4.2 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=(200, 200, 3), activation='relu')) 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.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.MaxPooling2D()) model.add(tf.keras.layers.Conv2D(512, (3, 3), activation='relu')) model.add(tf.keras.layers.MaxPooling2D()) model.add(tf.keras.layers.Conv2D(512, (3, 3), activation='relu')) model.add(tf.keras.layers.MaxPooling2D()) model.add(tf.keras.layers.Conv2D(1024, (3, 3), activation='relu')) model.add(tf.keras.layers.GlobalAveragePooling2D()) model.add(tf.keras.layers.Dense(1024, activation='relu')) model.add(tf.keras.layers.Dense(256, activation='relu')) model.add(tf.keras.layers.Dense(1, activation='sigmoid')) #%% model.summary() model.compile(optimizer=keras.optimizers.Adam(learning_rate=0.0001), loss='binary_crossentropy', metrics=['acc'])
Model: "sequential" _________________________________________________________________ Layer (type) Output Shape Param # ================================================================= conv2d (Conv2D) (None, 198, 198, 64) 1792 _________________________________________________________________ conv2d_1 (Conv2D) (None, 196, 196, 64) 36928 _________________________________________________________________ max_pooling2d (MaxPooling2D) (None, 98, 98, 64) 0 _________________________________________________________________ conv2d_2 (Conv2D) (None, 96, 96, 128) 73856 _________________________________________________________________ conv2d_3 (Conv2D) (None, 94, 94, 128) 147584 _________________________________________________________________ max_pooling2d_1 (MaxPooling2 (None, 47, 47, 128) 0 _________________________________________________________________ conv2d_4 (Conv2D) (None, 45, 45, 256) 295168 _________________________________________________________________ conv2d_5 (Conv2D) (None, 43, 43, 256) 590080 _________________________________________________________________ max_pooling2d_2 (MaxPooling2 (None, 21, 21, 256) 0 _________________________________________________________________ conv2d_6 (Conv2D) (None, 19, 19, 512) 1180160 _________________________________________________________________ max_pooling2d_3 (MaxPooling2 (None, 9, 9, 512) 0 _________________________________________________________________ conv2d_7 (Conv2D) (None, 7, 7, 512) 2359808 _________________________________________________________________ max_pooling2d_4 (MaxPooling2 (None, 3, 3, 512) 0 _________________________________________________________________ conv2d_8 (Conv2D) (None, 1, 1, 1024) 4719616 _________________________________________________________________ global_average_pooling2d (Gl (None, 1024) 0 _________________________________________________________________ dense (Dense) (None, 1024) 1049600 _________________________________________________________________ dense_1 (Dense) (None, 256) 262400 _________________________________________________________________ dense_2 (Dense) (None, 1) 257 ================================================================= Total params: 10,717,249 Trainable params: 10,717,249 Non-trainable params: 0 _________________________________________________________________
steps_per_eooch = train_count//batch_size validation_steps = test_count//batch_size history = model.fit(train_ds,epochs=100,steps_per_epoch=steps_per_eooch,validation_data=test_ds,validation_steps=validation_steps)
Epoch 1/100 66/66 [==============================] - 11s 171ms/step - loss: 0.6936 - acc: 0.5208 - val_loss: 0.6931 - val_acc: 0.4833 Epoch 2/100 66/66 [==============================] - 11s 169ms/step - loss: 0.6902 - acc: 0.5312 - val_loss: 0.6919 - val_acc: 0.5502 Epoch 3/100 66/66 [==============================] - 11s 168ms/step - loss: 0.6851 - acc: 0.5611 - val_loss: 0.6799 - val_acc: 0.6105 Epoch 4/100 66/66 [==============================] - 17s 254ms/step - loss: 0.6650 - acc: 0.6080 - val_loss: 0.6568 - val_acc: 0.6228 Epoch 5/100 66/66 [==============================] - 17s 257ms/step - loss: 0.6341 - acc: 0.6501 - val_loss: 0.6466 - val_acc: 0.6395 ....... Epoch 98/100 66/66 [==============================] - 17s 256ms/step - loss: 4.7811e-08 - acc: 1.0000 - val_loss: 3.8404 - val_acc: 0.7422 Epoch 99/100 66/66 [==============================] - 17s 259ms/step - loss: 5.1801e-08 - acc: 1.0000 - val_loss: 3.8481 - val_acc: 0.7433 Epoch 100/100 66/66 [==============================] - 17s 256ms/step - loss: 4.4815e-08 - acc: 1.0000 - val_loss: 3.8532 - val_acc: 0.7444
4.3 分析评估
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。
history.history.keys() plt.plot(history.epoch, history.history.get('acc'), label='acc') plt.plot(history.epoch, history.history.get('val_acc'), label='val_acc') plt.legend()
plt.plot(history.epoch, history.history.get('loss'), label='loss') plt.plot(history.epoch, history.history.get('val_loss'), label='val_loss') plt.legend()
最后评估一下在测试集上的误差
#test_ds = test_data.batch(batch_size) loss,acc = model.evaluate(test_data.batch(batch_size))
32/32 [==============================] - 4s 111ms/step - loss: 3.8820 - acc: 0.7300
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73%,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。
