- Resnet Core -> hash layers
- Resnet Core -> (decoder, hash layers)
- Encoder -> Decoder
-
Train Code and Architecture changed from previous codes of autoencoder
-
Used GAP to join 32 feature tensors of encoder to a linear layer(in=32, out=len_hash_vector) Difficult for model to backprop. Hash loss is difficult to minimize
- Alternatively, direct connection or pooling can pe done Seems better as of now
- Currently on linear bottleneck between encoder and decoder with bad image reconstruction but okay hashes
-
Losses namely: cosine cross entropy, cauchy quantization (better than mse for hashes), mse, mssim (seems useless as of now)
- Concept of triplet loss, and modified cauchy losses from MICCAI2019 paper are NOT IMPLEMENTED