1. 运行环境
Python 2.7
pip install scikit-learn
pip install Pillow
pip install keras==1.2.1 （有版本要求，最好与这个一致）
pip install theano==0.9 （之前是1.0.4会报错）
1.1 keras中获取shape的正确方法
在keras的网络中，如果用layer_name.shape的方式获取shape信息将会返还tensorflow.python.framework.tensor_shape.TensorShape其中包含的是tensorflow.python.framework.tensor_shape.Dimension
正确的方式是使用
import keras.backend as K
K.int_shape(laye_name)
2. 代码解析
2.1 语句解析
代码：model_raw = eval(net)(input_tensor=inputs, include_top=False, weights='imagenet')
上一句代码执行时会下载如下文件（可以直接下载好，放到对应文件夹下）：
Downloading data from https://github.com/fchollet/deep-learning-models/releases/download/v0.1/vgg16_weights_th_dim_ordering_th_kernels_notop.h5
文件夹目录：C:\Users\xpb.keras\models
2.2
Layer (type) Output Shape Param # Connected to
====================================================================================================
input_1 (InputLayer) (None, 3, 448, 448) 0
____________________________________________________________________________________________________
block1_conv1 (Convolution2D) (None, 64, 448, 448) 1792 input_1[0][0]
____________________________________________________________________________________________________
..........................省略普通的网络结构.......................
____________________________________________________________________________________________________
block5_conv3 (Convolution2D) (None, 512, 28, 28) 2359808 block5_conv2[0][0]
____________________________________________________________________________________________________
block5_pool (MaxPooling2D) (None, 512, 14, 14) 0 block5_conv3[0][0]
____________________________________________________________________________________________________
reshape_layer (Reshape) (None, 512, 196) 0 block5_pool[0][0]
____________________________________________________________________________________________________
permute_1 (Permute) (None, 196, 512) 0 reshape_layer[0][0]
____________________________________________________________________________________________________
convolution1d_1 (Convolution1D) (None, 196, 512) 262656 permute_1[0][0]
____________________________________________________________________________________________________
activation_57 (Activation) (None, 196, 512) 0 batchnormalization_57[0][0]
____________________________________________________________________________________________________
p0_avg_pool (GlobalAveragePoolin (None, 512) 0 activation_1[0][0]
____________________________________________________________________________________________________
attr0_avg_pool (GlobalAveragePoo (None, 512) 0 activation_3[0][0]
____________________________________________________________________________________________________
attr1_avg_pool (GlobalAveragePoo (None, 512) 0 activation_5[0][0]
region_attention (Activation) (None, 196)
3.
3.1 keras和TensorFlow版本不一致
ImportError: Keras requires TensorFlow 2.2 or higher. Install TensorFlow via pip install tensorflow
TensorFlow 1.10
pip install keras==2.2.0
pip install keras==1.2.1
3.2
ValueError: Negative dimension size caused by subtracting 2 from 1 for 'block2_pool/MaxPool' (op: 'MaxPool') with input shapes: [?,1,224,128].
3.3 如何将python的keras backend换为theano
keras backend默认为TensorFlow ，换为theano，
import os
os.environ['KERAS_BACKEND']='theano'
import keras
import keras.backend as K
K.set_image_dim_ordering('th')
3.4
IOError: Unable to open file (truncated file: eof = 294912, sblock->base_addr = 0, stored_eof = 58889096)
原因是，下载模型的时候未完成，重新下载就行了。
解决方案
找到.keras/model这个文件夹，将下载未完成的.h5文件删除
我的路径为C:\Users.keras\models
3.5
Traceback (most recent call last):
File "F:/PycharmProjects/attribute-aware-attention-master/cub_demo.py", line 95, in
id_prob,id_pool,id_fea_map = init_classification(share_fea_map, emb_dim, nb_classes, name='p0')
File "F:/PycharmProjects/attribute-aware-attention-master/cub_demo.py", line 66, in init_classification
fea_map = BatchNormalization(axis=2)(fea_map)
File "C:\Users\xpb\Anaconda3\envs\Python27\lib\site-packages\keras\engine\topology.py", line 572, in __call__
self.add_inbound_node(inbound_layers, node_indices, tensor_indices)
File "C:\Users\xpb\Anaconda3\envs\Python27\lib\site-packages\keras\engine\topology.py", line 635, in add_inbound_node
Node.create_node(self, inbound_layers, node_indices, tensor_indices)
File "C:\Users\xpb\Anaconda3\envs\Python27\lib\site-packages\keras\engine\topology.py", line 166, in create_node
output_tensors = to_list(outbound_layer.call(input_tensors[0], mask=input_masks[0]))
File "C:\Users\xpb\Anaconda3\envs\Python27\lib\site-packages\keras\layers\normalization.py", line 143, in call
x_normed = K.in_train_phase(x_normed, x_normed_running)
File "C:\Users\xpb\Anaconda3\envs\Python27\lib\site-packages\keras\backend\theano_backend.py", line 1168, in in_train_phase
x = theano.ifelse.ifelse(_LEARNING_PHASE, x, alt)
AttributeError: 'module' object has no attribute 'ifelse'
降级所使用的theano版本，重新安装0.9版本的theano
参考资料
[1] https://keras.io/zh/applications/#resnet
[2] 【Keras】常用的预训练模型权重下载及使用
[3] keras 预训练模型的使用方法
[4] # 版本问题---keras和tensorflow的版本对应关系
[5] tensorflow和keras版本对应关系
[6] keras backend 简单介绍
[7] keras-import keras.backend as K的意义
[8] keras修改backend的方法 解决问题
[9] 解决OSError: Unable to open file (truncated file: eof = 84336640, sblock->base_addr = 0, stored_eof = 解决问题
[10] Keras深度学习框架学习笔记(3) - AttributeError：’module’ object has no attribute ‘ifelse’错误信息的解决方法 解决问题
[11] 详解keras的model.summary()输出参数Param计算过程
[12] keras 中获取张量 tensor 的维度大小。 解决问题
[13] keras 获取某层的输入/输出 tensor 尺寸
[14] keras获得model中某一层的某一个Tensor的输出维度
Keras的网络层相关操作
[1] AveragePooling1D和GlobalAveragePooling1D的区别
Keras的Dot类
[1] 代码系列——keras.layers.Dot()解析
[2] Keras的Dot类
[3] keras.layers层dot维度计算的一些介绍
[4] # 深度学习（六）keras常用函数学习
[5] Docs » Layers » 融合层 Merge
论文下载
[1] Attribute-Aware Attention Model for Fine-grained Representation Learning
预训练模型
[1] https://github.com/fchollet/deep-learning-models/releases/tag/v0.2
[2] https://github.com/fchollet/deep-learning-models/releases/tag/v0.1
代码
[1] # iamhankai/attribute-aware-attention