怎样将CNN提取的特征输入SVM模型（tensorflow）

CNN框架如上图所示，模型summary如下:

Model: "sequential"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
conv2d (Conv2D)              (None, 223, 223, 64)      4864      
_________________________________________________________________
batch_normalization_1 (Batch (None, 223, 223, 64)      256       
_________________________________________________________________
max_pooling2d (MaxPooling2D) (None, 111, 111, 64)      0         
_________________________________________________________________
conv2d_1 (Conv2D)            (None, 107, 107, 64)      102464    
_________________________________________________________________
batch_normalization_2 (Batch (None, 107, 107, 64)      256       
_________________________________________________________________
max_pooling2d_1 (MaxPooling2 (None, 53, 53, 64)        0         
_________________________________________________________________
conv2d_2 (Conv2D)            (None, 49, 49, 128)       204928    
_________________________________________________________________
batch_normalization_3 (Batch (None, 49, 49, 128)       512       
_________________________________________________________________
max_pooling2d_2 (MaxPooling2 (None, 24, 24, 128)       0         
_________________________________________________________________
conv2d_3 (Conv2D)            (None, 22, 22, 256)       295168    
_________________________________________________________________
batch_normalization_4 (Batch (None, 22, 22, 256)       1024      
_________________________________________________________________
max_pooling2d_3 (MaxPooling2 (None, 11, 11, 256)       0         
_________________________________________________________________
conv2d_4 (Conv2D)            (None, 9, 9, 512)         1180160   
_________________________________________________________________
batch_normalization_5 (Batch (None, 9, 9, 512)         2048      
_________________________________________________________________
max_pooling2d_4 (MaxPooling2 (None, 4, 4, 512)         0         
_________________________________________________________________
conv2d_5 (Conv2D)            (None, 4, 4, 1024)        525312    
_________________________________________________________________
batch_normalization_6 (Batch (None, 4, 4, 1024)        4096      
_________________________________________________________________
max_pooling2d_5 (MaxPooling2 (None, 2, 2, 1024)        0         
_________________________________________________________________
conv2d_6 (Conv2D)            (None, 2, 2, 1024)        1049600   
_________________________________________________________________
batch_normalization_7 (Batch (None, 2, 2, 1024)        4096      
_________________________________________________________________
max_pooling2d_7 (MaxPooling2 (None, 1, 1, 1024)        0         
_________________________________________________________________
flatten (Flatten)            (None, 1024)              0         
_________________________________________________________________
dropout (Dropout)            (None, 1024)              0         
_________________________________________________________________
dense (Dense)                (None, 1024)              1049600   
_________________________________________________________________
dense_1 (Dense)              (None, 32)                32800     
=================================================================
Total params: 4,457,184
Trainable params: 4,451,040
Non-trainable params: 6,144

接下来我想使用SVM进行分类，请问应该如何操作

仅供参考


model = keras.Sequential([
tf.keras.Input(shape = (227, 227, 3)),
layers.Conv2D (64, (5,5), strides=(1, 1), padding='same', kernel_regularizer = regularizers.l2(0.001), 
               bias_regularizer = regularizers.l2(0.001)),
layers.ReLU(negative_slope= 1),
layers.BatchNormalization(),
layers.MaxPooling2D(pool_size=(2, 2), padding="same"),

layers.Conv2D (64, (5,5), strides=(1, 1), padding='same', kernel_regularizer = regularizers.l2(0.001), 
               bias_regularizer = regularizers.l2(0.001)),
layers.ReLU(negative_slope = 1),
layers.BatchNormalization(),
layers.MaxPooling2D(pool_size=(2, 2)),

layers.Conv2D (128, (5,5), strides=(1, 1), padding='same', kernel_regularizer = regularizers.l2(0.001), 
               bias_regularizer = regularizers.l2(0.001)),
layers.ReLU(negative_slope = 1),
layers.BatchNormalization(),
layers.MaxPooling2D(pool_size=(2, 2)),

layers.Conv2D (256, (3,3), strides=(1, 1), padding='same', kernel_regularizer = regularizers.l2(0.001), 
               bias_regularizer = regularizers.l2(0.001)),
layers.ReLU(negative_slope = 1),
layers.BatchNormalization(),
layers.MaxPooling2D(pool_size=(2, 2), padding="same"),

layers.Conv2D (512, (3,3), strides=(1, 1), padding='same', kernel_regularizer = regularizers.l2(0.001), 
               bias_regularizer = regularizers.l2(0.001)),
layers.ReLU(negative_slope = 1), #negative slope 1 is good
layers.BatchNormalization(),
layers.MaxPooling2D(pool_size=(2, 2), padding="same"),

layers.Conv2D (1024, (1,1), strides=(1, 1), padding='same', kernel_regularizer = regularizers.l2(0.001), 
               bias_regularizer = regularizers.l2(0.001)),
layers.ReLU(negative_slope = 1),
layers.BatchNormalization(),
layers.MaxPooling2D(pool_size=(2, 2), padding="same"),
layers.MaxPooling2D(pool_size=(3, 3), padding="same"),
layers.Reshape((1, 4096)),
layers.Dense(1024, kernel_regularizer = regularizers.l2(0.001), bias_regularizer = regularizers.l2(0.001)),
layers.ReLU(negative_slope = 1),
layers.Dropout(0.5),
layers.Dense(32, name="my_intermediate_layer", kernel_regularizer = regularizers.l2(0.001), 
             bias_regularizer = regularizers.l2(0.001)),
layers.ReLU(negative_slope = 1),
layers.Dense(5, activation='softmax', name="my_last_layer"),
])

model.summary()
model.compile(optimizer=Optimizer.Adam(lr=0.00005),loss='sparse_categorical_crossentropy',metrics=['accuracy'])

trained = model.fit(train_data,train_labels,epochs=30,validation_split=0.10)

model_feat = Model(inputs=model.input,outputs=model.get_layer('my_intermediate_layer').output)
feat_train = model_feat.predict(train_data)
print(feat_train.shape)

feat_train = feat_train.reshape(413,32)

from sklearn.svm import SVC

svm = SVC(kernel='rbf')
svm.fit(feat_train,train_labels)
svm.score(feat_train,train_labels)

完整代码：

import os
import numpy as np
import pandas as pd
import cv2
import warnings
warnings.filterwarnings('ignore')

from sklearn.utils import shuffle

import keras
import tensorflow as tf

from keras import layers
from keras import regularizers
from keras.models import Sequential, Model
import tensorflow.keras.optimizers as Optimizer

# Load data
train_dir = 
test_dir =
train_cate = pd.read_csv()
test_cate = pd.read_csv()

np.array(train_cate[train_cate['Image_name']=='IDRiD_001']['Retinopathy grade'])

# Get data and labels
train_data = []
train_labels = []

def get_images(image_dir, labels_dir):
    
    for image_file in os.listdir(image_dir):
        image = cv2.imread(image_dir+r'/'+image_file)
        image = cv2.resize(image,(227,227))
        train_data.append(image)
        labels = pd.read_csv(labels_dir)
        label = list(labels[labels['Image_name'] + '.jpg' ==  image_file]['Retinopathy grade'])
        train_labels.append(label)
    
    return shuffle(train_data,train_labels,random_state=7)

train_data, train_labels = get_images( , )

train_data = np.array(train_data)
train_labels = np.array(train_labels)

train_labels.reshape(413)

# Build cnn model
model = keras.Sequential([
tf.keras.Input(shape = (227, 227, 3)),
layers.Conv2D (64, (5,5), strides=(1, 1), padding='same', kernel_regularizer = regularizers.l2(0.001), 
               bias_regularizer = regularizers.l2(0.001)),
layers.ReLU(negative_slope= 1),
layers.BatchNormalization(),
layers.MaxPooling2D(pool_size=(2, 2), padding="same"),

layers.Conv2D (64, (5,5), strides=(1, 1), padding='same', kernel_regularizer = regularizers.l2(0.001), 
               bias_regularizer = regularizers.l2(0.001)),
layers.ReLU(negative_slope = 1),
layers.BatchNormalization(),
layers.MaxPooling2D(pool_size=(2, 2)),

layers.Conv2D (128, (5,5), strides=(1, 1), padding='same', kernel_regularizer = regularizers.l2(0.001), 
               bias_regularizer = regularizers.l2(0.001)),
layers.ReLU(negative_slope = 1),
layers.BatchNormalization(),
layers.MaxPooling2D(pool_size=(2, 2)),

layers.Conv2D (256, (3,3), strides=(1, 1), padding='same', kernel_regularizer = regularizers.l2(0.001), 
               bias_regularizer = regularizers.l2(0.001)),
layers.ReLU(negative_slope = 1),
layers.BatchNormalization(),
layers.MaxPooling2D(pool_size=(2, 2), padding="same"),

layers.Conv2D (512, (3,3), strides=(1, 1), padding='same', kernel_regularizer = regularizers.l2(0.001), 
               bias_regularizer = regularizers.l2(0.001)),
layers.ReLU(negative_slope = 1), #negative slope 1 is good
layers.BatchNormalization(),
layers.MaxPooling2D(pool_size=(2, 2), padding="same"),

layers.Conv2D (1024, (1,1), strides=(1, 1), padding='same', kernel_regularizer = regularizers.l2(0.001), 
               bias_regularizer = regularizers.l2(0.001)),
layers.ReLU(negative_slope = 1),
layers.BatchNormalization(),
layers.MaxPooling2D(pool_size=(2, 2), padding="same"),
layers.MaxPooling2D(pool_size=(3, 3), padding="same"),
layers.Reshape((1, 4096)),
layers.Dense(1024, kernel_regularizer = regularizers.l2(0.001), bias_regularizer = regularizers.l2(0.001)),
layers.ReLU(negative_slope = 1),
layers.Dropout(0.5),
layers.Dense(32, name="my_intermediate_layer", kernel_regularizer = regularizers.l2(0.001), 
             bias_regularizer = regularizers.l2(0.001)),
layers.ReLU(negative_slope = 1),
layers.Dense(5, activation='softmax', name="my_last_layer"),
])

model.summary()

model.compile(optimizer=Optimizer.Adam(lr=0.00005),loss='sparse_categorical_crossentropy',metrics=['accuracy'])

# Train
trained = model.fit(train_data,train_labels,epochs=30,validation_split=0.10)

# SVM
model_feat = Model(inputs=model.input,outputs=model.get_layer('my_intermediate_layer').output)
feat_train = model_feat.predict(train_data)
print(feat_train.shape)


feat_train = feat_train.reshape(413,32)

from sklearn.svm import SVC

svm = SVC(kernel='rbf')
svm.fit(feat_train,train_labels)
svm.score(feat_train,train_labels)

数据集：

Leaderboard - Grand Challenge This challenge evaluates automated techniques for analysis of fundus photographs. We target segmentation of retinal lesions like exudates, microaneurysms, and hemorrhages and detection of the optic disc and fovea. Also, we seek grading of fundus images according to the severity level of DR and DME. https://idrid.grand-challenge.org/leaderboard/

你好，我也遇到同样问题，请问您解决了吗？

Indian Diabetic Retinopathy Image Dataset (IDRiD) | IEEE DataPort Access the dataset for images of typical diabetic retinopathy lesions and also normal retinal structures annotated at a pixel level, focused on an Indian population. This dataset provides information on the disease severity of diabetic retinopathy, and diabetic macular edema for each image.

https://ieee-dataport.org/open-access/indian-diabetic-retinopathy-image-dataset-idrid

数据集的其它下载方法，需要注册账号，直接注册即可