CNN模型验证损失随机波动

Question

This is the model accuracy in train and validation

This is the model loss

批处理生成器：

train_batches = ImageDataGenerator(rotation_range=8).flow_from_directory(train_path,target_size=(224,224),  classes=['Covid','Normal','Pneumonia'],batch_size=64)
valid_batches = ImageDataGenerator().flow_from_directory(valid_path,target_size=(224,224), classes=['Covid','Normal','Pneumonia'],batch_size=32)

我使用的是预训练模型：

model=keras.applications.resnet.ResNet50(include_top=False, weights='imagenet', input_tensor=None,  input_shape=(224, 224, 3), pooling=None, classes=1000)

执行正规化： 正则化子=tf.keras.regularizers.l1（0.0001）

for layer in model.layers:
    for attr in ['kernel_regularizer']:
        if hasattr(layer, attr):
          setattr(layer, attr, regularizer)

用于切断预训练模型的最后一个完全连接的层，并添加以下层：

x = AveragePooling2D(pool_size=(4, 4))(last_layer)
x = Flatten(name="flatten")(x)
x = Dense(64, activation="relu",kernel_regularizer=regularizers.l2(0.001))(x)
x = Dropout(0.6)(x)
# x = Dropout(0.6)(x)

out = Dense(3, activation="softmax",name='output_layer')(x)

冻结上层：

for layer in custom_resnet_model.layers[:-7]:
    layer.trainable = False

使用Adam优化器：

custom_resnet_model.compile(Adam(lr=.0001),loss='binary_crossentropy',metrics=['accuracy'])

并且模型适合：

history = custom_resnet_model.fit_generator(train_batches, steps_per_epoch=36,
                    validation_data=valid_batches, validation_steps=18, epochs=25, verbose=2)

正如您在下面看到的，在接近尾声时，验证损失无处不在：

Epoch 1/25
 - 67s - loss: 0.7458 - accuracy: 0.7076 - val_loss: 0.7266 - val_accuracy: 0.7584
Epoch 2/25

 - 64s - loss: 0.5467 - accuracy: 0.8139 - val_loss: 0.5276 - val_accuracy: 0.8022
Epoch 3/25
 - 62s - loss: 0.4723 - accuracy: 0.8543 - val_loss: 0.4393 - val_accuracy: 0.8336
Epoch 4/25
 - 62s - loss: 0.4274 - accuracy: 0.8800 - val_loss: 0.6082 - val_accuracy: 0.8384
Epoch 5/25
 - 62s - loss: 0.4017 - accuracy: 0.8862 - val_loss: 0.5227 - val_accuracy: 0.8490
Epoch 6/25
 - 62s - loss: 0.3698 - accuracy: 0.9004 - val_loss: 0.5691 - val_accuracy: 0.8532
Epoch 7/25
 - 63s - loss: 0.3524 - accuracy: 0.9093 - val_loss: 0.4616 - val_accuracy: 0.8425
Epoch 8/25
 - 63s - loss: 0.3379 - accuracy: 0.9183 - val_loss: 0.4604 - val_accuracy: 0.8467
Epoch 9/25
 - 62s - loss: 0.3206 - accuracy: 0.9248 - val_loss: 0.5499 - val_accuracy: 0.8526
Epoch 10/25
 - 61s - loss: 0.3240 - accuracy: 0.9244 - val_loss: 0.4745 - val_accuracy: 0.8526
Epoch 11/25
 - 63s - loss: 0.3134 - accuracy: 0.9297 - val_loss: 0.4533 - val_accuracy: 0.8567
Epoch 12/25
 - 62s - loss: 0.2995 - accuracy: 0.9337 - val_loss: 0.5668 - val_accuracy: 0.8555
Epoch 13/25
 - 63s - loss: 0.2898 - accuracy: 0.9404 - val_loss: 0.6349 - val_accuracy: 0.8603
Epoch 14/25
 - 62s - loss: 0.2845 - accuracy: 0.9386 - val_loss: 0.5612 - val_accuracy: 0.8650
Epoch 15/25
 - 63s - loss: 0.2961 - accuracy: 0.9330 - val_loss: 0.7284 - val_accuracy: 0.8579
Epoch 16/25
 - 64s - loss: 0.2759 - accuracy: 0.9429 - val_loss: 0.4720 - val_accuracy: 0.8650
Epoch 17/25
 - 62s - loss: 0.2707 - accuracy: 0.9482 - val_loss: 0.9979 - val_accuracy: 0.8650
Epoch 18/25
 - 63s - loss: 0.2744 - accuracy: 0.9416 - val_loss: 0.8098 - val_accuracy: 0.8733
Epoch 19/25
 - 63s - loss: 0.2771 - accuracy: 0.9428 - val_loss: 0.1989 - val_accuracy: 0.8662
Epoch 20/25
 - 62s - loss: 0.2647 - accuracy: 0.9440 - val_loss: 0.8921 - val_accuracy: 0.8686
Epoch 21/25
 - 63s - loss: 0.2566 - accuracy: 0.9478 - val_loss: 0.3362 - val_accuracy: 0.8745
Epoch 22/25
 - 62s - loss: 0.2645 - accuracy: 0.9402 - val_loss: 1.2044 - val_accuracy: 0.8662
Epoch 23/25
 - 63s - loss: 0.2550 - accuracy: 0.9472 - val_loss: 0.6615 - val_accuracy: 0.8745
Epoch 24/25
 - 62s - loss: 0.2486 - accuracy: 0.9519 - val_loss: 0.4722 - val_accuracy: 0.8674
Epoch 25/25
 - 62s - loss: 0.2542 - accuracy: 0.9507 - val_loss: 0.8232 - val_accuracy: 0.8721

我已经发布了代码，以便有人能指出我是否做错了什么。

Answer 1

尝试增加验证数据集。 波动的原因可能是“非代表性验证数据集”。 如果它解决了你的问题，请告诉我

Unrepresentative Validation Dataset