Deep Supervised Hashing in pyTorch

This is a pyTorch implementation of paper: Liu et al. - Deep Supervised Hashing for Fast Image Retrieval (CVPR 2016)

Requirements

• python >=3.6
• pytorch >=1.1 (must havetensorboard included)

Model

• Uses a simple CNN layer, but in this repository the model is a small ResNet.
• Accepts the number of bits it should encode the hash into.

Loss Function

• The loss function is the pairwise ranking loss function with margin and regularization (There is a nice blog post about the description of various loss functions).
• The pairwise loss function forces the distance of similar pairs to be small, and that of dissimilar pairs to be large, with a margin so that the model does not leverage on dissimilar pairs.
• The third term, the regularization term, leads the output binary-like vectors' elements have absolute values close to 1 (an L1 loss with ones tensor).
• Large alpha value introduces more penalty if vector's elements are not close to 1.

Experiments

• The experiment is done with MNIST dataset, and the model is told to encode the class of the image in 8 bits (which can hold max 256 classes)
• The UMAP and t-SNE results show clear clustering of the output vectors.

Hash values

• Here are the list of hash values generated by taking the output vector of the model and encoding as follows:

if element >= 0:
    bit = 1
else:
    bit = 0

class	hash (binary)
0	0xFB (1111 1011)
1	0x39 (0011 1001)
2	0x77 (0111 0111)
3	0xF7 (1111 0111)
4	0x5E (0101 1110)
5	0xED (1110 1101)
6	0xE9 (1110 1001)
7	0xBE (1011 1110)
8	0xC7 (1100 0111)
9	0xFE (1111 1110)

• There exists variations to the generated hash values, so that the output hashes from the same class do not form an exact match, but they span a space within small Hamming distance.
• For example, other output hashes for 3 are: 0xE7, 0xBF, 0xFF which are all within 2 Hamming distance with 0xF7.
• However, the output hash for 2 and 3 have Hamming distance of 1, as well as that of output hashes of 5 and 6.
• This problem could be mitigated by telling the model to produce a vector with larger dimensions, generating larger hash in size.