tebsorflow2.0 tf.keras猫狗识别(2)—自定义训练

在上一篇文章中，我们介绍了利用tensorflow封装好的神经网络来训练猫狗数据集。但是在科研中有时我们需要对网络进行修改，这是我们就需要自定义网络训练
第一部分我们已经介绍了图片的读取，这里就不在一一赘述tebsorflow2.0 tf.keras猫狗识别，我们直接从图片的预处理。

1. 图片预处理

在这一部分我们采用from_tensor_slices的方法对图片数据集进行构建，对比tf1.x版本采用队列形式读取数据，这一种方法比较简单切易于理解。

#构建一个tf.data.Dataset
#一个图片数据集构建 tf.data.Dataset 最简单的方法就是使用 from_tensor_slices 方法。
#将字符串数组切片，得到一个字符串数据集：
train_path_ds =  tf.data.Dataset.from_tensor_slices(train_all_image_path)
print(train_path_ds)
test_path_ds =  tf.data.Dataset.from_tensor_slices(test_all_image_path)
#现在创建一个新的数据集，通过在路径数据集上映射preprocess_image来动态加载和格式化图片。
AUTOTUNE = tf.data.experimental.AUTOTUNE
train_image_ds = train_path_ds.map(load_and_preprocess_image,num_parallel_calls=AUTOTUNE)
test_image_ds = test_path_ds.map(load_and_preprocess_image,num_parallel_calls=AUTOTUNE)
train_lable_ds = tf.data.Dataset.from_tensor_slices(tf.cast(train_image_label,tf.int64))
test_lable_ds = tf.data.Dataset.from_tensor_slices(tf.cast(test_image_label,tf.int64))
for label in train_lable_ds.take(5):
    print(lable_names[label.numpy()])
#%%构建一个（图片，标签）对数据集
#因为这些数据集顺序相同，可以将他们打包起来
image_label_ds = tf.data.Dataset.zip((train_image_ds,train_lable_ds))
test_data = tf.data.Dataset.zip((test_image_ds,test_lable_ds))
print(test_data)

<TensorSliceDataset shapes: (), types: tf.string>
cat
dog
dog
cat
dog
<ZipDataset shapes: ((256, 256, 3), ()), types: (tf.float32, tf.int64)>

2. 训练阶段

batch_size = 32
# 设置一个和数据集大小一致的 shuffle buffer size（随机缓冲区大小）以保证数据被充分打乱。
train_ds = image_label_ds.shuffle(buffer_size=train_image_count).batch(batch_size)
test_ds = test_data.batch(batch_size)

train_ds = train_ds.prefetch(buffer_size=AUTOTUNE)
test_ds = test_ds.prefetch(buffer_size=AUTOTUNE)

2.1 构建模型并训练

model = tf.keras.Sequential()   #顺序模型
model.add(tf.keras.layers.Conv2D(64, (3, 3), input_shape=(256, 256, 3), activation='relu'))
model.add(tf.keras.layers.BatchNormalization())
#model.add(tf.keras.layers.Conv2D(64, (3, 3), activation='relu'))
model.add(tf.keras.layers.MaxPooling2D())

#model.add(tf.keras.layers.Conv2D(128, (3, 3), activation='relu'))
model.add(tf.keras.layers.Conv2D(128, (3, 3), activation='relu'))
model.add(tf.keras.layers.BatchNormalization())
model.add(tf.keras.layers.MaxPooling2D())

#model.add(tf.keras.layers.Conv2D(256, (3, 3), activation='relu'))
model.add(tf.keras.layers.Conv2D(256, (3, 3), activation='relu'))
model.add(tf.keras.layers.BatchNormalization())
model.add(tf.keras.layers.MaxPooling2D())

model.add(tf.keras.layers.Conv2D(512, (3, 3), activation='relu'))
model.add(tf.keras.layers.BatchNormalization())
model.add(tf.keras.layers.MaxPooling2D())
model.add(tf.keras.layers.Conv2D(512, (3, 3), activation='relu'))
model.add(tf.keras.layers.BatchNormalization())
model.add(tf.keras.layers.MaxPooling2D())

model.add(tf.keras.layers.Conv2D(1024, (3, 3), activation='relu'))
model.add(tf.keras.layers.BatchNormalization())
model.add(tf.keras.layers.GlobalAveragePooling2D())

model.add(tf.keras.layers.Dense(1024, activation='relu'))
model.add(tf.keras.layers.BatchNormalization())
model.add(tf.keras.layers.Dense(256, activation='relu'))
model.add(tf.keras.layers.BatchNormalization())
model.add(tf.keras.layers.Dense(1))

#%%
model.summary()

Model: "sequential_8"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
conv2d_66 (Conv2D)           (None, 254, 254, 64)      1792      
_________________________________________________________________
batch_normalization (BatchNo (None, 254, 254, 64)      256       
_________________________________________________________________
max_pooling2d_40 (MaxPooling (None, 127, 127, 64)      0         
_________________________________________________________________
conv2d_67 (Conv2D)           (None, 125, 125, 128)     73856     
_________________________________________________________________
batch_normalization_1 (Batch (None, 125, 125, 128)     512       
_________________________________________________________________
max_pooling2d_41 (MaxPooling (None, 62, 62, 128)       0         
_________________________________________________________________
conv2d_68 (Conv2D)           (None, 60, 60, 256)       295168    
_________________________________________________________________
batch_normalization_2 (Batch (None, 60, 60, 256)       1024      
_________________________________________________________________
max_pooling2d_42 (MaxPooling (None, 30, 30, 256)       0         
_________________________________________________________________
conv2d_69 (Conv2D)           (None, 28, 28, 512)       1180160   
_________________________________________________________________
batch_normalization_3 (Batch (None, 28, 28, 512)       2048      
_________________________________________________________________
max_pooling2d_43 (MaxPooling (None, 14, 14, 512)       0         
_________________________________________________________________
conv2d_70 (Conv2D)           (None, 12, 12, 512)       2359808   
_________________________________________________________________
batch_normalization_4 (Batch (None, 12, 12, 512)       2048      
_________________________________________________________________
max_pooling2d_44 (MaxPooling (None, 6, 6, 512)         0         
_________________________________________________________________
conv2d_71 (Conv2D)           (None, 4, 4, 1024)        4719616   
_________________________________________________________________
batch_normalization_5 (Batch (None, 4, 4, 1024)        4096      
_________________________________________________________________
global_average_pooling2d_8 ( (None, 1024)              0         
_________________________________________________________________
dense_24 (Dense)             (None, 1024)              1049600   
_________________________________________________________________
batch_normalization_6 (Batch (None, 1024)              4096      
_________________________________________________________________
dense_25 (Dense)             (None, 256)               262400    
_________________________________________________________________
batch_normalization_7 (Batch (None, 256)               1024      
_________________________________________________________________
dense_26 (Dense)             (None, 1)                 257       
=================================================================
Total params: 9,957,761
Trainable params: 9,950,209
Non-trainable params: 7,552
_________________________________________________________________

optimizer = tf.keras.optimizers.Adam(learning_rate=0.0001)
#自定义损失,Sparse是可调用的对象
loss_fuc = tf.keras.losses.BinaryCrossentropy(from_logits=True)# 这是可调用的方法,因为我们没有加入激活函数，所以from_logits=true

#定义损失
def loss(model,x,y):
  y_ = model(x)
  y = tf.expand_dims(tf.cast(y,dtype=tf.float32),axis=1)
  return loss_fuc(y_,y)

#定义平均训练集和测试集平均损失和准确率
train_loss = tf.keras.metrics.Mean("train_loss")
train_accuracy = tf.keras.metrics.Accuracy("train_accuracy") 
test_loss = tf.keras.metrics.Mean("test_loss")
test_accuracy = tf.keras.metrics.Accuracy("test_accuracy") 

#定义train_step
def train_step(model,image,labels):
  with tf.GradientTape() as t:
    pred = model(image)
    loss_step = loss_fuc(tf.expand_dims(tf.cast(labels,dtype=tf.float32),axis=1),pred)
  grads = t.gradient(loss_step,model.trainable_variables)
  optimizer.apply_gradients(zip(grads,model.trainable_variables))
  train_loss(loss_step)
  train_accuracy(labels,tf.cast(pred>0,tf.int64))
#定义train_step
def test_step(model,image,labels):
  pred = model(image)
  loss_step = loss_fuc(tf.expand_dims(tf.cast(labels,dtype=tf.float32),axis=1),pred)
  test_loss(loss_step)
  test_accuracy(labels,tf.cast(pred>0,tf.int64))

def train(model,train_ds,test_ds):
  train_loss_sca = []
  test_loss_sca = []
  train_acc_sca = []
  test_acc_sca = []
  for epoch in range(30):
    for (batch,(image,labels)) in enumerate(train_ds):
      #print(".")
      #进行异步训连
      train_step(model,image,labels)
    for (batch,(image,labels)) in enumerate(test_ds):
      test_step(model,image,labels)
    train_loss_sca.append(train_loss.result())
    test_loss_sca.append(test_loss.result())
    train_acc_sca.append(train_accuracy.result())
    test_acc_sca.append(test_accuracy.result())

    print("epoch{} train_loss is {};train_accuracy is {};test_loss is {};test_accuracy is {}".format(epoch+1,
                               train_loss.result(),
                               train_accuracy.result(),
                               test_loss.result(),
                               test_accuracy.result()
                               ))
    
    train_loss.reset_states()
    train_accuracy.reset_states()
    test_loss.reset_states()
    test_accuracy.reset_states()
  return (train_loss_sca,test_loss_sca,train_acc_sca,test_acc_sca)
  
#训练
train_loss_sca,test_loss_sca,train_acc_sca,test_acc_sca = train(model,train_ds,test_ds)

epoch1 train_loss is 0.6930190324783325;train_accuracy is 0.5039578080177307;test_loss is 0.6916858553886414;test_accuracy is 0.5
epoch2 train_loss is 0.6807847023010254;train_accuracy is 0.5712401270866394;test_loss is 0.6401413083076477;test_accuracy is 0.6240000128746033
epoch3 train_loss is 0.6488288640975952;train_accuracy is 0.6144459247589111;test_loss is 0.6341478824615479;test_accuracy is 0.6399999856948853
epoch4 train_loss is 0.6222826838493347;train_accuracy is 0.6510553956031799;test_loss is 0.608300507068634;test_accuracy is 0.6790000200271606
epoch5 train_loss is 0.6008259654045105;train_accuracy is 0.6721636056900024;test_loss is 0.6023556590080261;test_accuracy is 0.6660000085830688
epoch6 train_loss is 0.5754649639129639;train_accuracy is 0.6952506303787231;test_loss is 0.5704880952835083;test_accuracy is 0.7039999961853027
epoch7 train_loss is 0.5785166025161743;train_accuracy is 0.6995382308959961;test_loss is 0.5662873983383179;test_accuracy is 0.7039999961853027
epoch8 train_loss is 0.5402986407279968;train_accuracy is 0.7272427678108215;test_loss is 0.5656307935714722;test_accuracy is 0.7049999833106995
epoch9 train_loss is 0.5293075442314148;train_accuracy is 0.7358179688453674;test_loss is 0.5394512414932251;test_accuracy is 0.7279999852180481
epoch10 train_loss is 0.5094398260116577;train_accuracy is 0.7569261193275452;test_loss is 0.5466101169586182;test_accuracy is 0.7369999885559082
epoch11 train_loss is 0.49413660168647766;train_accuracy is 0.7575857639312744;test_loss is 0.4992522597312927;test_accuracy is 0.7630000114440918
epoch12 train_loss is 0.46310922503471375;train_accuracy is 0.7833113670349121;test_loss is 0.48759791254997253;test_accuracy is 0.7620000243186951
epoch13 train_loss is 0.43853995203971863;train_accuracy is 0.7974933981895447;test_loss is 0.4716934263706207;test_accuracy is 0.7950000166893005
epoch14 train_loss is 0.3964250981807709;train_accuracy is 0.8258575201034546;test_loss is 0.48196572065353394;test_accuracy is 0.781000018119812
epoch15 train_loss is 0.3641780912876129;train_accuracy is 0.8390501141548157;test_loss is 0.4671226739883423;test_accuracy is 0.8080000281333923
epoch16 train_loss is 0.3265441060066223;train_accuracy is 0.857189953327179;test_loss is 0.4624425172805786;test_accuracy is 0.7960000038146973
epoch17 train_loss is 0.3088551461696625;train_accuracy is 0.8680738806724548;test_loss is 0.48780468106269836;test_accuracy is 0.7919999957084656
epoch18 train_loss is 0.2565159499645233;train_accuracy is 0.899076521396637;test_loss is 0.4692841172218323;test_accuracy is 0.8169999718666077
epoch19 train_loss is 0.23743189871311188;train_accuracy is 0.9027044773101807;test_loss is 0.4935546815395355;test_accuracy is 0.8209999799728394
epoch20 train_loss is 0.22497089207172394;train_accuracy is 0.9099604487419128;test_loss is 0.514469563961029;test_accuracy is 0.8100000023841858
epoch21 train_loss is 0.16318537294864655;train_accuracy is 0.9370052814483643;test_loss is 0.5123884677886963;test_accuracy is 0.8330000042915344
epoch22 train_loss is 0.13359478116035461;train_accuracy is 0.9495382308959961;test_loss is 0.5613532066345215;test_accuracy is 0.8190000057220459
epoch23 train_loss is 0.10728871077299118;train_accuracy is 0.9617414474487305;test_loss is 0.5727055668830872;test_accuracy is 0.8399999737739563
epoch24 train_loss is 0.08103378862142563;train_accuracy is 0.9696570038795471;test_loss is 0.608410656452179;test_accuracy is 0.8320000171661377
epoch25 train_loss is 0.04331161081790924;train_accuracy is 0.9858179688453674;test_loss is 0.7248072624206543;test_accuracy is 0.8309999704360962
epoch26 train_loss is 0.06562910228967667;train_accuracy is 0.977902352809906;test_loss is 0.7443087100982666;test_accuracy is 0.777999997138977
epoch27 train_loss is 0.05616709962487221;train_accuracy is 0.9818601608276367;test_loss is 0.7444125413894653;test_accuracy is 0.8190000057220459
epoch28 train_loss is 0.02497768960893154;train_accuracy is 0.9910950064659119;test_loss is 0.8051329851150513;test_accuracy is 0.8180000185966492
epoch29 train_loss is 0.029464619234204292;train_accuracy is 0.9887862801551819;test_loss is 0.9197209477424622;test_accuracy is 0.8240000009536743
epoch30 train_loss is 0.012036236003041267;train_accuracy is 0.9970316886901855;test_loss is 0.8360891342163086;test_accuracy is 0.8450000286102295

2.2 模型评估

我们对比一下在训练集和验证集的准确度和损失的变化曲线，我们可以发现，本网络有些过拟合。

plt.plot([i for i in range(30)], train_loss_sca, label='train_loss')
plt.plot([i for i in range(30)], test_loss_sca, label='test_aloss')
plt.legend()

plt.plot([i for i in range(30)], train_acc_sca, label='traib_acc')
plt.plot([i for i in range(30)], test_acc_sca, label='test_acc')
plt.legend()

总结一下：本测试的结果有些过拟合，考虑到图片较少，不能提取更为高级的特征，所以在测试集的效果只有83%,后期我们可以通过增加数据集，数据增强，并引入正则化和丢弃法等方法来抑制过拟合并提高精度。

3.数据增强

我们更改一下读取图片的代码就可以实现，如下所示

#确定每个图像的标签
lable_names = sorted(item.name for item in data_dir.glob("train/*/"))
#为每个标签分配索引,构建字典
lable_to_index = dict((name,index) for index,name in enumerate(lable_names))
print(lable_to_index)
#创建一个列表，包含每个文件的标签索引
train_image_label = [lable_to_index[pathlib.Path(path).parent.name] for path in train_all_image_path]
test_image_label = [lable_to_index[pathlib.Path(path).parent.name] for path in test_all_image_path]
#包装为函数，以备后用,图片的预处理
def preprocess_image(image,is_train=True):
    image = tf.image.decode_jpeg(image, channels=3)
    if is_train:
      image = tf.image.resize(image, [360, 360])    #图像剪裁
      image = tf.image.random_crop(image,[256,256,3]) #要写通道
      image = tf.image.random_flip_left_right(image)  #左右翻转
      imgae = tf.image.random_flip_up_down(image)   #上下翻转
      #image = tf.image.random_brightness(image,0.5)
      #image = tf.image.random_contrast(image,0,1)
    else:
      image = tf.image.resize(image,[256,256])
    image /= 255.0  # normalize to [0,1] range
    return image
#加载图片
def load_and_preprocess_train_image(path,is_train=True):
    image = tf.io.read_file(path)
    return preprocess_image(image,is_train)
def load_and_preprocess_test_image(path,is_train=False):
    image = tf.io.read_file(path)
    return preprocess_image(image,is_train)


image_path = test_all_image_path[11]
label = test_image_label[11]

plt.imshow(load_and_preprocess_train_image(image_path))
plt.grid(False)
##plt.xlabel(caption_image(image_path))
plt.title(lable_names[label].title())
plt.axis("off")
print()

训练集和验证集的准确度和损失的变化曲线，我们可以看出，不仅提高了准确率而且对过拟合也有较好的抑制作用。