首页 > 分享 > 《python深度学习》笔记

《python深度学习》笔记

萌宠菠菠乐园
2024-10-04 18:47

《python深度学习》笔记---5.3-4、猫狗分类（使用预训练网络-微调模型）

打赏

一、总结一句话总结：1、模型微调（fine-tuning）？2、只有上面的分类器已经训练好了，才能微调卷积基的顶部几层？3、微调网络的步骤？4、为什么不微调更多层？为什么不微调整个卷积基？5、微调使用的优化器的特点？6、注意，从损失曲线上看不出与之前相比有任何真正的提高（实际上还在变差）。你可能感到奇怪，如果损失没有降低，那么精度怎么能保持稳定或提高呢？7、使用预训练网络知识点总结？8、计算测试集准确率？二、5.3-4、猫狗分类（使用预训练网络-微调模型）

一、总结

一句话总结：

【微调是指将预训练网络的前几层解冻】：微调是指将预训练网络的前几层解冻，在猫狗分类的例子中，可以看到准确率从90到94，提高了4个百分点【先训练好分类层】：微调的话，注意先把dense层（分类层）训练好，再低学习率的训练网络最后几层

# 冻结直到某一层的所有层 conv_base.trainable = True set_trainable = False for layer in conv_base.layers: if layer.name == 'block5_conv1': set_trainable = True if set_trainable: layer.trainable = True else: layer.trainable = False

1、模型微调（fine-tuning）？

【一种模型复用方法】：另一种广泛使用的模型复用方法是模型微调（fine-tuning），与特征提取互为补充。【微调是指将前几层解冻】：对于用于特征提取的冻结的模型基，微调是指将其顶部的几层“解冻”，并将这解冻的几层和新增加的部分（本例中是全连接分类器）联合训练【略微调整抽象表示】：之所以叫作微调，是因为它只是略微调整了所复用模型中更加抽象的表示，以便让这些表示与手头的问题更加相关。

2、只有上面的分类器已经训练好了，才能微调卷积基的顶部几层？

【只有分类器训练好了才能微调】：前面说过，冻结VGG16 的卷积基是为了能够在上面训练一个随机初始化的分类器。同理，只有上面的分类器已经训练好了，才能微调卷积基的顶部几层。【没训练好，微调的话，误差很大】：如果分类器没有训练好，那么训练期间通过网络传播的误差信号会特别大，微调的几层之前学到的表示都会被破坏。

3、微调网络的步骤？

(1) 在已经训练好的基网络（base network）上添加自定义网络。 (2) 冻结基网络。 (3) 训练所添加的部分。 (4) 解冻基网络的一些层。 (5) 联合训练解冻的这些层和添加的部分。

4、为什么不微调更多层？为什么不微调整个卷积基？

【靠底部的层编码的是更加通用的可复用特征】：卷积基中更靠底部的层编码的是更加通用的可复用特征，而更靠顶部的层编码的是更专业化的特征。【微调顶部的层更有用】：微调这些更专业化的特征更加有用，因为它们需要在你的新问题上改变用途。微调更靠底部的层，得到的回报会更少。训练的参数越多，过拟合的风险越大。卷积基有1500 万个参数，所以在你的小型数据集上训练这么多参数是有风险的。

5、微调使用的优化器的特点？

【学习率非常小的 RMSProp 优化器】：现在你可以开始微调网络。我们将使用学习率非常小的 RMSProp 优化器来实现。之所以让学习率很小，是因为对于微调的三层表示，我们希望其变化范围不要太大。太大的权重更新可能会破坏这些表示。

model.compile(loss='binary_crossentropy', optimizer=optimizers.RMSprop(lr=1e-5), metrics=['acc'])

6、注意，从损失曲线上看不出与之前相比有任何真正的提高（实际上还在变差）。你可能感到奇怪，如果损失没有降低，那么精度怎么能保持稳定或提高呢？

【影响精度的是损失值的分布，而非平均】：图中展示的是逐点（pointwise）损失值的平均值，但影响精度的是损失值的分布，而不是平均值，因为精度是模型预测的类别概率的二进制阈值。即使从平均损失中无法看出，但模型也仍然可能在改进。

7、使用预训练网络知识点总结？

【卷积神经网络在小数据集上也挺好用】：卷积神经网络是用于计算机视觉任务的最佳机器学习模型。即使在非常小的数据集上也可以从头开始训练一个卷积神经网络，而且得到的结果还不错。【数据增强解决小数据集的过拟合】：在小型数据集上的主要问题是过拟合。在处理图像数据时，数据增强是一种降低过拟合的强大方法。【使用预训练网络来实现特征提取】：利用特征提取，可以很容易将现有的卷积神经网络复用于新的数据集。对于小型图像数据集，这是一种很有价值的方法。【对预训练网络实现微调】：作为特征提取的补充，你还可以使用微调，将现有模型之前学到的一些数据表示应用于新问题。这种方法可以进一步提高模型性能。

8、计算测试集准确率？

model的evaluate方法：test_loss, test_acc = model.evaluate(test_generator, steps=50)

test_generator = test_datagen.flow_from_directory( test_dir, target_size=(150, 150), batch_size=20, class_mode='binary') test_loss, test_acc = model.evaluate(test_generator, steps=50) print('test acc:', test_acc)

二、5.3-4、猫狗分类（使用预训练网络-微调模型）

博客对应课程的视频位置：

import pandas as pd import numpy as np import tensorflow as tf import matplotlib.pyplot as plt

In [2]:

from tensorflow.keras.applications import VGG16 # 把vgg模型弄过来 conv_base = VGG16(weights='imagenet', # include_top=False表示不包含dense层 include_top=False, input_shape=(150, 150, 3)) # C:UsersFan Renyi.kerasmodelsvgg16_weights_tf_dim_ordering_tf_kernels_notop.h5

In [3]:

import os base_dir = 'E:78_recorded_lesson01_course_githubAI_dataSetdogs-vs-catscats_and_dogs_small' train_dir = os.path.join(base_dir, 'train') validation_dir = os.path.join(base_dir, 'validation') test_dir = os.path.join(base_dir, 'test')

In [4]:

from tensorflow.keras import models from tensorflow.keras import layers model = models.Sequential() model.add(conv_base) model.add(layers.Flatten()) model.add(layers.Dense(256, activation='relu')) model.add(layers.Dense(1, activation='sigmoid')) model.summary()

Model: "sequential" _________________________________________________________________ Layer (type) Output Shape Param # ================================================================= vgg16 (Functional) (None, 4, 4, 512) 14714688 _________________________________________________________________ flatten (Flatten) (None, 8192) 0 _________________________________________________________________ dense (Dense) (None, 256) 2097408 _________________________________________________________________ dense_1 (Dense) (None, 1) 257 ================================================================= Total params: 16,812,353 Trainable params: 16,812,353 Non-trainable params: 0 _________________________________________________________________

In [5]:

conv_base.trainable = False

In [6]:

from tensorflow.keras.preprocessing.image import ImageDataGenerator from tensorflow.keras import optimizers train_datagen = ImageDataGenerator( rescale=1./255, rotation_range=40, width_shift_range=0.2, height_shift_range=0.2, shear_range=0.2, zoom_range=0.2, horizontal_flip=True, fill_mode='nearest') # 注意，不能增强验证数据 test_datagen = ImageDataGenerator(rescale=1./255) train_generator = train_datagen.flow_from_directory( train_dir, # 目标目录 target_size=(150, 150), # 将所有图像的大小调整为 150×150 batch_size=20, class_mode='binary') # 因为使用了binary_crossentropy 损失，所以需要用二进制标签 validation_generator = test_datagen.flow_from_directory( validation_dir, target_size=(150, 150), batch_size=20, class_mode='binary') model.compile(loss='binary_crossentropy', optimizer=optimizers.RMSprop(lr=2e-5), metrics=['acc'])

Found 2000 images belonging to 2 classes. Found 1000 images belonging to 2 classes.

In [7]:

history = model.fit( train_generator, steps_per_epoch=100, epochs=30, validation_data=validation_generator, validation_steps=50)

Epoch 1/30 100/100 [==============================] - 42s 423ms/step - loss: 0.5806 - acc: 0.7040 - val_loss: 0.4488 - val_acc: 0.8140 Epoch 2/30 100/100 [==============================] - 19s 187ms/step - loss: 0.4826 - acc: 0.7850 - val_loss: 0.3632 - val_acc: 0.8580 Epoch 3/30 100/100 [==============================] - 19s 189ms/step - loss: 0.4398 - acc: 0.7970 - val_loss: 0.3246 - val_acc: 0.8680 Epoch 4/30 100/100 [==============================] - 19s 190ms/step - loss: 0.4014 - acc: 0.8195 - val_loss: 0.3056 - val_acc: 0.8810 Epoch 5/30 100/100 [==============================] - 19s 191ms/step - loss: 0.3747 - acc: 0.8330 - val_loss: 0.2898 - val_acc: 0.8830 Epoch 6/30 100/100 [==============================] - 19s 194ms/step - loss: 0.3599 - acc: 0.8430 - val_loss: 0.2752 - val_acc: 0.8960 Epoch 7/30 100/100 [==============================] - 20s 202ms/step - loss: 0.3449 - acc: 0.8445 - val_loss: 0.2713 - val_acc: 0.8950 Epoch 8/30 100/100 [==============================] - 20s 195ms/step - loss: 0.3396 - acc: 0.8550 - val_loss: 0.2597 - val_acc: 0.8990 Epoch 9/30 100/100 [==============================] - 19s 194ms/step - loss: 0.3313 - acc: 0.8525 - val_loss: 0.2543 - val_acc: 0.8960 Epoch 10/30 100/100 [==============================] - 20s 199ms/step - loss: 0.3263 - acc: 0.8625 - val_loss: 0.2501 - val_acc: 0.9030 Epoch 11/30 100/100 [==============================] - 20s 199ms/step - loss: 0.3264 - acc: 0.8655 - val_loss: 0.2485 - val_acc: 0.8970 Epoch 12/30 100/100 [==============================] - 20s 197ms/step - loss: 0.3368 - acc: 0.8545 - val_loss: 0.2500 - val_acc: 0.8970 Epoch 13/30 100/100 [==============================] - 20s 204ms/step - loss: 0.3348 - acc: 0.8535 - val_loss: 0.2472 - val_acc: 0.9060 Epoch 14/30 100/100 [==============================] - 20s 198ms/step - loss: 0.3185 - acc: 0.8680 - val_loss: 0.2533 - val_acc: 0.8940 Epoch 15/30 100/100 [==============================] - 20s 199ms/step - loss: 0.3137 - acc: 0.8615 - val_loss: 0.2439 - val_acc: 0.9060 Epoch 16/30 100/100 [==============================] - 20s 201ms/step - loss: 0.3147 - acc: 0.8660 - val_loss: 0.2413 - val_acc: 0.9010 Epoch 17/30 100/100 [==============================] - 21s 206ms/step - loss: 0.2998 - acc: 0.8730 - val_loss: 0.2406 - val_acc: 0.9000 Epoch 18/30 100/100 [==============================] - 19s 191ms/step - loss: 0.3185 - acc: 0.8580 - val_loss: 0.2415 - val_acc: 0.9020 Epoch 19/30 100/100 [==============================] - 19s 190ms/step - loss: 0.3115 - acc: 0.8610 - val_loss: 0.2406 - val_acc: 0.8970 Epoch 20/30 100/100 [==============================] - 19s 191ms/step - loss: 0.3036 - acc: 0.8705 - val_loss: 0.2412 - val_acc: 0.9000 Epoch 21/30 100/100 [==============================] - 19s 191ms/step - loss: 0.2945 - acc: 0.8695 - val_loss: 0.2421 - val_acc: 0.8980 Epoch 22/30 100/100 [==============================] - 19s 192ms/step - loss: 0.2923 - acc: 0.8750 - val_loss: 0.2396 - val_acc: 0.9000 Epoch 23/30 100/100 [==============================] - 19s 191ms/step - loss: 0.2822 - acc: 0.8775 - val_loss: 0.2411 - val_acc: 0.9050 Epoch 24/30 100/100 [==============================] - 19s 191ms/step - loss: 0.2969 - acc: 0.8770 - val_loss: 0.2347 - val_acc: 0.9060 Epoch 25/30 100/100 [==============================] - 19s 191ms/step - loss: 0.2897 - acc: 0.8750 - val_loss: 0.2468 - val_acc: 0.8880 Epoch 26/30 100/100 [==============================] - 19s 192ms/step - loss: 0.2938 - acc: 0.8745 - val_loss: 0.2380 - val_acc: 0.9000 Epoch 27/30 100/100 [==============================] - 19s 193ms/step - loss: 0.2869 - acc: 0.8790 - val_loss: 0.2435 - val_acc: 0.9060 Epoch 28/30 100/100 [==============================] - 19s 192ms/step - loss: 0.2713 - acc: 0.8900 - val_loss: 0.2439 - val_acc: 0.8900 Epoch 29/30 100/100 [==============================] - 19s 192ms/step - loss: 0.2850 - acc: 0.8815 - val_loss: 0.2434 - val_acc: 0.9060 Epoch 30/30 100/100 [==============================] - 19s 193ms/step - loss: 0.2723 - acc: 0.8810 - val_loss: 0.2345 - val_acc: 0.9030

In [8]:

acc = history.history['acc'] val_acc = history.history['val_acc'] loss = history.history['loss'] val_loss = history.history['val_loss'] epochs = range(1, len(acc) + 1) plt.plot(epochs, acc, 'b--', label='Training acc') plt.plot(epochs, val_acc, 'r-', label='Validation acc') plt.title('Training and validation accuracy') plt.legend() plt.figure() plt.plot(epochs, loss, 'b--', label='Training loss') plt.plot(epochs, val_loss, 'r-', label='Validation loss') plt.title('Training and validation loss') plt.legend() plt.show()

仅微调卷积基最后的两三层

In [9]:

In [11]:

for layer in conv_base.layers: print(layer) print(layer.trainable)

<tensorflow.python.keras.engine.input_layer.InputLayer object at 0x000001FA568E0988> False <tensorflow.python.keras.layers.convolutional.Conv2D object at 0x000001FA56AE3CC8> False <tensorflow.python.keras.layers.convolutional.Conv2D object at 0x000001FA56698608> False <tensorflow.python.keras.layers.pooling.MaxPooling2D object at 0x000001FA56B64588> False <tensorflow.python.keras.layers.convolutional.Conv2D object at 0x000001FA56B6B088> False <tensorflow.python.keras.layers.convolutional.Conv2D object at 0x000001FA56B6EA48> False <tensorflow.python.keras.layers.pooling.MaxPooling2D object at 0x000001FA56B76588> False <tensorflow.python.keras.layers.convolutional.Conv2D object at 0x000001FA56B7BFC8> False <tensorflow.python.keras.layers.convolutional.Conv2D object at 0x000001FA56B71D48> False <tensorflow.python.keras.layers.convolutional.Conv2D object at 0x000001FA5D053548> False <tensorflow.python.keras.layers.pooling.MaxPooling2D object at 0x000001FA5D05D0C8> False <tensorflow.python.keras.layers.convolutional.Conv2D object at 0x000001FA5D062B88> False <tensorflow.python.keras.layers.convolutional.Conv2D object at 0x000001FA5D0653C8> False <tensorflow.python.keras.layers.convolutional.Conv2D object at 0x000001FA5D06C188> False <tensorflow.python.keras.layers.pooling.MaxPooling2D object at 0x000001FA5D06ED88> False <tensorflow.python.keras.layers.convolutional.Conv2D object at 0x000001FA5D077888> True <tensorflow.python.keras.layers.convolutional.Conv2D object at 0x000001FA5D07A348> True <tensorflow.python.keras.layers.convolutional.Conv2D object at 0x000001FA5D07DF48> True <tensorflow.python.keras.layers.pooling.MaxPooling2D object at 0x000001FA5D085B48> True

微调模型

In [12]:

model.compile(loss='binary_crossentropy', optimizer=optimizers.RMSprop(lr=1e-5), metrics=['acc'])

In [13]:

history = model.fit( train_generator, steps_per_epoch=100, epochs=100, validation_data=validation_generator, validation_steps=50)

Epoch 1/100 100/100 [==============================] - 18s 184ms/step - loss: 0.2999 - acc: 0.8730 - val_loss: 0.2151 - val_acc: 0.9170 Epoch 2/100 100/100 [==============================] - 18s 184ms/step - loss: 0.2738 - acc: 0.8815 - val_loss: 0.2047 - val_acc: 0.9200 Epoch 3/100 100/100 [==============================] - 19s 191ms/step - loss: 0.2460 - acc: 0.8980 - val_loss: 0.2385 - val_acc: 0.9090 Epoch 4/100 100/100 [==============================] - 19s 191ms/step - loss: 0.2231 - acc: 0.9075 - val_loss: 0.2395 - val_acc: 0.9000 Epoch 5/100 100/100 [==============================] - 19s 194ms/step - loss: 0.2025 - acc: 0.9140 - val_loss: 0.2632 - val_acc: 0.8970 Epoch 6/100 100/100 [==============================] - 20s 195ms/step - loss: 0.1970 - acc: 0.9190 - val_loss: 0.1916 - val_acc: 0.9290 Epoch 7/100 100/100 [==============================] - 19s 194ms/step - loss: 0.1845 - acc: 0.9235 - val_loss: 0.2077 - val_acc: 0.9230 Epoch 8/100 100/100 [==============================] - 19s 194ms/step - loss: 0.1688 - acc: 0.9365 - val_loss: 0.2087 - val_acc: 0.9250 Epoch 9/100 100/100 [==============================] - 20s 195ms/step - loss: 0.1587 - acc: 0.9350 - val_loss: 0.2156 - val_acc: 0.9200 Epoch 10/100 100/100 [==============================] - 20s 197ms/step - loss: 0.1558 - acc: 0.9400 - val_loss: 0.2253 - val_acc: 0.9170 Epoch 11/100 100/100 [==============================] - 19s 194ms/step - loss: 0.1503 - acc: 0.9345 - val_loss: 0.2221 - val_acc: 0.9160 Epoch 12/100 100/100 [==============================] - 19s 195ms/step - loss: 0.1430 - acc: 0.9470 - val_loss: 0.1952 - val_acc: 0.9290 Epoch 13/100 100/100 [==============================] - 20s 197ms/step - loss: 0.1367 - acc: 0.9480 - val_loss: 0.1805 - val_acc: 0.9340 Epoch 14/100 100/100 [==============================] - 20s 196ms/step - loss: 0.1243 - acc: 0.9520 - val_loss: 0.1899 - val_acc: 0.9350 Epoch 15/100 100/100 [==============================] - 20s 197ms/step - loss: 0.1102 - acc: 0.9595 - val_loss: 0.2190 - val_acc: 0.9260 Epoch 16/100 100/100 [==============================] - 20s 197ms/step - loss: 0.1191 - acc: 0.9575 - val_loss: 0.2498 - val_acc: 0.9210 Epoch 17/100 100/100 [==============================] - 20s 197ms/step - loss: 0.1210 - acc: 0.9515 - val_loss: 0.1917 - val_acc: 0.9300 Epoch 18/100 100/100 [==============================] - 20s 198ms/step - loss: 0.1224 - acc: 0.9495 - val_loss: 0.1885 - val_acc: 0.9280 Epoch 19/100 100/100 [==============================] - 20s 199ms/step - loss: 0.1040 - acc: 0.9595 - val_loss: 0.1839 - val_acc: 0.9270 Epoch 20/100 100/100 [==============================] - 20s 198ms/step - loss: 0.0983 - acc: 0.9585 - val_loss: 0.2617 - val_acc: 0.9170 Epoch 21/100 100/100 [==============================] - 20s 204ms/step - loss: 0.1028 - acc: 0.9635 - val_loss: 0.3127 - val_acc: 0.9090 Epoch 22/100 100/100 [==============================] - 20s 198ms/step - loss: 0.0953 - acc: 0.9595 - val_loss: 0.2070 - val_acc: 0.9350 Epoch 23/100 100/100 [==============================] - 20s 199ms/step - loss: 0.0859 - acc: 0.9695 - val_loss: 0.2136 - val_acc: 0.9300 Epoch 24/100 100/100 [==============================] - 20s 201ms/step - loss: 0.0758 - acc: 0.9705 - val_loss: 0.2023 - val_acc: 0.9330 Epoch 25/100 100/100 [==============================] - 20s 199ms/step - loss: 0.0990 - acc: 0.9615 - val_loss: 0.1950 - val_acc: 0.9310 Epoch 26/100 100/100 [==============================] - 20s 199ms/step - loss: 0.0873 - acc: 0.9670 - val_loss: 0.1810 - val_acc: 0.9370 Epoch 27/100 100/100 [==============================] - 20s 199ms/step - loss: 0.0791 - acc: 0.9695 - val_loss: 0.2265 - val_acc: 0.9290 Epoch 28/100 100/100 [==============================] - 20s 199ms/step - loss: 0.0751 - acc: 0.9680 - val_loss: 0.1840 - val_acc: 0.9410 Epoch 29/100 100/100 [==============================] - 20s 199ms/step - loss: 0.0659 - acc: 0.9740 - val_loss: 0.2356 - val_acc: 0.9290 Epoch 30/100 100/100 [==============================] - 20s 200ms/step - loss: 0.0603 - acc: 0.9770 - val_loss: 0.2066 - val_acc: 0.9330 Epoch 31/100 100/100 [==============================] - 20s 198ms/step - loss: 0.0612 - acc: 0.9735 - val_loss: 0.1972 - val_acc: 0.9350 Epoch 32/100 100/100 [==============================] - 20s 199ms/step - loss: 0.0626 - acc: 0.9745 - val_loss: 0.1989 - val_acc: 0.9300 Epoch 33/100 100/100 [==============================] - 20s 199ms/step - loss: 0.0628 - acc: 0.9765 - val_loss: 0.2080 - val_acc: 0.9400 Epoch 34/100 100/100 [==============================] - 20s 199ms/step - loss: 0.0651 - acc: 0.9740 - val_loss: 0.2322 - val_acc: 0.9330 Epoch 35/100 100/100 [==============================] - 20s 199ms/step - loss: 0.0576 - acc: 0.9765 - val_loss: 0.1911 - val_acc: 0.9370 Epoch 36/100 100/100 [==============================] - 20s 199ms/step - loss: 0.0664 - acc: 0.9770 - val_loss: 0.1941 - val_acc: 0.9350 Epoch 37/100 100/100 [==============================] - 20s 199ms/step - loss: 0.0585 - acc: 0.9790 - val_loss: 0.2717 - val_acc: 0.9230 Epoch 38/100 100/100 [==============================] - 20s 200ms/step - loss: 0.0666 - acc: 0.9760 - val_loss: 0.1788 - val_acc: 0.9390 Epoch 39/100 100/100 [==============================] - 20s 200ms/step - loss: 0.0556 - acc: 0.9800 - val_loss: 0.2031 - val_acc: 0.9370 Epoch 40/100 100/100 [==============================] - 20s 198ms/step - loss: 0.0433 - acc: 0.9835 - val_loss: 0.2189 - val_acc: 0.9350 Epoch 41/100 100/100 [==============================] - 20s 200ms/step - loss: 0.0493 - acc: 0.9810 - val_loss: 0.2113 - val_acc: 0.9330 Epoch 42/100 100/100 [==============================] - 20s 200ms/step - loss: 0.0478 - acc: 0.9815 - val_loss: 0.2069 - val_acc: 0.9430 Epoch 43/100 100/100 [==============================] - 20s 200ms/step - loss: 0.0475 - acc: 0.9830 - val_loss: 0.2268 - val_acc: 0.9380 Epoch 44/100 100/100 [==============================] - 20s 200ms/step - loss: 0.0395 - acc: 0.9860 - val_loss: 0.3817 - val_acc: 0.9080 Epoch 45/100 100/100 [==============================] - 20s 200ms/step - loss: 0.0430 - acc: 0.9825 - val_loss: 0.2086 - val_acc: 0.9370 Epoch 46/100 100/100 [==============================] - 20s 200ms/step - loss: 0.0449 - acc: 0.9815 - val_loss: 0.1946 - val_acc: 0.9390 Epoch 47/100 100/100 [==============================] - 20s 201ms/step - loss: 0.0372 - acc: 0.9880 - val_loss: 0.2430 - val_acc: 0.9330 Epoch 48/100 100/100 [==============================] - 20s 200ms/step - loss: 0.0395 - acc: 0.9875 - val_loss: 0.2147 - val_acc: 0.9330 Epoch 49/100 100/100 [==============================] - 20s 200ms/step - loss: 0.0411 - acc: 0.9835 - val_loss: 0.2187 - val_acc: 0.9430 Epoch 50/100 100/100 [==============================] - 20s 200ms/step - loss: 0.0420 - acc: 0.9865 - val_loss: 0.2184 - val_acc: 0.9390 Epoch 51/100 100/100 [==============================] - 20s 200ms/step - loss: 0.0355 - acc: 0.9855 - val_loss: 0.2217 - val_acc: 0.9360 Epoch 52/100 100/100 [==============================] - 20s 200ms/step - loss: 0.0317 - acc: 0.9865 - val_loss: 0.2387 - val_acc: 0.9410 Epoch 53/100 100/100 [==============================] - 20s 201ms/step - loss: 0.0392 - acc: 0.9865 - val_loss: 0.2358 - val_acc: 0.9380 Epoch 54/100 100/100 [==============================] - 20s 200ms/step - loss: 0.0305 - acc: 0.9885 - val_loss: 0.2586 - val_acc: 0.9410 Epoch 55/100 100/100 [==============================] - 20s 199ms/step - loss: 0.0324 - acc: 0.9865 - val_loss: 0.2093 - val_acc: 0.9420 Epoch 56/100 100/100 [==============================] - 20s 200ms/step - loss: 0.0339 - acc: 0.9885 - val_loss: 0.2098 - val_acc: 0.9430 Epoch 57/100 100/100 [==============================] - 20s 201ms/step - loss: 0.0342 - acc: 0.9875 - val_loss: 0.2273 - val_acc: 0.9450 Epoch 58/100 100/100 [==============================] - 20s 201ms/step - loss: 0.0354 - acc: 0.9880 - val_loss: 0.2951 - val_acc: 0.9360 Epoch 59/100 100/100 [==============================] - 20s 200ms/step - loss: 0.0280 - acc: 0.9915 - val_loss: 0.2696 - val_acc: 0.9420 Epoch 60/100 100/100 [==============================] - 20s 199ms/step - loss: 0.0347 - acc: 0.9880 - val_loss: 0.2253 - val_acc: 0.9450 Epoch 61/100 100/100 [==============================] - 20s 199ms/step - loss: 0.0300 - acc: 0.9895 - val_loss: 0.4227 - val_acc: 0.9130 Epoch 62/100 100/100 [==============================] - 20s 199ms/step - loss: 0.0345 - acc: 0.9865 - val_loss: 0.2661 - val_acc: 0.9380 Epoch 63/100 100/100 [==============================] - 20s 199ms/step - loss: 0.0370 - acc: 0.9870 - val_loss: 0.3239 - val_acc: 0.9300 Epoch 64/100 100/100 [==============================] - 20s 199ms/step - loss: 0.0280 - acc: 0.9885 - val_loss: 0.2536 - val_acc: 0.9380 Epoch 65/100 100/100 [==============================] - 20s 203ms/step - loss: 0.0227 - acc: 0.9915 - val_loss: 0.2505 - val_acc: 0.9410 Epoch 66/100 100/100 [==============================] - 20s 199ms/step - loss: 0.0274 - acc: 0.9930 - val_loss: 0.2226 - val_acc: 0.9400 Epoch 67/100 100/100 [==============================] - 20s 200ms/step - loss: 0.0229 - acc: 0.9910 - val_loss: 0.2298 - val_acc: 0.9410 Epoch 68/100 100/100 [==============================] - 20s 199ms/step - loss: 0.0231 - acc: 0.9905 - val_loss: 0.2250 - val_acc: 0.9420 Epoch 69/100 100/100 [==============================] - 20s 201ms/step - loss: 0.0336 - acc: 0.9870 - val_loss: 0.2955 - val_acc: 0.9310 Epoch 70/100 100/100 [==============================] - 20s 200ms/step - loss: 0.0230 - acc: 0.9920 - val_loss: 0.4870 - val_acc: 0.9100 Epoch 71/100 100/100 [==============================] - 20s 201ms/step - loss: 0.0353 - acc: 0.9870 - val_loss: 0.2387 - val_acc: 0.9420 Epoch 72/100 100/100 [==============================] - 20s 201ms/step - loss: 0.0256 - acc: 0.9920 - val_loss: 0.2777 - val_acc: 0.9390 Epoch 73/100 100/100 [==============================] - 20s 200ms/step - loss: 0.0323 - acc: 0.9895 - val_loss: 0.2311 - val_acc: 0.9420 Epoch 74/100 100/100 [==============================] - 19s 193ms/step - loss: 0.0295 - acc: 0.9890 - val_loss: 0.3283 - val_acc: 0.9340 Epoch 75/100 100/100 [==============================] - 19s 194ms/step - loss: 0.0275 - acc: 0.9920 - val_loss: 0.3358 - val_acc: 0.9330 Epoch 76/100 100/100 [==============================] - 20s 199ms/step - loss: 0.0264 - acc: 0.9905 - val_loss: 0.2475 - val_acc: 0.9390 Epoch 77/100 100/100 [==============================] - 20s 200ms/step - loss: 0.0206 - acc: 0.9940 - val_loss: 0.2863 - val_acc: 0.9400 Epoch 78/100 100/100 [==============================] - 20s 201ms/step - loss: 0.0169 - acc: 0.9925 - val_loss: 0.2389 - val_acc: 0.9450 Epoch 79/100 100/100 [==============================] - 20s 199ms/step - loss: 0.0208 - acc: 0.9935 - val_loss: 0.2606 - val_acc: 0.9390 Epoch 80/100 100/100 [==============================] - 20s 201ms/step - loss: 0.0339 - acc: 0.9890 - val_loss: 0.2165 - val_acc: 0.9390 Epoch 81/100 100/100 [==============================] - 21s 207ms/step - loss: 0.0311 - acc: 0.9885 - val_loss: 0.2396 - val_acc: 0.9450 Epoch 82/100 100/100 [==============================] - 20s 200ms/step - loss: 0.0263 - acc: 0.9905 - val_loss: 0.2196 - val_acc: 0.9450 Epoch 83/100 100/100 [==============================] - 20s 200ms/step - loss: 0.0220 - acc: 0.9915 - val_loss: 0.3002 - val_acc: 0.9320 Epoch 84/100 100/100 [==============================] - 20s 201ms/step - loss: 0.0187 - acc: 0.9920 - val_loss: 0.2295 - val_acc: 0.9440 Epoch 85/100 100/100 [==============================] - 20s 200ms/step - loss: 0.0223 - acc: 0.9915 - val_loss: 0.2608 - val_acc: 0.9430 Epoch 86/100 100/100 [==============================] - 20s 202ms/step - loss: 0.0204 - acc: 0.9930 - val_loss: 0.2567 - val_acc: 0.9380 Epoch 87/100 100/100 [==============================] - 20s 202ms/step - loss: 0.0271 - acc: 0.9880 - val_loss: 0.2265 - val_acc: 0.9380 Epoch 88/100 100/100 [==============================] - 20s 202ms/step - loss: 0.0135 - acc: 0.9955 - val_loss: 0.2674 - val_acc: 0.9390 Epoch 89/100 100/100 [==============================] - 20s 201ms/step - loss: 0.0196 - acc: 0.9930 - val_loss: 0.4253 - val_acc: 0.9190 Epoch 90/100 100/100 [==============================] - 20s 201ms/step - loss: 0.0204 - acc: 0.9945 - val_loss: 0.2534 - val_acc: 0.9440 Epoch 91/100 100/100 [==============================] - 20s 202ms/step - loss: 0.0146 - acc: 0.9945 - val_loss: 0.3456 - val_acc: 0.9360 Epoch 92/100 100/100 [==============================] - 20s 202ms/step - loss: 0.0176 - acc: 0.9940 - val_loss: 0.2618 - val_acc: 0.9390 Epoch 93/100 100/100 [==============================] - 20s 203ms/step - loss: 0.0289 - acc: 0.9900 - val_loss: 0.3886 - val_acc: 0.9330 Epoch 94/100 100/100 [==============================] - 20s 201ms/step - loss: 0.0157 - acc: 0.9940 - val_loss: 0.2674 - val_acc: 0.9410 Epoch 95/100 100/100 [==============================] - 20s 200ms/step - loss: 0.0179 - acc: 0.9940 - val_loss: 0.2833 - val_acc: 0.9430 Epoch 96/100 100/100 [==============================] - 20s 202ms/step - loss: 0.0204 - acc: 0.9925 - val_loss: 0.2726 - val_acc: 0.9420 Epoch 97/100 100/100 [==============================] - 20s 200ms/step - loss: 0.0217 - acc: 0.9940 - val_loss: 0.3880 - val_acc: 0.9310 Epoch 98/100 100/100 [==============================] - 20s 202ms/step - loss: 0.0079 - acc: 0.9975 - val_loss: 0.3431 - val_acc: 0.9360 Epoch 99/100 100/100 [==============================] - 20s 202ms/step - loss: 0.0261 - acc: 0.9895 - val_loss: 0.2835 - val_acc: 0.9450 Epoch 100/100 100/100 [==============================] - 20s 201ms/step - loss: 0.0172 - acc: 0.9940 - val_loss: 0.2963 - val_acc: 0.9410

In [14]:

这些曲线看起来包含噪声。为了让图像更具可读性，你可以将每个损失和精度都替换为指数移动平均值，从而让曲线变得平滑。

使曲线变得平滑

In [15]:

def smooth_curve(points, factor=0.8): smoothed_points = [] for point in points: if smoothed_points: previous = smoothed_points[-1] # 上一个节点*0.8+当前节点*0.2 smoothed_points.append(previous * factor + point * (1 - factor)) else: # 添加point smoothed_points.append(point) return smoothed_points

In [17]:

plt.plot(epochs,smooth_curve(acc), 'b--', label='Smoothed training acc') plt.plot(epochs,smooth_curve(val_acc), 'r-', label='Smoothed validation acc') plt.title('Training and validation accuracy') plt.legend() plt.figure() plt.plot(epochs,smooth_curve(loss), 'b--', label='Smoothed training loss') plt.plot(epochs,smooth_curve(val_loss), 'r-', label='Smoothed validation loss') plt.title('Training and validation loss') plt.legend() plt.show()

测试数据上最终评估这个模型

In [18]:

test_generator = test_datagen.flow_from_directory( test_dir, target_size=(150, 150), batch_size=20, class_mode='binary') test_loss, test_acc = model.evaluate_generator(test_generator, steps=50) print('test acc:', test_acc)

Found 1000 images belonging to 2 classes. WARNING:tensorflow:From <ipython-input-18-a0a3d19fae47>:7: Model.evaluate_generator (from tensorflow.python.keras.engine.training) is deprecated and will be removed in a future version. Instructions for updating: Please use Model.evaluate, which supports generators. test acc: 0.9409999847412109

【费宝产品大全】Furbaby百科

湛江吴川宠物医疗

热点分享

布偶猫吃什么对毛发好原来这些食物就可以

对于布偶猫这种长毛猫来说，一般情况下，布偶猫这样的长毛猫咪毛...

这九种宠物既新奇又独特，看完你爱上没有？

这九种宠物既新奇又独特，看完你爱上没有？小猫，小狗，仓鼠...

推荐分享

缅因猫能长多大一种体型较大的猫

缅因猫能长多大?缅因猫是很多人都喜欢的一个品种，尤其是广大的女...

警惕狗贩的骗人损招星期狗的症状特征

警惕狗贩的骗人损招染色：这一招最多的是用在斑点狗、蝴蝶犬...

热门点击排行

2024“最流行”的9种猫，附上最新参考价格，别买贵了！

带宠物被阻，女子“脱光”大闹健身房，以为不要脸就能解决问题？

分享分类导航

萌宠日常

宠物饲养指南

宠物营养食谱

《python深度学习》笔记