Added ResNet to the list of validated models

README.md

@@ -19,10 +19,12 @@ See the [examples](examples/) folder for more details.
 
 ## Verification
 
-The following converted models have been verified on the ILSVRC2012 validation set.
+The following converted models have been verified on the ILSVRC2012 validation set using
+[validate.py](examples/imagenet/validate.py).
 
 | Model                                                 | Top 5 Accuracy |
 |:------------------------------------------------------|---------------:|
+| [ResNet 50](http://arxiv.org/abs/1512.03385)          |         92.02% |
 | [VGG 16](http://arxiv.org/abs/1409.1556)              |         89.88% |
 | [GoogLeNet](http://arxiv.org/abs/1409.4842)           |         89.06% |
 | [Network in Network](http://arxiv.org/abs/1312.4400)  |         81.21% |
@@ -46,3 +48,5 @@ The following converted models have been verified on the ILSVRC2012 validation s
 - Image rescaling can affect the ILSVRC2012 top 5 accuracy listed above slightly. VGG16 expects isotropic rescaling (anisotropic reduces accuracy to 88.45%) whereas BVLC's implementation of GoogLeNet expects anisotropic (isotropic reduces accuracy to 87.7%).
 
 - The support class `kaffe.tensorflow.Network` has no internal dependencies. It can be safely extracted and deployed without the rest of this library.
+
+- The ResNet model uses 1x1 convolutions with a stride of 2. This is currently only supported in the master branch of TensorFlow (the latest release at time of writing being v0.8.0, which does not support it).

examples/imagenet/models/helper.py

@@ -10,7 +10,7 @@
 from alexnet import AlexNet
 from caffenet import CaffeNet
 from nin import NiN
-
+from resnet import ResNet50, ResNet101, ResNet152
 
 class DataSpec(object):
     '''Input data specifications for an ImageNet model.'''
@@ -32,38 +32,37 @@ def __init__(self, batch_size, scale_size, crop_size, isotropic, channels=3, mea
         # However, using just the per-channel mean values instead doesn't affect things too much.
         self.mean = mean if mean is not None else np.array([104., 117., 124.])
 
+def alexnet_spec(batch_size=500):
+    '''Parameters used by AlexNet and its variants.'''
+    return DataSpec(batch_size=batch_size, scale_size=256, crop_size=227, isotropic=False)
+
+def std_spec(batch_size, isotropic=True):
+    '''Parameters commonly used by "post-AlexNet" architectures.'''
+    return DataSpec(batch_size=batch_size, scale_size=256, crop_size=224, isotropic=isotropic)
+
 # Collection of sample auto-generated models
-MODELS = (AlexNet, CaffeNet, GoogleNet, NiN, VGG16)
+MODELS = (AlexNet, CaffeNet, GoogleNet, NiN, ResNet50, ResNet101, ResNet152, VGG16)
 
 # The corresponding data specifications for the sample models
 # These specifications are based on how the models were trained.
 # The recommended batch size is based on a Titan X (12GB).
 MODEL_DATA_SPECS = {
 
-    AlexNet: DataSpec(batch_size=500,
-                      scale_size=256,
-                      crop_size=227,
-                      isotropic=False),
-
-    CaffeNet: DataSpec(batch_size=500,
-                       scale_size=256,
-                       crop_size=227,
-                       isotropic=False),
-
-    GoogleNet: DataSpec(batch_size=200,
-                        scale_size=256,
-                        crop_size=224,
-                        isotropic=False),
-
-    NiN: DataSpec(batch_size=500,
-                  scale_size=256,
-                  crop_size=224,
-                  isotropic=True),
-
-    VGG16: DataSpec(batch_size=25,
-                    scale_size=256,
-                    crop_size=224,
-                    isotropic=True),
+    AlexNet: alexnet_spec(),
+
+    CaffeNet: alexnet_spec(),
+
+    GoogleNet: std_spec(batch_size=200, isotropic=False),
+
+    ResNet50: std_spec(batch_size=25),
+
+    ResNet101: std_spec(batch_size=25),
+
+    ResNet152: std_spec(batch_size=25),
+
+    NiN: std_spec(batch_size=500),
+
+    VGG16: std_spec(batch_size=224)
 }