Add MobileNetv2 with pretrained weights

Author: taehoonlee

README.md

@@ -277,10 +277,16 @@ with tf.Session() as sess:
 | [DenseNet121](tensornets/densenets.py#L63)        | 25.480      | 8.022       | 6.842       | 8.1M   | 7.0M   | 202.9 | [[paper]](https://arxiv.org/abs/1608.06993) [[torch]](https://github.com/liuzhuang13/DenseNet/blob/master/models/densenet.lua) |
 | [DenseNet169](tensornets/densenets.py#L71)        | 23.926      | 6.892       | 6.140       | 14.3M  | 12.6M  | 219.1 | [[paper]](https://arxiv.org/abs/1608.06993) [[torch]](https://github.com/liuzhuang13/DenseNet/blob/master/models/densenet.lua) |
 | [DenseNet201](tensornets/densenets.py#L79)        | 22.936      | 6.542       | 5.724       | 20.2M  | 18.3M  | 272.0 | [[paper]](https://arxiv.org/abs/1608.06993) [[torch]](https://github.com/liuzhuang13/DenseNet/blob/master/models/densenet.lua) |
-| [MobileNet25](tensornets/mobilenets.py#L84)       | 48.418      | 24.208      | 21.196      | 0.5M   | 0.2M   | 29.27 | [[paper]](https://arxiv.org/abs/1704.04861) [[tf-slim]](https://github.com/tensorflow/models/blob/master/research/slim/nets/mobilenet_v1.py) |
-| [MobileNet50](tensornets/mobilenets.py#L91)       | 35.708      | 14.376      | 12.180      | 1.3M   | 0.8M   | 42.32 | [[paper]](https://arxiv.org/abs/1704.04861) [[tf-slim]](https://github.com/tensorflow/models/blob/master/research/slim/nets/mobilenet_v1.py) |
-| [MobileNet75](tensornets/mobilenets.py#L98)       | 31.588      | 11.758      | 9.878       | 2.6M   | 1.8M   | 57.23 | [[paper]](https://arxiv.org/abs/1704.04861) [[tf-slim]](https://github.com/tensorflow/models/blob/master/research/slim/nets/mobilenet_v1.py) |
-| [MobileNet100](tensornets/mobilenets.py#L105)     | 29.576      | 10.496      | 8.774       | 4.3M   | 3.2M   | 70.69 | [[paper]](https://arxiv.org/abs/1704.04861) [[tf-slim]](https://github.com/tensorflow/models/blob/master/research/slim/nets/mobilenet_v1.py) |
+| [MobileNet25](tensornets/mobilenets.py#L156)      | 48.418      | 24.208      | 21.196      | 0.5M   | 0.2M   | 34.46 | [[paper]](https://arxiv.org/abs/1704.04861) [[tf-slim]](https://github.com/tensorflow/models/blob/master/research/slim/nets/mobilenet_v1.py) |
+| [MobileNet50](tensornets/mobilenets.py#L163)      | 35.708      | 14.376      | 12.180      | 1.3M   | 0.8M   | 52.46 | [[paper]](https://arxiv.org/abs/1704.04861) [[tf-slim]](https://github.com/tensorflow/models/blob/master/research/slim/nets/mobilenet_v1.py) |
+| [MobileNet75](tensornets/mobilenets.py#L170)      | 31.588      | 11.758      | 9.878       | 2.6M   | 1.8M   | 70.11 | [[paper]](https://arxiv.org/abs/1704.04861) [[tf-slim]](https://github.com/tensorflow/models/blob/master/research/slim/nets/mobilenet_v1.py) |
+| [MobileNet100](tensornets/mobilenets.py#L177)     | 29.576      | 10.496      | 8.774       | 4.3M   | 3.2M   | 83.41 | [[paper]](https://arxiv.org/abs/1704.04861) [[tf-slim]](https://github.com/tensorflow/models/blob/master/research/slim/nets/mobilenet_v1.py) |
+| [MobileNet35v2](tensornets/mobilenets.py#L184)    | 39.914      | 17.568      | 15.422      | 1.7M   | 0.4M   | 57.04 | [[paper]](https://arxiv.org/abs/1801.04381) [[tf-slim]](https://github.com/tensorflow/models/blob/master/research/slim/nets/mobilenet/mobilenet_v2.py) |
+| [MobileNet50v2](tensornets/mobilenets.py#L191)    | 34.806      | 13.938      | 11.976      | 2.0M   | 0.7M   | 64.35 | [[paper]](https://arxiv.org/abs/1801.04381) [[tf-slim]](https://github.com/tensorflow/models/blob/master/research/slim/nets/mobilenet/mobilenet_v2.py) |
+| [MobileNet75v2](tensornets/mobilenets.py#L198)    | 30.468      | 10.824      | 9.188       | 2.7M   | 1.4M   | 88.68 | [[paper]](https://arxiv.org/abs/1801.04381) [[tf-slim]](https://github.com/tensorflow/models/blob/master/research/slim/nets/mobilenet/mobilenet_v2.py) |
+| [MobileNet100v2](tensornets/mobilenets.py#L205)   | 28.664      | 9.858       | 8.322       | 3.5M   | 2.3M   | 93.82 | [[paper]](https://arxiv.org/abs/1801.04381) [[tf-slim]](https://github.com/tensorflow/models/blob/master/research/slim/nets/mobilenet/mobilenet_v2.py) |
+| [MobileNet130v2](tensornets/mobilenets.py#L212)   | 25.320      | 7.878       | 6.728       | 5.4M   | 3.8M   | 130.4 | [[paper]](https://arxiv.org/abs/1801.04381) [[tf-slim]](https://github.com/tensorflow/models/blob/master/research/slim/nets/mobilenet/mobilenet_v2.py) |
+| [MobileNet140v2](tensornets/mobilenets.py#L219)   | 24.770      | 7.578       | 6.518       | 6.2M   | 4.4M   | 132.9 | [[paper]](https://arxiv.org/abs/1801.04381) [[tf-slim]](https://github.com/tensorflow/models/blob/master/research/slim/nets/mobilenet/mobilenet_v2.py) |
 | [SqueezeNet](tensornets/squeezenets.py#L45)       | 45.566      | 21.960      | 18.578      | 1.2M   | 0.7M   | 71.43 | [[paper]](https://arxiv.org/abs/1602.07360) [[caffe]](https://github.com/DeepScale/SqueezeNet/blob/master/SqueezeNet_v1.1/train_val.prototxt) |
 
 ### Object detection
@@ -310,13 +316,14 @@ with tf.Session() as sess:
 
 ## News 📰
 
+- The six variants of MobileNetv2 are released, [5 May 2018]().
 - YOLOv3 for COCO and VOC are released, [4 April 2018](https://github.com/taehoonlee/tensornets/commit/d8b2d8a54dc4b775a174035da63561028deb6624).
 - Generic object detection models for YOLOv2 and FasterRCNN are released, [26 March 2018](https://github.com/taehoonlee/tensornets/commit/67915e659d2097a96c82ba7740b9e43a8c69858d).
 
 ## Future work 🔥
 
 - Add training codes.
-- Add image classification models (MobileNetv2, PNASNet).
+- Add image classification models (PolyNet, PNASNet).
 - Add object detection models (MaskRCNN, SSD).
 - Add image segmentation models (FCN, UNet).
 - Add image datasets (COCO, OpenImages).

tensornets/__init__.py

@@ -36,6 +36,13 @@
 from .mobilenets import MobileNet75
 from .mobilenets import MobileNet100
 
+from .mobilenets import MobileNet35v2
+from .mobilenets import MobileNet50v2
+from .mobilenets import MobileNet75v2
+from .mobilenets import MobileNet100v2
+from .mobilenets import MobileNet130v2
+from .mobilenets import MobileNet140v2
+
 from .squeezenets import SqueezeNet
 
 from .capsulenets import CapsuleNet

tensornets/middles.py

@@ -1,6 +1,8 @@
 """Collection of representative endpoints for each model."""
 from __future__ import absolute_import
 
+from .utils import tf_later_than
+
 
 def names_inceptions(k, first_block, omit_first=False,
                      pool_last=False, resnet=False):
@@ -59,6 +61,37 @@ def names_darknets(k):
     return names
 
 
+def tuple_mobilenetv2():
+    def baseidx(b):
+        return [b, b + 3, b + 5]
+    indices = baseidx(2)
+    if tf_later_than('1.3.0'):
+        bn_name = 'FusedBatchNorm:0'
+    else:
+        bn_name = 'batchnorm/add_1:0'
+    names = ['conv1/Relu6:0', 'sconv1/Relu6:0', 'pconv1/bn/' + bn_name]
+    k = 10
+    l = 2
+    for (i, j) in enumerate([2, 3, 4, 3, 3, 1]):
+        indices += baseidx(k)
+        names += ["conv%d/conv/Relu6:0" % l,
+                  "conv%d/sconv/Relu6:0" % l,
+                  "conv%d/pconv/bn/%s" % (l, bn_name)]
+        k += 8
+        l += 1
+        for _ in range(j - 1):
+            indices += (baseidx(k) + [k + 6])
+            names += ["conv%d/conv/Relu6:0" % l,
+                      "conv%d/sconv/Relu6:0" % l,
+                      "conv%d/pconv/bn/%s" % (l, bn_name),
+                      "conv%d/out:0" % l]
+            k += 9
+            l += 1
+    indices += [k]
+    names += ["conv%d/Relu6:0" % l]
+    return (indices, names, -16)
+
+
 def direct(model_name):
     try:
         return __middles_dict__[model_name]
@@ -236,6 +269,12 @@ def direct(model_name):
         ['conv%d/conv/Relu6:0' % (i + 4) for i in range(11)],
         -3
     ),
+    'mobilenet35v2': tuple_mobilenetv2(),
+    'mobilenet50v2': tuple_mobilenetv2(),
+    'mobilenet75v2': tuple_mobilenetv2(),
+    'mobilenet100v2': tuple_mobilenetv2(),
+    'mobilenet130v2': tuple_mobilenetv2(),
+    'mobilenet140v2': tuple_mobilenetv2(),
     'squeezenet': (
         [9, 16, 17, 24, 31, 32] + list(range(39, 61, 7)),
         names_squeezenet(),

tensornets/mobilenets.py

@@ -1,17 +1,25 @@
 """Collection of MobileNet variants
 
-The reference paper:
+The reference papers:
 
+1. V1
  - MobileNets: Efficient Convolutional Neural Networks for Mobile Vision
    Applications, arXiv 2017
  - Andrew G. Howard et al.
  - https://arxiv.org/abs/1704.04861
+2. V2
+ - MobileNetV2: Inverted Residuals and Linear Bottlenecks, arXiv 2018
+ - Mark Sandler et al.
+ - https://arxiv.org/abs/1801.04381
 
-The reference implementation:
+The reference implementations:
 
-1. TF Slim
+1. (for v1) TF Slim
  - https://github.com/tensorflow/models/blob/master/research/slim/nets/
    mobilenet_v1.py
+2. (for v2) TF Slim
+ - https://github.com/tensorflow/models/blob/master/research/slim/nets/
+   mobilenet/mobilenet_v2.py
 """
 from __future__ import absolute_import
 
@@ -22,6 +30,7 @@
 from .layers import dropout
 from .layers import fc
 from .layers import separable_conv2d
+from .layers import convbn
 from .layers import convbnrelu6 as conv
 from .layers import sconvbnrelu6 as sconv
 
@@ -30,8 +39,8 @@
 from .utils import var_scope
 
 
-def __args__(is_training):
-    return [([batch_norm], {'decay': 0.9997, 'scale': True, 'epsilon': 0.001,
+def __base_args__(is_training, decay):
+    return [([batch_norm], {'decay': decay, 'scale': True, 'epsilon': 0.001,
                             'is_training': is_training, 'scope': 'bn'}),
             ([conv2d], {'padding': 'SAME', 'activation_fn': None,
                         'biases_initializer': None, 'scope': 'conv'}),
@@ -42,13 +51,33 @@ def __args__(is_training):
               'scope': 'sconv'})]
 
 
+def __args__(is_training):
+    return __base_args__(is_training, 0.9997)
+
+
+def __args_v2__(is_training):
+    return __base_args__(is_training, 0.999)
+
+
 @var_scope('block')
 def block(x, filters, stride=1, scope=None):
     x = sconv(x, None, 3, 1, stride=stride, scope='sconv')
     x = conv(x, filters, 1, stride=1, scope='conv')
     return x
 
 
+@var_scope('blockv2')
+def block2(x, filters, first=False, stride=1, scope=None):
+    shortcut = x
+    x = conv(x, 6 * x.shape[-1].value, 1, scope='conv')
+    x = sconv(x, None, 3, 1, stride=stride, scope='sconv')
+    x = convbn(x, filters, 1, stride=1, scope='pconv')
+    if stride == 1 and shortcut.shape[-1].value == filters:
+        return add(shortcut, x, name='out')
+    else:
+        return x
+
+
 def mobilenet(x, depth_multiplier, is_training, classes, stem,
               scope=None, reuse=None):
     def depth(d):
@@ -81,6 +110,49 @@ def depth(d):
     return x
 
 
+def mobilenetv2(x, depth_multiplier, is_training, classes, stem,
+                scope=None, reuse=None):
+    def depth(d):
+        d *= depth_multiplier
+        filters = max(8, int(d + 4) // 8 * 8)
+        if filters < 0.9 * d:
+            filters += 8
+        return filters
+    x = conv(x, depth(32), 3, stride=2, scope='conv1')
+    x = sconv(x, None, 3, 1, scope='sconv1')
+    x = convbn(x, depth(16), 1, scope='pconv1')
+
+    x = block2(x, depth(24), stride=2, scope='conv2')
+    x = block2(x, depth(24), scope='conv3')
+
+    x = block2(x, depth(32), stride=2, scope='conv4')
+    x = block2(x, depth(32), scope='conv5')
+    x = block2(x, depth(32), scope='conv6')
+
+    x = block2(x, depth(64), stride=2, scope='conv7')
+    x = block2(x, depth(64), scope='conv8')
+    x = block2(x, depth(64), scope='conv9')
+    x = block2(x, depth(64), scope='conv10')
+
+    x = block2(x, depth(96), scope='conv11')
+    x = block2(x, depth(96), scope='conv12')
+    x = block2(x, depth(96), scope='conv13')
+
+    x = block2(x, depth(160), stride=2, scope='conv14')
+    x = block2(x, depth(160), scope='conv15')
+    x = block2(x, depth(160), scope='conv16')
+
+    x = block2(x, depth(320), scope='conv17')
+    x = conv(x, 1280 * depth_multiplier if depth_multiplier > 1. else 1280, 1,
+             scope='conv18')
+    if stem: return x
+
+    x = reduce_mean(x, [1, 2], name='avgpool')
+    x = fc(x, classes, scope='logits')
+    x = softmax(x, name='probs')
+    return x
+
+
 @var_scope('mobilenet25')
 @set_args(__args__)
 def mobilenet25(x, is_training=False, classes=1000,
@@ -109,8 +181,56 @@ def mobilenet100(x, is_training=False, classes=1000,
     return mobilenet(x, 1.0, is_training, classes, stem, scope, reuse)
 
 
+@var_scope('mobilenet35v2')
+@set_args(__args_v2__)
+def mobilenet35v2(x, is_training=False, classes=1000,
+                  stem=False, scope=None, reuse=None):
+    return mobilenetv2(x, 0.35, is_training, classes, stem, scope, reuse)
+
+
+@var_scope('mobilenet50v2')
+@set_args(__args_v2__)
+def mobilenet50v2(x, is_training=False, classes=1000,
+                  stem=False, scope=None, reuse=None):
+    return mobilenetv2(x, 0.50, is_training, classes, stem, scope, reuse)
+
+
+@var_scope('mobilenet75v2')
+@set_args(__args_v2__)
+def mobilenet75v2(x, is_training=False, classes=1000,
+                  stem=False, scope=None, reuse=None):
+    return mobilenetv2(x, 0.75, is_training, classes, stem, scope, reuse)
+
+
+@var_scope('mobilenet100v2')
+@set_args(__args_v2__)
+def mobilenet100v2(x, is_training=False, classes=1000,
+                   stem=False, scope=None, reuse=None):
+    return mobilenetv2(x, 1.0, is_training, classes, stem, scope, reuse)
+
+
+@var_scope('mobilenet130v2')
+@set_args(__args_v2__)
+def mobilenet130v2(x, is_training=False, classes=1000,
+                   stem=False, scope=None, reuse=None):
+    return mobilenetv2(x, 1.3, is_training, classes, stem, scope, reuse)
+
+
+@var_scope('mobilenet140v2')
+@set_args(__args_v2__)
+def mobilenet140v2(x, is_training=False, classes=1000,
+                   stem=False, scope=None, reuse=None):
+    return mobilenetv2(x, 1.4, is_training, classes, stem, scope, reuse)
+
+
 # Simple alias.
 MobileNet25 = mobilenet25
 MobileNet50 = mobilenet50
 MobileNet75 = mobilenet75
 MobileNet100 = mobilenet100
+MobileNet35v2 = mobilenet35v2
+MobileNet50v2 = mobilenet50v2
+MobileNet75v2 = mobilenet75v2
+MobileNet100v2 = mobilenet100v2
+MobileNet130v2 = mobilenet130v2
+MobileNet140v2 = mobilenet140v2

tensornets/preprocess.py

@@ -165,6 +165,13 @@ def faster_rcnn_preprocess(x):
     'mobilenet50': tfslim_preprocess,
     'mobilenet75': tfslim_preprocess,
     'mobilenet100': tfslim_preprocess,
+    'mobilenetv2': tfslim_preprocess,
+    'mobilenet35v2': tfslim_preprocess,
+    'mobilenet50v2': tfslim_preprocess,
+    'mobilenet75v2': tfslim_preprocess,
+    'mobilenet100v2': tfslim_preprocess,
+    'mobilenet130v2': tfslim_preprocess,
+    'mobilenet140v2': tfslim_preprocess,
     'squeezenet': bair_preprocess,
     'zf': faster_rcnn_preprocess,
     'darknet19': darknet_preprocess,

tensornets/pretrained.py

@@ -530,6 +530,72 @@ def load_mobilenet100(scopes, return_fn=_assign):
     return return_fn(scopes, values)
 
 
+def load_mobilenet35v2(scopes, return_fn=_assign):
+    """Converted from the [TF Slim][2]."""
+    filename = 'mobilenet_v2_035_224.npz'
+    weights_path = get_file(
+        filename, __model_url__ + 'mobilenet/' + filename,
+        cache_subdir='models',
+        file_hash='cf758f7f8024d39365e553ec924bb395')
+    values = parse_weights(weights_path)
+    return return_fn(scopes, values)
+
+
+def load_mobilenet50v2(scopes, return_fn=_assign):
+    """Converted from the [TF Slim][2]."""
+    filename = 'mobilenet_v2_050_224.npz'
+    weights_path = get_file(
+        filename, __model_url__ + 'mobilenet/' + filename,
+        cache_subdir='models',
+        file_hash='218d51cd1b12b03ece24054029e7005b')
+    values = parse_weights(weights_path)
+    return return_fn(scopes, values)
+
+
+def load_mobilenet75v2(scopes, return_fn=_assign):
+    """Converted from the [TF Slim][2]."""
+    filename = 'mobilenet_v2_075_224.npz'
+    weights_path = get_file(
+        filename, __model_url__ + 'mobilenet/' + filename,
+        cache_subdir='models',
+        file_hash='25b5f6c93ebec7558a757e7a70b16b1c')
+    values = parse_weights(weights_path)
+    return return_fn(scopes, values)
+
+
+def load_mobilenet100v2(scopes, return_fn=_assign):
+    """Converted from the [TF Slim][2]."""
+    filename = 'mobilenet_v2_100_224.npz'
+    weights_path = get_file(
+        filename, __model_url__ + 'mobilenet/' + filename,
+        cache_subdir='models',
+        file_hash='ea55ba8d51df1df59b196d2508a3f262')
+    values = parse_weights(weights_path)
+    return return_fn(scopes, values)
+
+
+def load_mobilenet130v2(scopes, return_fn=_assign):
+    """Converted from the [TF Slim][2]."""
+    filename = 'mobilenet_v2_130_224.npz'
+    weights_path = get_file(
+        filename, __model_url__ + 'mobilenet/' + filename,
+        cache_subdir='models',
+        file_hash='a2470b36675853bbe107a406b50cd648')
+    values = parse_weights(weights_path)
+    return return_fn(scopes, values)
+
+
+def load_mobilenet140v2(scopes, return_fn=_assign):
+    """Converted from the [TF Slim][2]."""
+    filename = 'mobilenet_v2_140_224.npz'
+    weights_path = get_file(
+        filename, __model_url__ + 'mobilenet/' + filename,
+        cache_subdir='models',
+        file_hash='6988a66bb89a088ce20e2ae97adca88b')
+    values = parse_weights(weights_path)
+    return return_fn(scopes, values)
+
+
 def load_squeezenet(scopes, return_fn=_assign):
     """Converted from the [Caffe SqueezeNets][8]."""
     filename = 'squeezenet.npz'
@@ -704,6 +770,12 @@ def load_ref_faster_rcnn_vgg16_voc(scopes, return_fn=_assign):
     'mobilenet50': load_mobilenet50,
     'mobilenet75': load_mobilenet75,
     'mobilenet100': load_mobilenet100,
+    'mobilenet35v2': load_mobilenet35v2,
+    'mobilenet50v2': load_mobilenet50v2,
+    'mobilenet75v2': load_mobilenet75v2,
+    'mobilenet100v2': load_mobilenet100v2,
+    'mobilenet130v2': load_mobilenet130v2,
+    'mobilenet140v2': load_mobilenet140v2,
     'squeezenet': load_squeezenet,
     'zf': load_nothing,
     'darknet19': load_darknet19,