Add ResNeXt for ImageNet

examples/ResNet/README.md

@@ -22,6 +22,7 @@ baseline and they actually cannot beat this ResNet recipe.
 | ResNet50    | 6.85%       | 23.61%      | [:arrow_down:](http://models.tensorpack.com/ResNet/ImageNet-ResNet50.npz)    |
 | ResNet50-SE | 6.24%       | 22.64%      | [:arrow_down:](http://models.tensorpack.com/ResNet/ImageNet-ResNet50-SE.npz) |
 | ResNet101   | 6.04%       | 21.95%      | [:arrow_down:](http://models.tensorpack.com/ResNet/ImageNet-ResNet101.npz)   |
+| ResNeXt101-32x4d   | 5.73%       | 21.05%      |   [:arrow_down:](http://models.tensorpack.com/ResNet/ImageNet-ResNeXt101-32x4d.npz) |
 | ResNet152   | 5.78%       | 21.51%      | [:arrow_down:](http://models.tensorpack.com/ResNet/ImageNet-ResNet152.npz)   |
 
 To reproduce the above results,

examples/ResNet/imagenet-resnet.py

@@ -15,8 +15,9 @@ from tensorpack.utils.gpu import get_num_gpu
 
 from imagenet_utils import ImageNetModel, eval_on_ILSVRC12, get_imagenet_dataflow, get_imagenet_tfdata
 from resnet_model import (
-    preresnet_basicblock, preresnet_bottleneck, preresnet_group, resnet_backbone, resnet_basicblock, resnet_bottleneck,
-    resnet_group, se_resnet_bottleneck)
+    preresnet_basicblock, preresnet_bottleneck, preresnet_group,
+    resnet_backbone, resnet_group,
+    resnet_basicblock, resnet_bottleneck, resnext_32x4d_bottleneck, se_resnet_bottleneck)
 
 
 class Model(ImageNetModel):
@@ -28,6 +29,7 @@ class Model(ImageNetModel):
         basicblock = preresnet_basicblock if mode == 'preact' else resnet_basicblock
         bottleneck = {
             'resnet': resnet_bottleneck,
+            'resnext32x4d': resnext_32x4d_bottleneck,
             'preact': preresnet_bottleneck,
             'se': se_resnet_bottleneck}[mode]
         self.num_blocks, self.block_func = {
@@ -126,7 +128,7 @@ if __name__ == '__main__':
                         help="total batch size. "
                         "Note that it's best to keep per-GPU batch size in [32, 64] to obtain the best accuracy."
                         "Pretrained models listed in README were trained with batch=32x8.")
-    parser.add_argument('--mode', choices=['resnet', 'preact', 'se'],
+    parser.add_argument('--mode', choices=['resnet', 'preact', 'se', 'resnext32x4d'],
                         help='variants of resnet to use', default='resnet')
     args = parser.parse_args()
 

examples/ResNet/resnet_model.py

@@ -102,6 +102,15 @@ def se_resnet_bottleneck(l, ch_out, stride):
     return tf.nn.relu(out)
 
 
+def resnext_32x4d_bottleneck(l, ch_out, stride):
+    shortcut = l
+    l = Conv2D('conv1', l, ch_out * 2, 1, strides=1, activation=BNReLU)
+    l = Conv2D('conv2', l, ch_out * 2, 3, strides=stride, activation=BNReLU, split=32)
+    l = Conv2D('conv3', l, ch_out * 4, 1, activation=get_bn(zero_init=True))
+    out = l + resnet_shortcut(shortcut, ch_out * 4, stride, activation=get_bn(zero_init=False))
+    return tf.nn.relu(out)
+
+
 def resnet_group(name, l, block_func, features, count, stride):
     with tf.variable_scope(name):
         for i in range(0, count):