CIFAR-10 在 Keras 上使用 MobileNet/VGG16 的准确率无法超过 10%答案

【问题标题】：CIFAR-10 can't get above 10% Accuracy with MobileNet/VGG16 on KerasCIFAR-10 在 Keras 上使用 MobileNet/VGG16 的准确率无法超过 10%
【发布时间】：2020-07-17 15:47:48
【问题描述】：

我正在尝试使用 CIFAR10 数据集训练 mobileNet 和 VGG16 模型，但准确率不能超过 9.9%。我需要完整的模型 (include_top=True) 并且没有来自 imagenet 的 wights。

P.S.：我尝试增加/减少 dropout 和学习率，并更改了优化器，但我始终保持相同的准确度。

train_generator = ImageDataGenerator(
                                    rotation_range=2, 
                                    horizontal_flip=True,
                                    zoom_range=.1 )
val_generator = ImageDataGenerator(
                                    rotation_range=2, 
                                    horizontal_flip=True,
                                    zoom_range=.1)

train_generator.fit(x_train)
val_generator.fit(x_val)


batch_size= 100
epochs=50

learn_rate=.001

sgd=SGD(lr=learn_rate,momentum=.9,nesterov=False)
adam=Adam(lr=learn_rate, beta_1=0.9, beta_2=0.999, epsilon=None, decay=0.0, amsgrad=False)

model_1.compile(optimizer=adam,loss='sparse_categorical_crossentropy',metrics=['accuracy'])

model_1.fit_generator(train_generator.flow(x_train,y_train,batch_size=batch_size),
                      epochs=epochs,
                      steps_per_epoch=x_train.shape[0]//batch_size,
                      validation_data=val_generator.flow(x_val,y_val,batch_size=batch_size),validation_steps=250,
                      verbose=1)

MobileNet 的结果：

    Epoch 1/50
350/350 [==============================] - 17s 50ms/step - loss: nan - accuracy: 0.0990 - val_loss: nan - val_accuracy: 0.1021
Epoch 2/50
350/350 [==============================] - 17s 49ms/step - loss: nan - accuracy: 0.0990 - val_loss: nan - val_accuracy: 0.1030
Epoch 3/50
350/350 [==============================] - 17s 49ms/step - loss: nan - accuracy: 0.0990 - val_loss: nan - val_accuracy: 0.1016
Epoch 4/50
350/350 [==============================] - 17s 49ms/step - loss: nan - accuracy: 0.0990 - val_loss: nan - val_accuracy: 0.1014
Epoch 5/50
350/350 [==============================] - 17s 49ms/step - loss: nan - accuracy: 0.0990 - val_loss: nan - val_accuracy: 0.1040
Epoch 6/50
350/350 [==============================] - 17s 49ms/step - loss: nan - accuracy: 0.0990 - val_loss: nan - val_accuracy: 0.1009
Epoch 7/50
350/350 [==============================] - 17s 49ms/step - loss: nan - accuracy: 0.0990 - val_loss: nan - val_accuracy: 0.1035
Epoch 8/50
350/350 [==============================] - 17s 49ms/step - loss: nan - accuracy: 0.0990 - val_loss: nan - val_accuracy: 0.1013
Epoch 9/50
350/350 [==============================] - 17s 49ms/step - loss: nan - accuracy: 0.0990 - val_loss: nan - val_accuracy: 0.1029
Epoch 10/50
350/350 [==============================] - 17s 49ms/step - loss: nan - accuracy: 0.0990 - val_loss: nan - val_accuracy: 0.1023
Epoch 11/50
350/350 [==============================] - 17s 49ms/step - loss: nan - accuracy: 0.0990 - val_loss: nan - val_accuracy: 0.1017
Epoch 12/50
350/350 [==============================] - 17s 49ms/step - loss: nan - accuracy: 0.0990 - val_loss: nan - val_accuracy: 0.1020
Epoch 13/50
350/350 [==============================] - 17s 49ms/step - loss: nan - accuracy: 0.0990 - val_loss: nan - val_accuracy: 0.1020
Epoch 14/50
350/350 [==============================] - 17s 49ms/step - loss: nan - accuracy: 0.0990 - val_loss: nan - val_accuracy: 0.1033
Epoch 15/50
350/350 [==============================] - 17s 49ms/step - loss: nan - accuracy: 0.0990 - val_loss: nan - val_accuracy: 0.1011
Epoch 16/50
350/350 [==============================] - 17s 49ms/step - loss: nan - accuracy: 0.0990 - val_loss: nan - val_accuracy: 0.1016
Epoch 17/50
350/350 [==============================] - 17s 49ms/step - loss: nan - accuracy: 0.0990 - val_loss: nan - val_accuracy: 0.1024
Epoch 18/50
350/350 [==============================] - 17s 49ms/step - loss: nan - accuracy: 0.0990 - val_loss: nan - val_accuracy: 0.1024
Epoch 19/50
350/350 [==============================] - 17s 49ms/step - loss: nan - accuracy: 0.0990 - val_loss: nan - val_accuracy: 0.1041
Epoch 20/50
350/350 [==============================] - 17s 49ms/step - loss: nan - accuracy: 0.0990 - val_loss: nan - val_accuracy: 0.1010
Epoch 21/50
350/350 [==============================] - 17s 49ms/step - loss: nan - accuracy: 0.0990 - val_loss: nan - val_accuracy: 0.1022
Epoch 22/50
350/350 [==============================] - 17s 49ms/step - loss: nan - accuracy: 0.0990 - val_loss: nan - val_accuracy: 0.1014
Epoch 23/50
350/350 [==============================] - 17s 49ms/step - loss: nan - accuracy: 0.0990 - val_loss: nan - val_accuracy: 0.1035
Epoch 24/50
350/350 [==============================] - 17s 49ms/step - loss: nan - accuracy: 0.0990 - val_loss: nan - val_accuracy: 0.1032
Epoch 25/50
350/350 [==============================] - 17s 49ms/step - loss: nan - accuracy: 0.0990 - val_loss: nan - val_accuracy: 0.1012
Epoch 26/50
350/350 [==============================] - 17s 49ms/step - loss: nan - accuracy: 0.0990 - val_loss: nan - val_accuracy: 0.1018
Epoch 27/50
350/350 [==============================] - 17s 49ms/step - loss: nan - accuracy: 0.0990 - val_loss: nan - val_accuracy: 0.1022
Epoch 28/50
350/350 [==============================] - 17s 49ms/step - loss: nan - accuracy: 0.0990 - val_loss: nan - val_accuracy: 0.1031
Epoch 29/50
350/350 [==============================] - 17s 49ms/step - loss: nan - accuracy: 0.0990 - val_loss: nan - val_accuracy: 0.1022
Epoch 30/50
350/350 [==============================] - 17s 49ms/step - loss: nan - accuracy: 0.0990 - val_loss: nan - val_accuracy: 0.1015
Epoch 31/50
350/350 [==============================] - 17s 49ms/step - loss: nan - accuracy: 0.0990 - val_loss: nan - val_accuracy: 0.1028
Epoch 32/50
350/350 [==============================] - 17s 49ms/step - loss: nan - accuracy: 0.0990 - val_loss: nan - val_accuracy: 0.1015
Epoch 33/50
350/350 [==============================] - 17s 49ms/step - loss: nan - accuracy: 0.0990 - val_loss: nan - val_accuracy: 0.1030
Epoch 34/50
350/350 [==============================] - 17s 49ms/step - loss: nan - accuracy: 0.0990 - val_loss: nan - val_accuracy: 0.1003
Epoch 35/50
350/350 [==============================] - 17s 49ms/step - loss: nan - accuracy: 0.0990 - val_loss: nan - val_accuracy: 0.1044
Epoch 36/50
350/350 [==============================] - 17s 49ms/step - loss: nan - accuracy: 0.0990 - val_loss: nan - val_accuracy: 0.1012
Epoch 37/50
350/350 [==============================] - 17s 49ms/step - loss: nan - accuracy: 0.0990 - val_loss: nan - val_accuracy: 0.1022
Epoch 38/50
350/350 [==============================] - 17s 49ms/step - loss: nan - accuracy: 0.0990 - val_loss: nan - val_accuracy: 0.1021
Epoch 39/50
350/350 [==============================] - 17s 49ms/step - loss: nan - accuracy: 0.0990 - val_loss: nan - val_accuracy: 0.1028
Epoch 40/50
350/350 [==============================] - 17s 49ms/step - loss: nan - accuracy: 0.0990 - val_loss: nan - val_accuracy: 0.1012
Epoch 41/50
350/350 [==============================] - 17s 49ms/step - loss: nan - accuracy: 0.0990 - val_loss: nan - val_accuracy: 0.1035
Epoch 42/50
350/350 [==============================] - 17s 49ms/step - loss: nan - accuracy: 0.0990 - val_loss: nan - val_accuracy: 0.1009
Epoch 43/50
350/350 [==============================] - 17s 49ms/step - loss: nan - accuracy: 0.0990 - val_loss: nan - val_accuracy: 0.1034
Epoch 44/50
350/350 [==============================] - 17s 49ms/step - loss: nan - accuracy: 0.0990 - val_loss: nan - val_accuracy: 0.1024
Epoch 45/50
350/350 [==============================] - 17s 49ms/step - loss: nan - accuracy: 0.0990 - val_loss: nan - val_accuracy: 0.1016
Epoch 46/50
350/350 [==============================] - 17s 49ms/step - loss: nan - accuracy: 0.0990 - val_loss: nan - val_accuracy: 0.1028
Epoch 47/50
350/350 [==============================] - 17s 49ms/step - loss: nan - accuracy: 0.0990 - val_loss: nan - val_accuracy: 0.1016
Epoch 48/50
350/350 [==============================] - 17s 49ms/step - loss: nan - accuracy: 0.0990 - val_loss: nan - val_accuracy: 0.1033
Epoch 49/50
350/350 [==============================] - 17s 49ms/step - loss: nan - accuracy: 0.0990 - val_loss: nan - val_accuracy: 0.1018
Epoch 50/50
350/350 [==============================] - 17s 49ms/step - loss: nan - accuracy: 0.0990 - val_loss: nan - val_accuracy: 0.1023

<tensorflow.python.keras.callbacks.History at 0x7fa30b188e48>

有没有办法解决这个问题？

【问题讨论】：

你能告诉我们模型代码，你是如何创建的吗？
我猜这些层没有设置为可训练
@mujjiga 我没有创建我只是导入了它，模型集成在Keras中。
@AndrewHolmgren 我真的这么认为..
@mujjiga 这里：model_1 = MobileNet(include_top=True, weights=None, input_shape=(32,32,3), classes=y_train.shape[1])

标签： python tensorflow machine-learning keras deep-learning

【解决方案1】：

我经常使用 MobileNet 模型，它运行良好。要使用它，请参阅下面的代码。

 
 mobile = tf.keras.applications.mobilenet.MobileNet( include_top=False,
                                                           input_shape=(image_size,image_size,3),
                                                           pooling='max', weights='imagenet',
                                                           alpha=1, depth_multiplier=1,dropout=.4)
  x=mobile.layers[-1].output
  x=keras.layers.BatchNormalization(axis=-1, momentum=0.99, epsilon=0.001 )(x)
  predictions=Dense (len(classes), activation='softmax')(x)
  model = Model(inputs=mobile.input, outputs=predictions)    
  for layer in model.layers:
      layer.trainable=True
  model.compile(Adamax(lr=lr_rate), loss='categorical_crossentropy', metrics=['accuracy'])

Classes 是数据集中的类数。图像大小是图像的大小，以像素为单位。s

【讨论】：

这实际上不是问题，使用 weights='imagenet' 和 include_top=False 我可以达到 90% 以上的准确率，但我想在没有这些参数的情况下训练模型。
如果您离开 top=True，您的最后一层将具有与原始 VGG16 模型一样多的类，我相信是 1000。也许这就是为什么您的损失是 nan（不是数字）我没有t 看了，但我相信 CIFAR10 数据集没有 1000 个类。您必须定制顶层以拥有与类一样多的节点。
想多了。我想网络可以使用冻结的随机权重进行学习。顶部=假。 Thr VGG 网络将对每个图像应用一个固定的变换，也许密集层仍然可以学习。