DCNN实施

Question

我正在此会议论文。

附件的图是模型体系结构： 输入图像大小为（256,256,12），训练图像的总数为90。 分类类别为2。（Pixelwise分类）

这是我的代码： （我不确定HOE能够在纸张中实现50％步伐的设置。）

cnn = Sequential()
cnn.add(Conv2D(100, (8,8), input_shape = (256, 256, 12), activation = 'relu', padding = 'SAME'))
cnn.add(Conv2D(50, (1,1), activation = 'relu'))
cnn.add(Conv2D(30, (1,1), activation = 'relu'))
cnn.add(Conv2D(15, (1,1), activation = 'relu'))
cnn.add(Conv2D(1, (1,1), activation = 'softmax'))

cnn.compile(optimizer = 'adam', loss = 'binary_crossentropy', metrics = ['accuracy'])

model_checkpoint_callback = tf.keras.callbacks.ModelCheckpoint(
    filepath="dcnn_model.h5", 
    verbose=1,
    save_weights_only=True,
    monitor='val_accuracy',
    mode='max',
    save_best_only=True)

cnn.fit(x_train,y_train,
      batch_size=5,
      epochs=30,
      shuffle=True,
      validation_data=(x_val,y_val),
      callbacks=[model_checkpoint_callback])

但是，准确性和损失似乎卡在相同的值处。

Epoch 1/10
30/30 [==============================] - 10s 318ms/step - loss: 3.5711 - accuracy: 0.7658 - val_loss: 3.7894 - val_accuracy: 0.7515

Epoch 00001: val_accuracy improved from -inf to 0.75150, saving model to dcnn_model.h5
Epoch 2/10
30/30 [==============================] - 10s 330ms/step - loss: 3.5711 - accuracy: 0.7658 - val_loss: 3.7894 - val_accuracy: 0.7515

Epoch 00002: val_accuracy did not improve from 0.75150
Epoch 3/10
30/30 [==============================] - 10s 334ms/step - loss: 3.5711 - accuracy: 0.7658 - val_loss: 3.7894 - val_accuracy: 0.7515

Epoch 00003: val_accuracy did not improve from 0.75150
Epoch 4/10
30/30 [==============================] - 10s 334ms/step - loss: 3.5711 - accuracy: 0.7658 - val_loss: 3.7894 - val_accuracy: 0.7515

Epoch 00004: val_accuracy did not improve from 0.75150
Epoch 5/10
30/30 [==============================] - 10s 332ms/step - loss: 3.5711 - accuracy: 0.7658 - val_loss: 3.7894 - val_accuracy: 0.7515

Epoch 00005: val_accuracy did not improve from 0.75150
Epoch 6/10
30/30 [==============================] - 10s 331ms/step - loss: 3.5711 - accuracy: 0.7658 - val_loss: 3.7894 - val_accuracy: 0.7515

Epoch 00006: val_accuracy did not improve from 0.75150
Epoch 7/10
30/30 [==============================] - 10s 334ms/step - loss: 3.5711 - accuracy: 0.7658 - val_loss: 3.7894 - val_accuracy: 0.7515

Epoch 00007: val_accuracy did not improve from 0.75150
Epoch 8/10
30/30 [==============================] - 10s 334ms/step - loss: 3.5711 - accuracy: 0.7658 - val_loss: 3.7894 - val_accuracy: 0.7515

Epoch 00008: val_accuracy did not improve from 0.75150
Epoch 9/10
30/30 [==============================] - 10s 333ms/step - loss: 3.5711 - accuracy: 0.7658 - val_loss: 3.7894 - val_accuracy: 0.7515

如何修改代码？

Answer 1

您拥有的体系结构并不完全像论文，但主要问题是您具有conv2d带有softmax激活功能，而不是sigmoid 。如果您有二进制分类问题，则应使用sigmoid激活功能和1 store>密集 layer，而不是soft> softmax < /代码>。您不能使用softmax和1个密度节点，因为当您使用softmax时，概率的总和必须为1，如果您只有1个节点，则概率永远是1是没有意义的1。