Representation Learning Based on Autoencoder and Deep Adaptive Clustering for Image Clustering

Table 3

Network architecture in experiments.

Fashion-MNIST
Encoder	3 × 3 conv., stride 2, 8 BN Relu; 3 × 3 conv., stride 2, 16 BN Relu; 3 × 3 conv., stride 2, 32 BN Relu; 3 × 3 conv., stride 2, 64 BN Relu
Decoder	3 × 3 deconv., stride 2, 32 BN Relu	Clustering	(256,64) fc., BN Relu; (64,10) fc., BN Relu; Softmax
	3 × 3 deconv., stride 2, 16 BN Relu
	3 × 3 deconv., stride 2, 8 BN Relu
	3 × 3 deconv., stride 2, 1
Cifar-10 and Cifar-100 dataset
Encoder	3 × 3 conv., stride 2, 8 BN Relu; 3 × 3 conv., stride 2, 16 BN Relu; 3 × 3 conv., stride 2, 32 BN Relu; 3 × 3 conv., stride 2, 64 BN Relu
Decoder	3 × 3 deconv., stride 2, 32 BN Relu	Clustering	(256,64) fc., BN Relu; (64,k) fc., BN Relu; Softmax
	3 × 3 deconv., stride 2, 16 BN Relu
	3 × 3 deconv., stride 2, 8 BN Relu
	3 × 3 deconv., stride 2, 3
STL-10 dataset
Encoder	5 × 5 conv., stride 4, 16 BN Relu; 3 × 3 conv., stride 2, 32 BN Relu; 3 × 3 conv., stride 2, 64 BN Relu; 3 × 3 conv., stride 2, 128 BN Relu; 3 × 3 conv., stride 1, 256 BN Relu
Decoder	3 × 3 deconv., stride 2, 128 BN Relu	Clustering	FC (256,64) fc., BN Relu; (64,10) fc., BN Relu; Softmax
	3 × 3 deconv., stride 2, 64 BN Relu
	3 × 3 deconv., stride 2, 32 BN Relu
	3 × 3 deconv., stride 2, 16 BN Relu
	3 × 3 deconv., stride 2, 3