Merge pull request #539 from will-am/inception_v4

Implement Inception-v4 for image classification

README.cn.md

@@ -98,14 +98,15 @@ PaddlePaddle提供了丰富的运算单元，帮助大家以模块化的方式
 
 图像相比文字能够提供更加生动、容易理解及更具艺术感的信息，是人们转递与交换信息的重要来源。图像分类是根据图像的语义信息对不同类别图像进行区分，是计算机视觉中重要的基础问题，也是图像检测、图像分割、物体跟踪、行为分析等其他高层视觉任务的基础，在许多领域都有着广泛的应用。如：安防领域的人脸识别和智能视频分析等，交通领域的交通场景识别，互联网领域基于内容的图像检索和相册自动归类，医学领域的图像识别等。
 
-在图像分类任务中，我们向大家介绍如何训练AlexNet、VGG、GoogLeNet、ResNet和Inception-Resnet-V2模型。同时提供了能够将Caffe或TensorFlow训练好的模型文件转换为PaddlePaddle模型文件的模型转换工具。
+在图像分类任务中，我们向大家介绍如何训练AlexNet、VGG、GoogLeNet、ResNet、Inception-v4和Inception-Resnet-V2模型。同时提供了能够将Caffe或TensorFlow训练好的模型文件转换为PaddlePaddle模型文件的模型转换工具。
 
 - 11.1 [将Caffe模型文件转换为PaddlePaddle模型文件](https://github.com/PaddlePaddle/models/tree/develop/image_classification/caffe2paddle)
 - 11.2 [将TensorFlow模型文件转换为PaddlePaddle模型文件](https://github.com/PaddlePaddle/models/tree/develop/image_classification/tf2paddle)
 - 11.3 [AlexNet](https://github.com/PaddlePaddle/models/tree/develop/image_classification)
 - 11.4 [VGG](https://github.com/PaddlePaddle/models/tree/develop/image_classification)
 - 11.5 [Residual Network](https://github.com/PaddlePaddle/models/tree/develop/image_classification)
-- 11.6 [Inception-Resnet-V2](https://github.com/PaddlePaddle/models/tree/develop/image_classification)
+- 11.6 [Inception-v4](https://github.com/PaddlePaddle/models/tree/develop/image_classification)
+- 11.7 [Inception-Resnet-V2](https://github.com/PaddlePaddle/models/tree/develop/image_classification)
 
 ## 12. 目标检测
 

README.md

@@ -72,13 +72,14 @@ As an example for sequence-to-sequence learning, we take the machine translation
 
 ## 9. Image classification
 
-For the example of image classification, we show you how to train AlexNet, VGG, GoogLeNet, ResNet and Inception-Resnet-V2 models in PaddlePaddle. It also provides model conversion tools that convert Caffe or TensorFlow trained model files into PaddlePaddle model files.
+For the example of image classification, we show you how to train AlexNet, VGG, GoogLeNet, ResNet, Inception-v4 and Inception-Resnet-V2 models in PaddlePaddle. It also provides model conversion tools that convert Caffe or TensorFlow trained model files into PaddlePaddle model files.
 
 - 9.1 [convert Caffe model file to PaddlePaddle model file](https://github.com/PaddlePaddle/models/tree/develop/image_classification/caffe2paddle)
 - 9.2 [convert TensorFlow model file to PaddlePaddle model file](https://github.com/PaddlePaddle/models/tree/develop/image_classification/tf2paddle)
 - 9.3 [AlexNet](https://github.com/PaddlePaddle/models/tree/develop/image_classification)
 - 9.4 [VGG](https://github.com/PaddlePaddle/models/tree/develop/image_classification)
 - 9.5 [Residual Network](https://github.com/PaddlePaddle/models/tree/develop/image_classification)
-- 9.6 [Inception-Resnet-V2](https://github.com/PaddlePaddle/models/tree/develop/image_classification)
+- 9.6 [Inception-v4](https://github.com/PaddlePaddle/models/tree/develop/image_classification)
+- 9.7 [Inception-Resnet-V2](https://github.com/PaddlePaddle/models/tree/develop/image_classification)
 
 This tutorial is contributed by [PaddlePaddle](https://github.com/PaddlePaddle/Paddle) and licensed under the [Apache-2.0 license](LICENSE).

image_classification/README.md

 图像分类
 =======================
 
-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)。
+这里将介绍如何在PaddlePaddle下使用AlexNet、VGG、GoogLeNet、ResNet、Inception-v4和Inception-ResNet-v2模型进行图像分类。图像分类问题的描述和这些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
 
@@ -20,6 +20,7 @@ import vgg
 import resnet
 import alexnet
 import googlenet
+import inception_v4
 import inception_resnet_v2
 
 
@@ -45,7 +46,7 @@ lbl = paddle.layer.data(
 
 ### 获得所用模型
 
-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。
+这里可以选择使用AlexNet、VGG、GoogLeNet、ResNet、Inception-v4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
 
@@ -90,7 +91,16 @@ 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
+5. 使用Inception-v4模型
+
+Inception-v4模型可以通过下面的代码获取, 本例中使用的模型输入大小为`3 * 224 * 224` (原文献中使用的输入大小为`3 * 299 * 299`)：
+
+```python
+out = inception_v4.inception_v4(image, class_dim=CLASS_DIM)
+```
+
+
+6. 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：
 
@@ -188,7 +198,7 @@ def event_handler(event):
 
 ### 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：
+对于AlexNet、VGG、ResNet、Inception-v4和Inception-ResNet-v2，可以按下面的代码定义训练方法：
 
 ```python
 # Create trainer

image_classification/inception_v4.py

+import paddle.v2 as paddle
+
+__all__ = ['inception_v4']
+
+
+def img_conv(name,
+             input,
+             num_filters,
+             filter_size,
+             stride,
+             padding,
+             num_channels=None):
+    conv = paddle.layer.img_conv(
+        name=name,
+        input=input,
+        num_channels=num_channels,
+        num_filters=num_filters,
+        filter_size=filter_size,
+        stride=stride,
+        padding=padding,
+        act=paddle.activation.Linear())
+    norm = paddle.layer.batch_norm(
+        name=name + '_norm', input=conv, act=paddle.activation.Relu())
+    return norm
+
+
+def stem(input):
+    conv0 = img_conv(
+        name='stem_conv0',
+        input=input,
+        num_channels=3,
+        num_filters=32,
+        filter_size=3,
+        stride=2,
+        padding=1)
+    conv1 = img_conv(
+        name='stem_conv1',
+        input=conv0,
+        num_channels=32,
+        num_filters=32,
+        filter_size=3,
+        stride=1,
+        padding=1)
+    conv2 = img_conv(
+        name='stem_conv2',
+        input=conv1,
+        num_channels=32,
+        num_filters=64,
+        filter_size=3,
+        stride=1,
+        padding=1)
+
+    def block0(input):
+        pool0 = paddle.layer.img_pool(
+            name='stem_branch0_pool0',
+            input=input,
+            num_channels=64,
+            pool_size=3,
+            stride=2,
+            pool_type=paddle.pooling.Max())
+        conv0 = img_conv(
+            name='stem_branch0_conv0',
+            input=input,
+            num_channels=64,
+            num_filters=96,
+            filter_size=3,
+            stride=2,
+            padding=1)
+        return paddle.layer.concat(input=[pool0, conv0])
+
+    def block1(input):
+        l_conv0 = img_conv(
+            name='stem_branch1_l_conv0',
+            input=input,
+            num_channels=160,
+            num_filters=64,
+            filter_size=1,
+            stride=1,
+            padding=0)
+        l_conv1 = img_conv(
+            name='stem_branch1_l_conv1',
+            input=l_conv0,
+            num_channels=64,
+            num_filters=96,
+            filter_size=3,
+            stride=1,
+            padding=1)
+        r_conv0 = img_conv(
+            name='stem_branch1_r_conv0',
+            input=input,
+            num_channels=160,
+            num_filters=64,
+            filter_size=1,
+            stride=1,
+            padding=0)
+        r_conv1 = img_conv(
+            name='stem_branch1_r_conv1',
+            input=r_conv0,
+            num_channels=64,
+            num_filters=64,
+            filter_size=(7, 1),
+            stride=1,
+            padding=(3, 0))
+        r_conv2 = img_conv(
+            name='stem_branch1_r_conv2',
+            input=r_conv1,
+            num_channels=64,
+            num_filters=64,
+            filter_size=(1, 7),
+            stride=1,
+            padding=(0, 3))
+        r_conv3 = img_conv(
+            name='stem_branch1_r_conv3',
+            input=r_conv2,
+            num_channels=64,
+            num_filters=96,
+            filter_size=3,
+            stride=1,
+            padding=1)
+        return paddle.layer.concat(input=[l_conv1, r_conv3])
+
+    def block2(input):
+        conv0 = img_conv(
+            name='stem_branch2_conv0',
+            input=input,
+            num_channels=192,
+            num_filters=192,
+            filter_size=3,
+            stride=2,
+            padding=1)
+        pool0 = paddle.layer.img_pool(
+            name='stem_branch2_pool0',
+            input=input,
+            num_channels=192,
+            pool_size=3,
+            stride=2,
+            pool_type=paddle.pooling.Max())
+        return paddle.layer.concat(input=[conv0, pool0])
+
+    conv3 = block0(conv2)
+    conv4 = block1(conv3)
+    conv5 = block2(conv4)
+    return conv5
+
+
+def Inception_A(input, depth):
+    b0_pool0 = paddle.layer.img_pool(
+        name='inceptA{0}_branch0_pool0'.format(depth),
+        input=input,
+        num_channels=384,
+        pool_size=3,
+        stride=1,
+        padding=1,
+        pool_type=paddle.pooling.Avg())
+    b0_conv0 = img_conv(
+        name='inceptA{0}_branch0_conv0'.format(depth),
+        input=b0_pool0,
+        num_channels=384,
+        num_filters=96,
+        filter_size=1,
+        stride=1,
+        padding=0)
+    b1_conv0 = img_conv(
+        name='inceptA{0}_branch1_conv0'.format(depth),
+        input=input,
+        num_channels=384,
+        num_filters=96,
+        filter_size=1,
+        stride=1,
+        padding=0)
+    b2_conv0 = img_conv(
+        name='inceptA{0}_branch2_conv0'.format(depth),
+        input=input,
+        num_channels=384,
+        num_filters=64,
+        filter_size=1,
+        stride=1,
+        padding=0)
+    b2_conv1 = img_conv(
+        name='inceptA{0}_branch2_conv1'.format(depth),
+        input=b2_conv0,
+        num_channels=64,
+        num_filters=96,
+        filter_size=3,
+        stride=1,
+        padding=1)
+    b3_conv0 = img_conv(
+        name='inceptA{0}_branch3_conv0'.format(depth),
+        input=input,
+        num_channels=384,
+        num_filters=64,
+        filter_size=1,
+        stride=1,
+        padding=0)
+    b3_conv1 = img_conv(
+        name='inceptA{0}_branch3_conv1'.format(depth),
+        input=b3_conv0,
+        num_channels=64,
+        num_filters=96,
+        filter_size=3,
+        stride=1,
+        padding=1)
+    b3_conv2 = img_conv(
+        name='inceptA{0}_branch3_conv2'.format(depth),
+        input=b3_conv1,
+        num_channels=96,
+        num_filters=96,
+        filter_size=3,
+        stride=1,
+        padding=1)
+    return paddle.layer.concat(input=[b0_conv0, b1_conv0, b2_conv1, b3_conv2])
+
+
+def Inception_B(input, depth):
+    b0_pool0 = paddle.layer.img_pool(
+        name='inceptB{0}_branch0_pool0'.format(depth),
+        input=input,
+        num_channels=1024,
+        pool_size=3,
+        stride=1,
+        padding=1,
+        pool_type=paddle.pooling.Avg())
+    b0_conv0 = img_conv(
+        name='inceptB{0}_branch0_conv0'.format(depth),
+        input=b0_pool0,
+        num_channels=1024,
+        num_filters=128,
+        filter_size=1,
+        stride=1,
+        padding=0)
+    b1_conv0 = img_conv(
+        name='inceptB{0}_branch1_conv0'.format(depth),
+        input=input,
+        num_channels=1024,
+        num_filters=384,
+        filter_size=1,
+        stride=1,
+        padding=0)
+    b2_conv0 = img_conv(
+        name='inceptB{0}_branch2_conv0'.format(depth),
+        input=input,
+        num_channels=1024,
+        num_filters=192,
+        filter_size=1,
+        stride=1,
+        padding=0)
+    b2_conv1 = img_conv(
+        name='inceptB{0}_branch2_conv1'.format(depth),
+        input=b2_conv0,
+        num_channels=192,
+        num_filters=224,
+        filter_size=(1, 7),
+        stride=1,
+        padding=(0, 3))
+    b2_conv2 = img_conv(
+        name='inceptB{0}_branch2_conv2'.format(depth),
+        input=b2_conv1,
+        num_channels=224,
+        num_filters=256,
+        filter_size=(7, 1),
+        stride=1,
+        padding=(3, 0))
+    b3_conv0 = img_conv(
+        name='inceptB{0}_branch3_conv0'.format(depth),
+        input=input,
+        num_channels=1024,
+        num_filters=192,
+        filter_size=1,
+        stride=1,
+        padding=0)
+    b3_conv1 = img_conv(
+        name='inceptB{0}_branch3_conv1'.format(depth),
+        input=b3_conv0,
+        num_channels=192,
+        num_filters=192,
+        filter_size=(1, 7),
+        stride=1,
+        padding=(0, 3))
+    b3_conv2 = img_conv(
+        name='inceptB{0}_branch3_conv2'.format(depth),
+        input=b3_conv1,
+        num_channels=192,
+        num_filters=224,
+        filter_size=(7, 1),
+        stride=1,
+        padding=(3, 0))
+    b3_conv3 = img_conv(
+        name='inceptB{0}_branch3_conv3'.format(depth),
+        input=b3_conv2,
+        num_channels=224,
+        num_filters=224,
+        filter_size=(1, 7),
+        stride=1,
+        padding=(0, 3))
+    b3_conv4 = img_conv(
+        name='inceptB{0}_branch3_conv4'.format(depth),
+        input=b3_conv3,
+        num_channels=224,
+        num_filters=256,
+        filter_size=(7, 1),
+        stride=1,
+        padding=(3, 0))
+    return paddle.layer.concat(input=[b0_conv0, b1_conv0, b2_conv2, b3_conv4])
+
+
+def Inception_C(input, depth):
+    b0_pool0 = paddle.layer.img_pool(
+        name='inceptC{0}_branch0_pool0'.format(depth),
+        input=input,
+        num_channels=1536,
+        pool_size=3,
+        stride=1,
+        padding=1,
+        pool_type=paddle.pooling.Avg())
+    b0_conv0 = img_conv(
+        name='inceptC{0}_branch0_conv0'.format(depth),
+        input=b0_pool0,
+        num_channels=1536,
+        num_filters=256,
+        filter_size=1,
+        stride=1,
+        padding=0)
+    b1_conv0 = img_conv(
+        name='inceptC{0}_branch1_conv0'.format(depth),
+        input=input,
+        num_channels=1536,
+        num_filters=256,
+        filter_size=1,
+        stride=1,
+        padding=0)
+    b2_conv0 = img_conv(
+        name='inceptC{0}_branch2_conv0'.format(depth),
+        input=input,
+        num_channels=1536,
+        num_filters=384,
+        filter_size=1,
+        stride=1,
+        padding=0)
+    b2_conv1 = img_conv(
+        name='inceptC{0}_branch2_conv1'.format(depth),
+        input=b2_conv0,
+        num_channels=384,
+        num_filters=256,
+        filter_size=(1, 3),
+        stride=1,
+        padding=(0, 1))
+    b2_conv2 = img_conv(
+        name='inceptC{0}_branch2_conv2'.format(depth),
+        input=b2_conv0,
+        num_channels=384,
+        num_filters=256,
+        filter_size=(3, 1),
+        stride=1,
+        padding=(1, 0))
+    b3_conv0 = img_conv(
+        name='inceptC{0}_branch3_conv0'.format(depth),
+        input=input,
+        num_channels=1536,
+        num_filters=384,
+        filter_size=1,
+        stride=1,
+        padding=0)
+    b3_conv1 = img_conv(
+        name='inceptC{0}_branch3_conv1'.format(depth),
+        input=b3_conv0,
+        num_channels=384,
+        num_filters=448,
+        filter_size=(1, 3),
+        stride=1,
+        padding=(0, 1))
+    b3_conv2 = img_conv(
+        name='inceptC{0}_branch3_conv2'.format(depth),
+        input=b3_conv1,
+        num_channels=448,
+        num_filters=512,
+        filter_size=(3, 1),
+        stride=1,
+        padding=(1, 0))
+    b3_conv3 = img_conv(
+        name='inceptC{0}_branch3_conv3'.format(depth),
+        input=b3_conv2,
+        num_channels=512,
+        num_filters=256,
+        filter_size=(3, 1),
+        stride=1,
+        padding=(1, 0))
+    b3_conv4 = img_conv(
+        name='inceptC{0}_branch3_conv4'.format(depth),
+        input=b3_conv2,
+        num_channels=512,
+        num_filters=256,
+        filter_size=(1, 3),
+        stride=1,
+        padding=(0, 1))
+    return paddle.layer.concat(
+        input=[b0_conv0, b1_conv0, b2_conv1, b2_conv2, b3_conv3, b3_conv4])
+
+
+def Reduction_A(input):
+    b0_pool0 = paddle.layer.img_pool(
+        name='ReductA_branch0_pool0',
+        input=input,
+        num_channels=384,
+        pool_size=3,
+        stride=2,
+        pool_type=paddle.pooling.Max())
+    b1_conv0 = img_conv(
+        name='ReductA_branch1_conv0',
+        input=input,
+        num_channels=384,
+        num_filters=384,
+        filter_size=3,
+        stride=2,
+        padding=1)
+    b2_conv0 = img_conv(
+        name='ReductA_branch2_conv0',
+        input=input,
+        num_channels=384,
+        num_filters=192,
+        filter_size=1,
+        stride=1,
+        padding=0)
+    b2_conv1 = img_conv(
+        name='ReductA_branch2_conv1',
+        input=b2_conv0,
+        num_channels=192,
+        num_filters=224,
+        filter_size=3,
+        stride=1,
+        padding=1)
+    b2_conv2 = img_conv(
+        name='ReductA_branch2_conv2',
+        input=b2_conv1,
+        num_channels=224,
+        num_filters=256,
+        filter_size=3,
+        stride=2,
+        padding=1)
+    return paddle.layer.concat(input=[b0_pool0, b1_conv0, b2_conv2])
+
+
+def Reduction_B(input):
+    b0_pool0 = paddle.layer.img_pool(
+        name='ReductB_branch0_pool0',
+        input=input,
+        num_channels=1024,
+        pool_size=3,
+        stride=2,
+        pool_type=paddle.pooling.Max())
+    b1_conv0 = img_conv(
+        name='ReductB_branch1_conv0',
+        input=input,
+        num_channels=1024,
+        num_filters=192,
+        filter_size=1,
+        stride=1,
+        padding=0)
+    b1_conv1 = img_conv(
+        name='ReductB_branch1_conv1',
+        input=b1_conv0,
+        num_channels=192,
+        num_filters=192,
+        filter_size=3,
+        stride=2,
+        padding=1)
+    b2_conv0 = img_conv(
+        name='ReductB_branch2_conv0',
+        input=input,
+        num_channels=1024,
+        num_filters=256,
+        filter_size=1,
+        stride=1,
+        padding=0)
+    b2_conv1 = img_conv(
+        name='ReductB_branch2_conv1',
+        input=b2_conv0,
+        num_channels=256,
+        num_filters=256,
+        filter_size=(1, 7),
+        stride=1,
+        padding=(0, 3))
+    b2_conv2 = img_conv(
+        name='ReductB_branch2_conv2',
+        input=b2_conv1,
+        num_channels=256,
+        num_filters=320,
+        filter_size=(7, 1),
+        stride=1,
+        padding=(3, 0))
+    b2_conv3 = img_conv(
+        name='ReductB_branch2_conv3',
+        input=b2_conv2,
+        num_channels=320,
+        num_filters=320,
+        filter_size=3,
+        stride=2,
+        padding=1)
+    return paddle.layer.concat(input=[b0_pool0, b1_conv1, b2_conv3])
+
+
+def inception_v4(input, class_dim):
+    conv = stem(input)
+
+    for i in range(4):
+        conv = Inception_A(conv, i)
+    conv = Reduction_A(conv)
+    for i in range(7):
+        conv = Inception_B(conv, i)
+    conv = Reduction_B(conv)
+    for i in range(3):
+        conv = Inception_C(conv, i)
+
+    pool = paddle.layer.img_pool(
+        name='incept_avg_pool',
+        input=conv,
+        num_channels=1536,
+        pool_size=7,
+        stride=1,
+        pool_type=paddle.pooling.Avg())
+    drop = paddle.layer.dropout(input=pool, dropout_rate=0.2)
+    out = paddle.layer.fc(
+        name='incept_fc',
+        input=drop,
+        size=class_dim,
+        act=paddle.activation.Softmax())
+    return out

image_classification/infer.py

@@ -10,6 +10,7 @@ import vgg
 import resnet
 import alexnet
 import googlenet
+import inception_v4
 import inception_resnet_v2
 
 DATA_DIM = 3 * 224 * 224  # Use 3 * 331 * 331 or 3 * 299 * 299 for Inception-ResNet-v2.
@@ -28,7 +29,7 @@ def main():
         help='The model for image classification',
         choices=[
             'alexnet', 'vgg13', 'vgg16', 'vgg19', 'resnet', 'googlenet',
-            'inception-resnet-v2'
+            'inception-resnet-v2', 'inception_v4'
         ])
     parser.add_argument(
         'params_path', help='The file which stores the parameters')
@@ -56,6 +57,8 @@ def main():
         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)
+    elif args.model == 'inception_v4':
+        out = inception_v4.inception_v4(image, class_dim=CLASS_DIM)
 
     # load parameters
     with gzip.open(args.params_path, 'r') as f:

image_classification/train.py

@@ -8,6 +8,7 @@ import vgg
 import resnet
 import alexnet
 import googlenet
+import inception_v4
 import inception_resnet_v2
 
 DATA_DIM = 3 * 224 * 224  # Use 3 * 331 * 331 or 3 * 299 * 299 for Inception-ResNet-v2.
@@ -23,7 +24,7 @@ def main():
         help='The model for image classification',
         choices=[
             'alexnet', 'vgg13', 'vgg16', 'vgg19', 'resnet', 'googlenet',
-            'inception-resnet-v2'
+            'inception-resnet-v2', 'inception_v4'
         ])
     args = parser.parse_args()
 
@@ -61,6 +62,8 @@ def main():
         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)
+    elif args.model == 'inception_v4':
+        out = inception_v4.inception_v4(image, class_dim=CLASS_DIM)
 
     cost = paddle.layer.classification_cost(input=out, label=lbl)