代码之家 › 专栏 › 技术社区 › mad

用异常cnn训练自定义图像的奇怪错误

deep-learning keras tensorflow python

mad · 技术社区 · 5 年前

我有一系列的小图片,我想运行一个例外CNN的培训使用它们。训练集和验证集分别具有以下形状:

>(63787, 72, 72, 3) 

>(155, 72, 72, 3)

换句话说,我的图像满足71,71,3最小形状的异常输入要求。

这就是我构建模型的方式

    base_model=xception.Xception(include_top=False, weights=None, input_shape=(72, 72, 3))
    x = base_model.output
    x = Dense(256, activation='relu')(x)
    predictions = Dense(classes, activation='softmax')(x)
    model = Model(inputs=base_model.input, outputs=predictions)
    opt= SGD(lr=0.0001, decay=1e-6, momentum=0.9, nesterov=True)
    model.compile(loss="sparse_categorical_crossentropy", optimizer=opt, metrics=["accuracy"])

我就是这样训练模特的

checkpoint = ModelCheckpoint(filename, monitor='val_acc', verbose=1, save_best_only=True, save_weights_only=False, mode='max', period=self.__model_history_period)
lr_reducer = ReduceLROnPlateau(factor=np.sqrt(0.1), cooldown=0, patience=1000, min_lr=0.5e-6)
early_stopper = EarlyStopping(min_delta=0.0001, patience=10000)
callbacks_list = [checkpoint, lr_reducer, early_stopper]

datagen = ImageDataGenerator(width_shift_range=0.1, height_shift_range=0.1, horizontal_flip=True, vertical_flip=True) 
datagen.fit(x_train)
 # Fit the model on the batches generated by datagen.flow().
H=model.fit_generator(datagen.flow(X_train, y_train,batch_size=self.__bs), steps_per_epoch=X_train.shape[0] // self.__bs, validation_data=(x_val, y_val), epochs=self.__epochs, verbose=1, max_q_size=100, callbacks=[checkpoint, lr_reducer, early_stopper])

但是,当我运行CNN培训时,我有以下错误:

> Traceback (most recent call last):
  File "esperimento_paper.py", line 86, in <module>
    vgg.run_2D()
  File "Desktop/PhD-Market-Nets/src/classes/VggHandler.py", line 662, in run_2D
    model, H, n_epochs = self.__train_2D(x_train=x_train, y_train=y_train, x_val=x_val, y_val=y_val, index_net=index_net, index_walk=index_walk)
  File "Desktop/PhD-Market-Nets/src/classes/VggHandler.py", line 267, in __train_2D
    callbacks=[checkpoint, lr_reducer, early_stopper])  
  File "Desktop/PhD-Market-Nets/venv/lib/python3.6/site-packages/keras/legacy/interfaces.py", line 91, in wrapper
    return func(*args, **kwargs)
  File "Desktop/PhD-Market-Nets/venv/lib/python3.6/site-packages/keras/engine/training.py", line 1418, in fit_generator
    initial_epoch=initial_epoch)
  File "Desktop/PhD-Market-Nets/venv/lib/python3.6/site-packages/keras/engine/training_generator.py", line 144, in fit_generator
    val_x, val_y, val_sample_weight)
  File "Desktop/PhD-Market-Nets/venv/lib/python3.6/site-packages/keras/engine/training.py", line 789, in _standardize_user_data
    exception_prefix='target')
  File "Desktop/PhD-Market-Nets/venv/lib/python3.6/site-packages/keras/engine/training_utils.py", line 128, in standardize_input_data
    'with shape ' + str(data_shape))
ValueError: Error when checking target: expected dense_2 to have 4 dimensions, but got array with shape (155, 1)

我错过了什么?

1 回复 | 直到 5 年前

Marcin MoÅ¼ejko 5 年前

问题就在这一行:

x = base_model.output
x = Dense(256, activation='relu')(x)

正如你所设定的 include_top=False -然后返回的是形状的特征映射 [number of examples, h, w, number of features] 是的。当你申请 Dense 对于这个特性图-您只将它应用于最后一个维度(它类似于1x1卷积)。这就是为什么输出是 4D 是的。为了克服这种尝试:

x = base_model.output
x = Flatten()(x)
x = Dense(256, activation='relu')(x)

Flatten 会压扁 h 我是说, w 和 number of features 维度到单个维度。多亏了这一点,你的网络应该运行良好。

另外,你也可以尝试使用 GlobalMaxPooling2D 或者它的平均版本。这将跳过过滤器的空间位置,但将显著降低模型的内存占用率。