如何解决这种自动编码器架构的损耗binary_crossentropy停滞在0.601附近
我正在研究一个无监督的图像分类问题,该数据集包含约4700张食肉动物照片。我想到了通过构造自动编码器并获取图像嵌入,然后应用余弦相似度来实现此任务的方法。我没有太大的进步。这是我的自动编码器体系结构:
Model: "functional_75"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
input_19 (InputLayer) [(None,128,3)] 0
_________________________________________________________________
conv2d_126 (Conv2D) (None,64) 1792
_________________________________________________________________
max_pooling2d_54 (MaxPooling (None,64,64) 0
_________________________________________________________________
conv2d_127 (Conv2D) (None,32) 18464
_________________________________________________________________
max_pooling2d_55 (MaxPooling (None,32,32) 0
_________________________________________________________________
conv2d_128 (Conv2D) (None,16) 4624
_________________________________________________________________
max_pooling2d_56 (MaxPooling (None,16,16) 0
_________________________________________________________________
conv2d_129 (Conv2D) (None,16) 2320
_________________________________________________________________
up_sampling2d_54 (UpSampling (None,16) 0
_________________________________________________________________
conv2d_130 (Conv2D) (None,32) 4640
_________________________________________________________________
up_sampling2d_55 (UpSampling (None,32) 0
_________________________________________________________________
conv2d_131 (Conv2D) (None,64) 18496
_________________________________________________________________
conv2d_132 (Conv2D) (None,3) 1731
_________________________________________________________________
up_sampling2d_56 (UpSampling (None,3) 0
=================================================================
Total params: 52,067
Trainable params: 52,067
Non-trainable params: 0
_________________________________________________________________
请提出一些改进建议。
版权声明:本文内容由互联网用户自发贡献,该文观点与技术仅代表作者本人。本站仅提供信息存储空间服务,不拥有所有权,不承担相关法律责任。如发现本站有涉嫌侵权/违法违规的内容, 请发送邮件至 dio@foxmail.com 举报,一经查实,本站将立刻删除。