互联网上的大多数例子都是关于 multi-label
图像分类仅基于一个 few
标签。例如,与 6
类,我们得到。
model = models.Sequential()
model.add(layer=base)
model.add(layer=layers.Flatten())
model.add(layer=layers.Dense(units=256, activation="relu"))
model.add(layer=layers.Dense(units=6, activation="sigmoid"))
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
vgg16 (Model) (None, 7, 7, 512) 14714688
_________________________________________________________________
flatten_1 (Flatten) (None, 25088) 0
_________________________________________________________________
dense_1 (Dense) (None, 256) 6422784
_________________________________________________________________
dense_2 (Dense) (None, 6) 1542
=================================================================
Total params: 21,139,014
Trainable params: 13,503,750
Non-trainable params: 7,635,264
然而,对于数据集与 significantly
更多的标签,训练的大小 parameters
爆炸,最终训练过程失败,并伴有 ResourceExhaustedError
错误。例如,对于 3047
我们得到的标签。
model = models.Sequential()
model.add(layer=base)
model.add(layer=layers.Flatten())
model.add(layer=layers.Dense(units=256, activation="relu"))
model.add(layer=layers.Dense(units=3047, activation="sigmoid"))
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
vgg16 (Model) (None, 7, 7, 512) 14714688
_________________________________________________________________
flatten_1 (Flatten) (None, 25088) 0
_________________________________________________________________
dense_1 (Dense) (None, 256) 6422784
_________________________________________________________________
dense_2 (Dense) (None, 3047) 783079
=================================================================
Total params: 21,920,551
Trainable params: 14,285,287
Non-trainable params: 7,635,264
_________________________________________________________________
很明显,我的网络出了点问题 但不知道如何解决这个问题...
资源耗尽错误与内存问题有关。要么是你的系统中没有足够的内存,要么是代码的其他部分造成了内存问题。