tensorflow - 使用 Keras 进行图像着色 - val_loss 振荡

Question

我有这个顺序模型：

conv2d (Conv2D)              (None, 256, 256, 32)      320
_________________________________________________________________
conv2d_1 (Conv2D)            (None, 128, 128, 64)      18496
_________________________________________________________________
conv2d_2 (Conv2D)            (None, 128, 128, 64)      36928
_________________________________________________________________
conv2d_3 (Conv2D)            (None, 64, 64, 128)       73856
_________________________________________________________________
conv2d_4 (Conv2D)            (None, 64, 64, 128)       147584
_________________________________________________________________
conv2d_5 (Conv2D)            (None, 32, 32, 256)       295168
_________________________________________________________________
conv2d_6 (Conv2D)            (None, 32, 32, 256)       590080
_________________________________________________________________
conv2d_7 (Conv2D)            (None, 32, 32, 128)       295040
_________________________________________________________________
conv2d_transpose (Conv2DTran (None, 64, 64, 128)       147584
_________________________________________________________________
conv2d_8 (Conv2D)            (None, 64, 64, 64)        73792
_________________________________________________________________
conv2d_transpose_1 (Conv2DTr (None, 128, 128, 64)      36928
_________________________________________________________________
conv2d_9 (Conv2D)            (None, 128, 128, 32)      18464
_________________________________________________________________
conv2d_transpose_2 (Conv2DTr (None, 256, 256, 32)      9248
_________________________________________________________________
conv2d_10 (Conv2D)           (None, 256, 256, 16)      4624
_________________________________________________________________
conv2d_11 (Conv2D)           (None, 256, 256, 2)       290
=================================================================

我想给黑白（256x256）肖像上色。我的数据集大小：7650

我尝试使用 Adamax 和 RMSprop。准确度和损失还可以，但是 val_loss 和 val_accuracy 只是振荡。问题出在哪里？

标签： tensorflowkerasautoencoder

Answer 1

看来您的模型过拟合了，您应该减少网络中的参数数量：减少权重层或减少过滤器。

此外，您可以使用技术来减少过度拟合：Dropout、BatchNormalization、正则化......

最后，您可以使用数据增强（根据您的数据集）创建新图像，例如图像翻转、移位、旋转、裁剪......