Add the Inception-ResNet-v2 model

image_classification/README.md

 图像分类
 =======================
 
-这里将介绍如何在PaddlePaddle下使用AlexNet、VGG、GoogLeNet和ResNet模型进行图像分类。图像分类问题的描述和这四种模型的介绍可以参考[PaddlePaddle book](https://github.com/PaddlePaddle/book/tree/develop/03.image_classification)。
+这里将介绍如何在PaddlePaddle下使用AlexNet、VGG、GoogLeNet、ResNet和Inception-ResNet-v2模型进行图像分类。图像分类问题的描述和这五种模型的介绍可以参考[PaddlePaddle book](https://github.com/PaddlePaddle/book/tree/develop/03.image_classification)。
 
 ## 训练模型
 
@@ -11,6 +11,8 @@
 
 ```python
 import gzip
+import argparse
+
 import paddle.v2.dataset.flowers as flowers
 import paddle.v2 as paddle
 import reader
@@ -18,6 +20,7 @@ import vgg
 import resnet
 import alexnet
 import googlenet
+import inception_resnet_v2
 
 
 # PaddlePaddle init
@@ -29,7 +32,7 @@ paddle.init(use_gpu=False, trainer_count=1)
 设置算法参数（如数据维度、类别数目和batch size等参数），定义数据输入层`image`和类别标签`lbl`。
 
 ```python
-DATA_DIM = 3 * 224 * 224
+DATA_DIM = 3 * 224 * 224  # Use 3 * 331 * 331 or 3 * 299 * 299 for Inception-ResNet-v2.
 CLASS_DIM = 102
 BATCH_SIZE = 128
 
@@ -41,7 +44,7 @@ lbl = paddle.layer.data(
 
 ### 获得所用模型
 
-这里可以选择使用AlexNet、VGG、GoogLeNet和ResNet模型中的一个模型进行图像分类。通过调用相应的方法可以获得网络最后的Softmax层。
+这里可以选择使用AlexNet、VGG、GoogLeNet、ResNet和Inception-ResNet-v2模型中的一个模型进行图像分类。通过调用相应的方法可以获得网络最后的Softmax层。
 
 1. 使用AlexNet模型
 
@@ -86,6 +89,16 @@ ResNet模型可以通过下面的代码获取：
 out = resnet.resnet_imagenet(image, class_dim=CLASS_DIM)
 ```
 
+5. 使用Inception-ResNet-v2模型
+
+提供的Inception-ResNet-v2模型支持`3 * 331 * 331`和`3 * 299 * 299`两种大小的输入，同时可以自行设置dropout概率，可以通过如下的代码使用：
+
+```python
+out = inception_resnet_v2.inception_resnet_v2(image, class_dim=CLASS_DIM, dropout_rate=0.5, size=DATA_DIM)
+```
+
+注意，由于和其他几种模型输入大小不同，若配合提供的`reader.py`使用Inception-ResNet-v2时请先将`reader.py`中`paddle.image.simple_transform`中的参数为修改为相应大小。
+
 ### 定义损失函数
 
 ```python
@@ -173,7 +186,7 @@ def event_handler(event):
 
 ### 定义训练方法
 
-对于AlexNet、VGG和ResNet，可以按下面的代码定义训练方法：
+对于AlexNet、VGG、ResNet和Inception-ResNet-v2，可以按下面的代码定义训练方法：
 
 ```python
 # Create trainer

image_classification/inception_resnet_v2.py

+import paddle.v2 as paddle
+
+
+def conv_bn_layer(input,
+                  ch_out,
+                  filter_size,
+                  stride,
+                  padding=0,
+                  active_type=paddle.activation.Relu(),
+                  ch_in=None):
+    tmp = paddle.layer.img_conv(
+        input=input,
+        filter_size=filter_size,
+        num_channels=ch_in,
+        num_filters=ch_out,
+        stride=stride,
+        padding=padding,
+        act=paddle.activation.Linear(),
+        bias_attr=False)
+    return paddle.layer.batch_norm(input=tmp, epsilon=0.001, act=active_type)
+
+
+def sequential_block(input, *layers):
+    for layer in layers:
+        layer_func, layer_conf = layer
+        input = layer_func(input, **layer_conf)
+    return input
+
+
+def mixed_5b_block(input):
+    branch0 = conv_bn_layer(
+        input, ch_in=192, ch_out=96, filter_size=1, stride=1)
+    branch1 = sequential_block(input, (conv_bn_layer, {
+        "ch_in": 192,
+        "ch_out": 48,
+        "filter_size": 1,
+        "stride": 1
+    }), (conv_bn_layer, {
+        "ch_in": 48,
+        "ch_out": 64,
+        "filter_size": 5,
+        "stride": 1,
+        "padding": 2
+    }))
+    branch2 = sequential_block(input, (conv_bn_layer, {
+        "ch_in": 192,
+        "ch_out": 64,
+        "filter_size": 1,
+        "stride": 1
+    }), (conv_bn_layer, {
+        "ch_in": 64,
+        "ch_out": 96,
+        "filter_size": 3,
+        "stride": 1,
+        "padding": 1
+    }), (conv_bn_layer, {
+        "ch_in": 96,
+        "ch_out": 96,
+        "filter_size": 3,
+        "stride": 1,
+        "padding": 1
+    }))
+    branch3 = sequential_block(
+        input,
+        (paddle.layer.img_pool, {
+            "pool_size": 3,
+            "stride": 1,
+            "padding": 1,
+            "pool_type": paddle.pooling.Avg(),
+            "exclude_mode": False
+        }),
+        (conv_bn_layer, {
+            "ch_in": 192,
+            "ch_out": 64,
+            "filter_size": 1,
+            "stride": 1
+        }), )
+    out = paddle.layer.concat(input=[branch0, branch1, branch2, branch3])
+    return out
+
+
+def block35(input, scale=1.0):
+    branch0 = conv_bn_layer(
+        input, ch_in=320, ch_out=32, filter_size=1, stride=1)
+    branch1 = sequential_block(input, (conv_bn_layer, {
+        "ch_in": 320,
+        "ch_out": 32,
+        "filter_size": 1,
+        "stride": 1
+    }), (conv_bn_layer, {
+        "ch_in": 32,
+        "ch_out": 32,
+        "filter_size": 3,
+        "stride": 1,
+        "padding": 1
+    }))
+    branch2 = sequential_block(input, (conv_bn_layer, {
+        "ch_in": 320,
+        "ch_out": 32,
+        "filter_size": 1,
+        "stride": 1
+    }), (conv_bn_layer, {
+        "ch_in": 32,
+        "ch_out": 48,
+        "filter_size": 3,
+        "stride": 1,
+        "padding": 1
+    }), (conv_bn_layer, {
+        "ch_in": 48,
+        "ch_out": 64,
+        "filter_size": 3,
+        "stride": 1,
+        "padding": 1
+    }))
+    out = paddle.layer.concat(input=[branch0, branch1, branch2])
+    out = paddle.layer.img_conv(
+        input=out,
+        filter_size=1,
+        num_channels=128,
+        num_filters=320,
+        stride=1,
+        padding=0,
+        act=paddle.activation.Linear(),
+        bias_attr=None)
+    out = paddle.layer.slope_intercept(out, slope=scale, intercept=0.0)
+    out = paddle.layer.addto(input=[input, out], act=paddle.activation.Relu())
+    return out
+
+
+def mixed_6a_block(input):
+    branch0 = conv_bn_layer(
+        input, ch_in=320, ch_out=384, filter_size=3, stride=2)
+    branch1 = sequential_block(input, (conv_bn_layer, {
+        "ch_in": 320,
+        "ch_out": 256,
+        "filter_size": 1,
+        "stride": 1
+    }), (conv_bn_layer, {
+        "ch_in": 256,
+        "ch_out": 256,
+        "filter_size": 3,
+        "stride": 1,
+        "padding": 1
+    }), (conv_bn_layer, {
+        "ch_in": 256,
+        "ch_out": 384,
+        "filter_size": 3,
+        "stride": 2
+    }))
+    branch2 = paddle.layer.img_pool(
+        input,
+        num_channels=320,
+        pool_size=3,
+        stride=2,
+        pool_type=paddle.pooling.Max())
+    out = paddle.layer.concat(input=[branch0, branch1, branch2])
+    return out
+
+
+def block17(input, scale=1.0):
+    branch0 = conv_bn_layer(
+        input, ch_in=1088, ch_out=192, filter_size=1, stride=1)
+    branch1 = sequential_block(input, (conv_bn_layer, {
+        "ch_in": 1088,
+        "ch_out": 128,
+        "filter_size": 1,
+        "stride": 1
+    }), (conv_bn_layer, {
+        "ch_in": 128,
+        "ch_out": 160,
+        "filter_size": [7, 1],
+        "stride": 1,
+        "padding": [3, 0]
+    }), (conv_bn_layer, {
+        "ch_in": 160,
+        "ch_out": 192,
+        "filter_size": [1, 7],
+        "stride": 1,
+        "padding": [0, 3]
+    }))
+    out = paddle.layer.concat(input=[branch0, branch1])
+    out = paddle.layer.img_conv(
+        input=out,
+        filter_size=1,
+        num_channels=384,
+        num_filters=1088,
+        stride=1,
+        padding=0,
+        act=paddle.activation.Linear(),
+        bias_attr=None)
+    out = paddle.layer.slope_intercept(out, slope=scale, intercept=0.0)
+    out = paddle.layer.addto(input=[input, out], act=paddle.activation.Relu())
+    return out
+
+
+def mixed_7a_block(input):
+    branch0 = sequential_block(
+        input,
+        (conv_bn_layer, {
+            "ch_in": 1088,
+            "ch_out": 256,
+            "filter_size": 1,
+            "stride": 1
+        }),
+        (conv_bn_layer, {
+            "ch_in": 256,
+            "ch_out": 384,
+            "filter_size": 3,
+            "stride": 2
+        }), )
+    branch1 = sequential_block(
+        input,
+        (conv_bn_layer, {
+            "ch_in": 1088,
+            "ch_out": 256,
+            "filter_size": 1,
+            "stride": 1
+        }),
+        (conv_bn_layer, {
+            "ch_in": 256,
+            "ch_out": 288,
+            "filter_size": 3,
+            "stride": 2
+        }), )
+    branch2 = sequential_block(input, (conv_bn_layer, {
+        "ch_in": 1088,
+        "ch_out": 256,
+        "filter_size": 1,
+        "stride": 1
+    }), (conv_bn_layer, {
+        "ch_in": 256,
+        "ch_out": 288,
+        "filter_size": 3,
+        "stride": 1,
+        "padding": 1
+    }), (conv_bn_layer, {
+        "ch_in": 288,
+        "ch_out": 320,
+        "filter_size": 3,
+        "stride": 2
+    }))
+    branch3 = paddle.layer.img_pool(
+        input,
+        num_channels=1088,
+        pool_size=3,
+        stride=2,
+        pool_type=paddle.pooling.Max())
+    out = paddle.layer.concat(input=[branch0, branch1, branch2, branch3])
+    return out
+
+
+def block8(input, scale=1.0, no_relu=False):
+    branch0 = conv_bn_layer(
+        input, ch_in=2080, ch_out=192, filter_size=1, stride=1)
+    branch1 = sequential_block(input, (conv_bn_layer, {
+        "ch_in": 2080,
+        "ch_out": 192,
+        "filter_size": 1,
+        "stride": 1
+    }), (conv_bn_layer, {
+        "ch_in": 192,
+        "ch_out": 224,
+        "filter_size": [3, 1],
+        "stride": 1,
+        "padding": [1, 0]
+    }), (conv_bn_layer, {
+        "ch_in": 224,
+        "ch_out": 256,
+        "filter_size": [1, 3],
+        "stride": 1,
+        "padding": [0, 1]
+    }))
+    out = paddle.layer.concat(input=[branch0, branch1])
+    out = paddle.layer.img_conv(
+        input=out,
+        filter_size=1,
+        num_channels=448,
+        num_filters=2080,
+        stride=1,
+        padding=0,
+        act=paddle.activation.Linear(),
+        bias_attr=None)
+    out = paddle.layer.slope_intercept(out, slope=scale, intercept=0.0)
+    out = paddle.layer.addto(
+        input=[input, out],
+        act=paddle.activation.Linear() if no_relu else paddle.activation.Relu())
+    return out
+
+
+def inception_resnet_v2(input,
+                        class_dim,
+                        dropout_rate=0.5,
+                        data_dim=3 * 331 * 331):
+    conv2d_1a = conv_bn_layer(
+        input, ch_in=3, ch_out=32, filter_size=3, stride=2)
+    conv2d_2a = conv_bn_layer(
+        conv2d_1a, ch_in=32, ch_out=32, filter_size=3, stride=1)
+    conv2d_2b = conv_bn_layer(
+        conv2d_2a, ch_in=32, ch_out=64, filter_size=3, stride=1, padding=1)
+    maxpool_3a = paddle.layer.img_pool(
+        input=conv2d_2b, pool_size=3, stride=2, pool_type=paddle.pooling.Max())
+    conv2d_3b = conv_bn_layer(
+        maxpool_3a, ch_in=64, ch_out=80, filter_size=1, stride=1)
+    conv2d_4a = conv_bn_layer(
+        conv2d_3b, ch_in=80, ch_out=192, filter_size=3, stride=1)
+    maxpool_5a = paddle.layer.img_pool(
+        input=conv2d_4a, pool_size=3, stride=2, pool_type=paddle.pooling.Max())
+    mixed_5b = mixed_5b_block(maxpool_5a)
+    repeat = sequential_block(mixed_5b, *([(block35, {"scale": 0.17})] * 10))
+    mixed_6a = mixed_6a_block(repeat)
+    repeat1 = sequential_block(mixed_6a, *([(block17, {"scale": 0.10})] * 20))
+    mixed_7a = mixed_7a_block(repeat1)
+    repeat2 = sequential_block(mixed_7a, *([(block8, {"scale": 0.20})] * 9))
+    block_8 = block8(repeat2, no_relu=True)
+    conv2d_7b = conv_bn_layer(
+        block_8, ch_in=2080, ch_out=1536, filter_size=1, stride=1)
+    avgpool_1a = paddle.layer.img_pool(
+        input=conv2d_7b,
+        pool_size=8 if data_dim == 3 * 299 * 299 else 9,
+        stride=1,
+        pool_type=paddle.pooling.Avg(),
+        exclude_mode=False)
+    drop_out = paddle.layer.dropout(input=avgpool_1a, dropout_rate=dropout_rate)
+    out = paddle.layer.fc(
+        input=drop_out, size=class_dim, act=paddle.activation.Softmax())
+    return out

image_classification/infer.py

+import os
 import gzip
+import argparse
+import numpy as np
+from PIL import Image
+
 import paddle.v2 as paddle
 import reader
 import vgg
 import resnet
 import alexnet
 import googlenet
-import argparse
-import os
-from PIL import Image
-import numpy as np
+import inception_resnet_v2
 
-WIDTH = 224
-HEIGHT = 224
-DATA_DIM = 3 * WIDTH * HEIGHT
+DATA_DIM = 3 * 224 * 224  # Use 3 * 331 * 331 or 3 * 299 * 299 for Inception-ResNet-v2.
 CLASS_DIM = 102
 
 
@@ -26,7 +26,10 @@ def main():
     parser.add_argument(
         'model',
         help='The model for image classification',
-        choices=['alexnet', 'vgg13', 'vgg16', 'vgg19', 'resnet', 'googlenet'])
+        choices=[
+            'alexnet', 'vgg13', 'vgg16', 'vgg19', 'resnet', 'googlenet',
+            'inception-resnet-v2'
+        ])
     parser.add_argument(
         'params_path', help='The file which stores the parameters')
     args = parser.parse_args()
@@ -49,6 +52,10 @@ def main():
         out = resnet.resnet_imagenet(image, class_dim=CLASS_DIM)
     elif args.model == 'googlenet':
         out, _, _ = googlenet.googlenet(image, class_dim=CLASS_DIM)
+    elif args.model == 'inception-resnet-v2':
+        assert DATA_DIM == 3 * 331 * 331 or DATA_DIM == 3 * 299 * 299
+        out = inception_resnet_v2.inception_resnet_v2(
+            image, class_dim=CLASS_DIM, dropout_rate=0.5, data_dim=DATA_DIM)
 
     # load parameters
     with gzip.open(args.params_path, 'r') as f:

image_classification/train.py

 import gzip
+import argparse
+
 import paddle.v2.dataset.flowers as flowers
 import paddle.v2 as paddle
 import reader
@@ -6,9 +8,9 @@ import vgg
 import resnet
 import alexnet
 import googlenet
-import argparse
+import inception_resnet_v2
 
-DATA_DIM = 3 * 224 * 224
+DATA_DIM = 3 * 224 * 224  # Use 3 * 331 * 331 or 3 * 299 * 299 for Inception-ResNet-v2.
 CLASS_DIM = 102
 BATCH_SIZE = 128
 
@@ -19,7 +21,10 @@ def main():
     parser.add_argument(
         'model',
         help='The model for image classification',
-        choices=['alexnet', 'vgg13', 'vgg16', 'vgg19', 'resnet', 'googlenet'])
+        choices=[
+            'alexnet', 'vgg13', 'vgg16', 'vgg19', 'resnet', 'googlenet',
+            'inception-resnet-v2'
+        ])
     args = parser.parse_args()
 
     # PaddlePaddle init
@@ -52,6 +57,10 @@ def main():
             input=out2, label=lbl, coeff=0.3)
         paddle.evaluator.classification_error(input=out2, label=lbl)
         extra_layers = [loss1, loss2]
+    elif args.model == 'inception-resnet-v2':
+        assert DATA_DIM == 3 * 331 * 331 or DATA_DIM == 3 * 299 * 299
+        out = inception_resnet_v2.inception_resnet_v2(
+            image, class_dim=CLASS_DIM, dropout_rate=0.5, data_dim=DATA_DIM)
 
     cost = paddle.layer.classification_cost(input=out, label=lbl)