我正在尝试基于冻结的Inception_v3模型训练一个模型，其中有 3 个类作为输出。当我运行训练时，训练准确率上升，但验证准确度没有提高，验证准确度或多或少精确到 33.33%，即显示完全随机的预测。我不知道我的代码和/或方法中的错误在哪里
我在 Inception v3 核心之后尝试了各种形式的输出，完全没有区别。
# Model definition
# InceptionV3 frozen, flatten, dense 1024, dropout 50%, dense 1024, dense 3, lr 0.001 --> does not train
# InceptionV3 frozen, flatten, dense 1024, dense 3, lr 0.001 --> does not train
# InceptionV3 frozen, flatten, dense 1024, dense 3, lr 0.005 --> does not train
# InceptionV3 frozen, GlobalAvgPooling, dense 1024, dense 1024, dense 512, dense 3, lr 0.001 --> does not train
# InceptionV3 frozen, GlobalAvgPooling dropout 0.4 dense 3, lr 0.001, custom pre-process --> does not train
# InceptionV3 frozen, GlobalAvgPooling dropout 0.4 dense 3, lr 0.001, custom pre-process, batch=32 --> does not train
# InceptionV3 frozen, GlobalAvgPooling dropout 0.4 dense 3, lr 0.001, custom pre-process, batch=32, rebalance train/val sets --> does not train
IMAGE_SIZE = 150
BATCH_SIZE = 32
def build_model(image_size):
input_tensor = tf.keras.layers.Input(shape=(image_size, image_size, 3))
inception_base = InceptionV3(include_top=False, weights='imagenet', input_tensor=input_tensor)
for layer in inception_base.layers:
layer.trainable = False
x = inception_base.output
x = tf.keras.layers.GlobalAveragePooling2D()(x)
x = tf.keras.layers.Dropout(0.2)(x)
output_tensor = tf.keras.layers.Dense(3, activation="softmax")(x)
model = tf.keras.Model(inputs=input_tensor, outputs=output_tensor)
return model
model = build_model(IMAGE_SIZE)
model.compile(optimizer=RMSprop(lr=0.002), loss='categorical_crossentropy', metrics=['acc'])
# Data generators with Image augmentations
train_datagen = ImageDataGenerator(
rescale=1./255,
preprocessing_function=tf.keras.applications.inception_v3.preprocess_input,
rotation_range=40,
width_shift_range=0.2,
height_shift_range=0.2,
shear_range=0.2,
zoom_range=0.2,
horizontal_flip=True,
fill_mode='nearest')
# Do not augment validation!
validation_datagen = ImageDataGenerator(
rescale=1./255,
preprocessing_function=tf.keras.applications.inception_v3.preprocess_input)
train_generator = train_datagen.flow_from_directory(
train_dir,
target_size=(IMAGE_SIZE, IMAGE_SIZE),
batch_size=BATCH_SIZE,
class_mode='categorical')
validation_generator = validation_datagen.flow_from_directory(
valid_dir,
target_size=(IMAGE_SIZE, IMAGE_SIZE),
batch_size=BATCH_SIZE,
class_mode='categorical')
此单元格的输出为：
找到属于 3 类的 1697 张图像。 找到属于 3 类的 712 张图片。
最近两个训练时期的输出：
大纪元 19/20 23/23 [==========================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================7870 - 累计： 0.6912 - val_loss： 1.1930 - val_acc： 0.3174 大纪元 20/20 23/23 [===============================] -6s 255ms/步 - 损失： 1.1985 - 累计： 0.3160
54/54
[=val_acc val_loss===========================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================
唯一让我想到的大事是抛弃rescale=1./255ImageDataGenerators，因为这也由tf.keras.applications.inception_v3.preprocess_input处理，它将 从 -1 扩展到 1; 网络的预期输入。