Update some model presets (ResNet, XDenseNet)

chainer_/chainercv2/model_provider.py

@@ -100,12 +100,14 @@
     'resnet10': resnet10,
     'resnet12': resnet12,
     'resnet14': resnet14,
+    'resnetbc14b': resnetbc14b,
     'resnet16': resnet16,
     'resnet18_wd4': resnet18_wd4,
     'resnet18_wd2': resnet18_wd2,
     'resnet18_w3d4': resnet18_w3d4,
-
     'resnet18': resnet18,
+    'resnet26': resnet26,
+    'resnetbc26b': resnetbc26b,
     'resnet34': resnet34,
     'resnet50': resnet50,
     'resnet50b': resnet50b,
@@ -471,8 +473,10 @@
 
     'xdensenet40_2_k24_bc_cifar10': xdensenet40_2_k24_bc_cifar10,
     'xdensenet40_2_k24_bc_cifar100': xdensenet40_2_k24_bc_cifar100,
+    'xdensenet40_2_k24_bc_svhn': xdensenet40_2_k24_bc_svhn,
     'xdensenet40_2_k36_bc_cifar10': xdensenet40_2_k36_bc_cifar10,
     'xdensenet40_2_k36_bc_cifar100': xdensenet40_2_k36_bc_cifar100,
+    'xdensenet40_2_k36_bc_svhn': xdensenet40_2_k36_bc_svhn,
 
     'wrn16_10_cifar10': wrn16_10_cifar10,
     'wrn16_10_cifar100': wrn16_10_cifar100,

chainer_/chainercv2/models/resnet.py

@@ -3,9 +3,10 @@
     Original paper: 'Deep Residual Learning for Image Recognition,' https://arxiv.org/abs/1512.03385.
 """
 
-__all__ = ['ResNet', 'resnet10', 'resnet12', 'resnet14', 'resnet16', 'resnet18_wd4', 'resnet18_wd2', 'resnet18_w3d4',
-           'resnet18', 'resnet34', 'resnet50', 'resnet50b', 'resnet101', 'resnet101b', 'resnet152', 'resnet152b',
-           'resnet200', 'resnet200b', 'ResBlock', 'ResBottleneck', 'ResUnit', 'ResInitBlock']
+__all__ = ['ResNet', 'resnet10', 'resnet12', 'resnet14', 'resnetbc14b', 'resnet16', 'resnet18_wd4', 'resnet18_wd2',
+           'resnet18_w3d4', 'resnet18', 'resnet26', 'resnetbc26b', 'resnet34', 'resnet50', 'resnet50b', 'resnet101',
+           'resnet101b', 'resnet152', 'resnet152b', 'resnet200', 'resnet200b', 'ResBlock', 'ResBottleneck', 'ResUnit',
+           'ResInitBlock']
 
 import os
 import chainer.functions as F
@@ -280,6 +281,7 @@ def __call__(self, x):
 
 
 def get_resnet(blocks,
+               bottleneck=None,
                conv1_stride=True,
                width_scale=1.0,
                model_name=None,
@@ -293,6 +295,8 @@ def get_resnet(blocks,
     ----------
     blocks : int
         Number of blocks.
+    bottleneck : bool, default None
+        Whether to use a bottleneck or simple block in units.
     conv1_stride : bool, default True
         Whether to use stride in the first or the second convolution layer in units.
     width_scale : float, default 1.0
@@ -304,17 +308,25 @@ def get_resnet(blocks,
     root : str, default '~/.chainer/models'
         Location for keeping the model parameters.
     """
+    if bottleneck is None:
+        bottleneck = (blocks >= 50)
 
     if blocks == 10:
         layers = [1, 1, 1, 1]
     elif blocks == 12:
         layers = [2, 1, 1, 1]
-    elif blocks == 14:
+    elif blocks == 14 and not bottleneck:
         layers = [2, 2, 1, 1]
+    elif (blocks == 14) and bottleneck:
+        layers = [1, 1, 1, 1]
     elif blocks == 16:
         layers = [2, 2, 2, 1]
     elif blocks == 18:
         layers = [2, 2, 2, 2]
+    elif (blocks == 26) and not bottleneck:
+        layers = [3, 3, 3, 3]
+    elif (blocks == 26) and bottleneck:
+        layers = [2, 2, 2, 2]
     elif blocks == 34:
         layers = [3, 4, 6, 3]
     elif blocks == 50:
@@ -328,19 +340,21 @@ def get_resnet(blocks,
     else:
         raise ValueError("Unsupported ResNet with number of blocks: {}".format(blocks))
 
+    if bottleneck:
+        assert (sum(layers) * 3 + 2 == blocks)
+    else:
+        assert (sum(layers) * 2 + 2 == blocks)
+
     init_block_channels = 64
+    channels_per_layers = [64, 128, 256, 512]
 
-    if blocks < 50:
-        channels_per_layers = [64, 128, 256, 512]
-        bottleneck = False
-    else:
-        channels_per_layers = [256, 512, 1024, 2048]
-        bottleneck = True
+    if bottleneck:
+        bottleneck_factor = 4
+        channels_per_layers = [ci * bottleneck_factor for ci in channels_per_layers]
 
     channels = [[ci] * li for (ci, li) in zip(channels_per_layers, layers)]
 
     if width_scale != 1.0:
-        # channels = [[int(cij * width_scale) for cij in ci] for ci in channels]
         channels = [[int(cij * width_scale) if (i != len(channels) - 1) or (j != len(ci) - 1) else cij
                      for j, cij in enumerate(ci)] for i, ci in enumerate(channels)]
         init_block_channels = int(init_block_channels * width_scale)
@@ -410,6 +424,21 @@ def resnet14(**kwargs):
     return get_resnet(blocks=14, model_name="resnet14", **kwargs)
 
 
+def resnetbc14b(**kwargs):
+    """
+    ResNet-BC-14b model from 'Deep Residual Learning for Image Recognition,' https://arxiv.org/abs/1512.03385.
+    It's an experimental model (bottleneck compressed).
+
+    Parameters:
+    ----------
+    pretrained : bool, default False
+        Whether to load the pretrained weights for model.
+    root : str, default '~/.chainer/models'
+        Location for keeping the model parameters.
+    """
+    return get_resnet(blocks=14, bottleneck=True, conv1_stride=False, model_name="resnetbc14b", **kwargs)
+
+
 def resnet16(**kwargs):
     """
     ResNet-16 model from 'Deep Residual Learning for Image Recognition,' https://arxiv.org/abs/1512.03385.
@@ -484,6 +513,36 @@ def resnet18(**kwargs):
     return get_resnet(blocks=18, model_name="resnet18", **kwargs)
 
 
+def resnet26(**kwargs):
+    """
+    ResNet-26 model from 'Deep Residual Learning for Image Recognition,' https://arxiv.org/abs/1512.03385.
+    It's an experimental model.
+
+    Parameters:
+    ----------
+    pretrained : bool, default False
+        Whether to load the pretrained weights for model.
+    root : str, default '~/.chainer/models'
+        Location for keeping the model parameters.
+    """
+    return get_resnet(blocks=26, bottleneck=False, model_name="resnet26", **kwargs)
+
+
+def resnetbc26b(**kwargs):
+    """
+    ResNet-BC-26b model from 'Deep Residual Learning for Image Recognition,' https://arxiv.org/abs/1512.03385.
+    It's an experimental model (bottleneck compressed).
+
+    Parameters:
+    ----------
+    pretrained : bool, default False
+        Whether to load the pretrained weights for model.
+    root : str, default '~/.chainer/models'
+        Location for keeping the model parameters.
+    """
+    return get_resnet(blocks=26, bottleneck=True, conv1_stride=False, model_name="resnetbc26b", **kwargs)
+
+
 def resnet34(**kwargs):
     """
     ResNet-34 model from 'Deep Residual Learning for Image Recognition,' https://arxiv.org/abs/1512.03385.
@@ -627,12 +686,14 @@ def _test():
         resnet10,
         resnet12,
         resnet14,
+        resnetbc14b,
         resnet16,
         resnet18_wd4,
         resnet18_wd2,
         resnet18_w3d4,
-
         resnet18,
+        resnet26,
+        resnetbc26b,
         resnet34,
         resnet50,
         resnet50b,
@@ -652,11 +713,14 @@ def _test():
         assert (model != resnet10 or weight_count == 5418792)
         assert (model != resnet12 or weight_count == 5492776)
         assert (model != resnet14 or weight_count == 5788200)
+        assert (model != resnetbc14b or weight_count == 10064936)
         assert (model != resnet16 or weight_count == 6968872)
-        assert (model != resnet18_wd4 or weight_count == 3937400)  # 831096
-        assert (model != resnet18_wd2 or weight_count == 5804296)  # 3055880
-        assert (model != resnet18_w3d4 or weight_count == 8476056)  # 6675352
+        assert (model != resnet18_wd4 or weight_count == 3937400)
+        assert (model != resnet18_wd2 or weight_count == 5804296)
+        assert (model != resnet18_w3d4 or weight_count == 8476056)
         assert (model != resnet18 or weight_count == 11689512)
+        assert (model != resnet26 or weight_count == 17960232)
+        assert (model != resnetbc26b or weight_count == 15995176)
         assert (model != resnet34 or weight_count == 21797672)
         assert (model != resnet50 or weight_count == 25557032)
         assert (model != resnet50b or weight_count == 25557032)

chainer_/chainercv2/models/xdensenet_cifar.py

@@ -1,11 +1,12 @@
 """
-    X-DenseNet for CIFAR, implemented in Chainer.
+    X-DenseNet for CIFAR/SVHN, implemented in Chainer.
     Original paper: 'Deep Expander Networks: Efficient Deep Networks from Graph Theory,'
     https://arxiv.org/abs/1711.08757.
 """
 
 __all__ = ['CIFARXDenseNet', 'xdensenet40_2_k24_bc_cifar10', 'xdensenet40_2_k24_bc_cifar100',
-           'xdensenet40_2_k36_bc_cifar10', 'xdensenet40_2_k36_bc_cifar100']
+           'xdensenet40_2_k24_bc_svhn', 'xdensenet40_2_k36_bc_cifar10', 'xdensenet40_2_k36_bc_cifar100',
+           'xdensenet40_2_k36_bc_svhn']
 
 import os
 import chainer.functions as F
@@ -251,6 +252,24 @@ def xdensenet40_2_k24_bc_cifar100(classes=100, **kwargs):
                                model_name="xdensenet40_2_k24_bc_cifar100", **kwargs)
 
 
+def xdensenet40_2_k24_bc_svhn(classes=10, **kwargs):
+    """
+    X-DenseNet-BC-40-2 (k=24) model for SVHN from 'Deep Expander Networks: Efficient Deep Networks from Graph
+    Theory,' https://arxiv.org/abs/1711.08757.
+
+    Parameters:
+    ----------
+    classes : int, default 10
+        Number of classification classes.
+    pretrained : bool, default False
+        Whether to load the pretrained weights for model.
+    root : str, default '~/.chainer/models'
+        Location for keeping the model parameters.
+    """
+    return get_xdensenet_cifar(classes=classes, blocks=40, growth_rate=24, bottleneck=True,
+                               model_name="xdensenet40_2_k24_bc_svhn", **kwargs)
+
+
 def xdensenet40_2_k36_bc_cifar10(classes=10, **kwargs):
     """
     X-DenseNet-BC-40-2 (k=36) model for CIFAR-10 from 'Deep Expander Networks: Efficient Deep Networks from Graph
@@ -287,6 +306,24 @@ def xdensenet40_2_k36_bc_cifar100(classes=100, **kwargs):
                                model_name="xdensenet40_2_k36_bc_cifar100", **kwargs)
 
 
+def xdensenet40_2_k36_bc_svhn(classes=10, **kwargs):
+    """
+    X-DenseNet-BC-40-2 (k=36) model for SVHN from 'Deep Expander Networks: Efficient Deep Networks from Graph
+    Theory,' https://arxiv.org/abs/1711.08757.
+
+    Parameters:
+    ----------
+    classes : int, default 10
+        Number of classification classes.
+    pretrained : bool, default False
+        Whether to load the pretrained weights for model.
+    root : str, default '~/.chainer/models'
+        Location for keeping the model parameters.
+    """
+    return get_xdensenet_cifar(classes=classes, blocks=40, growth_rate=36, bottleneck=True,
+                               model_name="xdensenet40_2_k36_bc_svhn", **kwargs)
+
+
 def _test():
     import numpy as np
     import chainer
@@ -298,8 +335,10 @@ def _test():
     models = [
         (xdensenet40_2_k24_bc_cifar10, 10),
         (xdensenet40_2_k24_bc_cifar100, 100),
+        (xdensenet40_2_k24_bc_svhn, 10),
         (xdensenet40_2_k36_bc_cifar10, 10),
         (xdensenet40_2_k36_bc_cifar100, 100),
+        (xdensenet40_2_k36_bc_svhn, 10),
     ]
 
     for model, classes in models:
@@ -309,8 +348,10 @@ def _test():
         print("m={}, {}".format(model.__name__, weight_count))
         assert (model != xdensenet40_2_k24_bc_cifar10 or weight_count == 690346)
         assert (model != xdensenet40_2_k24_bc_cifar100 or weight_count == 714196)
+        assert (model != xdensenet40_2_k24_bc_svhn or weight_count == 690346)
         assert (model != xdensenet40_2_k36_bc_cifar10 or weight_count == 1542682)
         assert (model != xdensenet40_2_k36_bc_cifar100 or weight_count == 1578412)
+        assert (model != xdensenet40_2_k36_bc_svhn or weight_count == 1542682)
 
         x = np.zeros((1, 3, 32, 32), np.float32)
         y = net(x)

gluon/gluoncv2/model_provider.py

@@ -109,12 +109,14 @@
     'resnet10': resnet10,
     'resnet12': resnet12,
     'resnet14': resnet14,
+    'resnetbc14b': resnetbc14b,
     'resnet16': resnet16,
     'resnet18_wd4': resnet18_wd4,
     'resnet18_wd2': resnet18_wd2,
     'resnet18_w3d4': resnet18_w3d4,
-
     'resnet18': resnet18,
+    'resnet26': resnet26,
+    'resnetbc26b': resnetbc26b,
     'resnet34': resnet34,
     'resnet50': resnet50,
     'resnet50b': resnet50b,
@@ -125,10 +127,6 @@
     'resnet200': resnet200,
     'resnet200b': resnet200b,
 
-    'resnetbn14b': resnetbn14b,
-    'resnet26': resnet26,
-    'resnetbn26b': resnetbn26b,
-
     'preresnet10': preresnet10,
     'preresnet12': preresnet12,
     'preresnet14': preresnet14,
@@ -509,8 +507,10 @@
 
     'xdensenet40_2_k24_bc_cifar10': xdensenet40_2_k24_bc_cifar10,
     'xdensenet40_2_k24_bc_cifar100': xdensenet40_2_k24_bc_cifar100,
+    'xdensenet40_2_k24_bc_svhn': xdensenet40_2_k24_bc_svhn,
     'xdensenet40_2_k36_bc_cifar10': xdensenet40_2_k36_bc_cifar10,
     'xdensenet40_2_k36_bc_cifar100': xdensenet40_2_k36_bc_cifar100,
+    'xdensenet40_2_k36_bc_svhn': xdensenet40_2_k36_bc_svhn,
 
     'wrn16_10_cifar10': wrn16_10_cifar10,
     'wrn16_10_cifar100': wrn16_10_cifar100,