如何在使用 CNN 提取特征的同时使用 SVM 进行分类？

Question

   conv12 = Conv2D(32, (3, 3), activation='relu', padding='same')(up12)
   conv12 = Dropout(0.3)(conv12)
   conv12 = Conv2D(32, (3, 3), activation='relu', padding='same')(conv12)
   
   conv13 = Conv2D(1, (1, 1), activation='sigmoid')(conv12)

   model = Model(inputs=[inputs], outputs=[conv13])

   model.compile(optimizer=Adam(lr=.00045), loss=dice_coef_loss, metrics=[dice_coef])

   return model

After the conv13 layer i want to use SVM, how can i do that? I am new to this and can't figure this out.

Answer 1

Your neural network produces an output given an input (an image I assume). This output, as you set it to be the parameters from the conv13 layer, will be a vector of a certain size. You can now consider this output as input for your SVM classifier. However, you do not need to stick to Keras for this step, as libraries like scikit-learn have implemented an easier way to do that.

Let's say your CNN produces a set of vectors like X =[95, 25, ..., 45, 24] as output. If you then have a set of labels y = {0, 1} then you can do:

from sklearn.svm import SVC

clf = SVC() # Play with hyperparameters.

clf.predict([[95, 25, ..., 45, 24]]) # Output: array([1])

See: https://scikit-learn.org/stable/modules/svm.html