Keras猫狗大战五：采用全部数据集训练，精度提高到90%

Posted 2020-12-01 zhengbiqing

深度学习严重依赖训练数据量的大小，前面（https://www.cnblogs.com/zhengbiqing/p/11070783.html）只随机抽取猫狗图片各1000、500、200分别作为训练、验证、测试集，即使采用了数据增强，精度只达到83%。

采用kaggle 猫狗数据集全部25000张进行训练学习，随机选取猫狗图片各9000、2250、1250分别作为训练、验证、测试集，进行训练。

训练100次迭代：

history = model.fit_generator(
    train_generator,
    steps_per_epoch=train_generator.samples // batch_size,
    epochs=100,
    validation_data=validation_generator,
    validation_steps=validation_generator.samples // batch_size)

Epoch 1/100
281/281 [==============================] - 140s 498ms/step - loss: 0.6736 - acc: 0.5683 - val_loss: 0.6532 - val_acc: 0.6074
Epoch 2/100
281/281 [==============================] - 74s 262ms/step - loss: 0.6358 - acc: 0.6310 - val_loss: 0.5861 - val_acc: 0.6803
Epoch 3/100
281/281 [==============================] - 73s 261ms/step - loss: 0.6094 - acc: 0.6625 - val_loss: 0.5559 - val_acc: 0.7146
Epoch 4/100
281/281 [==============================] - 73s 261ms/step - loss: 0.5936 - acc: 0.6807 - val_loss: 0.5511 - val_acc: 0.7216
Epoch 5/100
281/281 [==============================] - 73s 260ms/step - loss: 0.5797 - acc: 0.6918 - val_loss: 0.5308 - val_acc: 0.7347
Epoch 6/100
281/281 [==============================] - 73s 261ms/step - loss: 0.5681 - acc: 0.7018 - val_loss: 0.5318 - val_acc: 0.7356
......

Epoch 96/100
281/281 [==============================] - 74s 264ms/step - loss: 0.2963 - acc: 0.8717 - val_loss: 0.3118 - val_acc: 0.8614
Epoch 97/100
281/281 [==============================] - 74s 263ms/step - loss: 0.2983 - acc: 0.8726 - val_loss: 0.2581 - val_acc: 0.8862
Epoch 98/100
281/281 [==============================] - 74s 263ms/step - loss: 0.2936 - acc: 0.8781 - val_loss: 0.2546 - val_acc: 0.8922
Epoch 99/100
281/281 [==============================] - 74s 263ms/step - loss: 0.3047 - acc: 0.8711 - val_loss: 0.2548 - val_acc: 0.8884
Epoch 100/100
281/281 [==============================] - 74s 263ms/step - loss: 0.3013 - acc: 0.8744 - val_loss: 0.2745 - val_acc: 0.8848

训练曲线：

用测试集对模型进行测试：

test_generator = test_datagen.flow_from_directory(
    test_dir,
    target_size=(150, 150),
    batch_size=batch_size,
    class_mode=‘binary‘)

test_loss, test_acc = model.evaluate_generator(test_generator, steps=test_generator.samples // batch_size)
print(‘test acc:‘, test_acc)

Found 2500 images belonging to 2 classes.
test acc: 0.9

 混淆矩阵：