictor
+
 
 Detection
----------
+~~~~~~~~~
 
 Detection links share a common method :meth:`predict` to detect objects in images.
 For more details, please read :func:`FasterRCNN.predict`.
@@ -16,7 +34,7 @@ For more details, please read :func:`FasterRCNN.predict`.
 
 
 Semantic Segmentation
----------------------
+~~~~~~~~~~~~~~~~~~~~~
 
 .. module:: chainercv.links.model.segnet
 
@@ -29,7 +47,7 @@ For more details, please read :func:`SegNetBasic.predict`.
 
 
 Classifiers
------------
+~~~~~~~~~~~
 
 .. toctree::
 

docs/source/reference/links/faster_rcnn.rst

@@ -45,10 +45,6 @@ RegionProposalNetwork
    :members:
    :special-members:  __call__
 
-VGG16FeatureExtractor
-~~~~~~~~~~~~~~~~~~~~~
-.. autoclass:: VGG16FeatureExtractor
-
 VGG16RoIHead
 ~~~~~~~~~~~~
 .. autoclass:: VGG16RoIHead

docs/source/reference/links/vgg.rst

@@ -0,0 +1,11 @@
+VGG
+===
+
+.. module:: chainercv.links.model.vgg
+
+
+VGG16
+-----
+
+.. autoclass:: VGG16
+   :members: