Diversity Denoising with Fully Convolutional Variational Autoencoders Reimplementation for Deep Learning Course in Koç University
This repo will contain the Julia implementation of the paper Diversity Denoising with Fully Convolutional Variational Autoencoders with the Koç University Deep Learning Framework (Knet).
To start training, you can use DenoiSegMouse and W2S modules. First include them by calling '''include("train_mouse.jl")''' or ''include("train_W2S.jl")'''.
usage: <DenoiSegMouse> [--seed SEED] [--batchsize BATCHSIZE]
[--epochs EPOCHS] [--nf NF] [--lr LR]
[--gauss_std GAUSS_STD] [--kl_anneal KL_ANNEAL]
[--minvar MINVAR] [--dataset DATASET]
Fully Unsupervised or Unsupervised Training on DenoiSeg Mouse and
DenoiSeg Mouse s&p datasets.
optional arguments:
--seed SEED random number seed: use a nonnegative int for
repeatable results (type: Int64, default: -1)
--batchsize BATCHSIZE
minibatch size (type: Int64, default: 32)
--epochs EPOCHS number of epochs for training (type: Int64,
default: 1000)
--nf NF number of filters for first conv layer of
encoder (type: Int64, default: 32)
--lr LR initial learning rate (type: Float64, default:
0.001)
--gauss_std GAUSS_STD
Gaussian noise assumption for the dataset,if
it is negative fully unsupervised traning is
performed (type: Float64, default: -1.0)
--kl_anneal KL_ANNEAL
How many epochs to perform kl annealing
starting from first epoch. [To prevent
posterior collapse, KL annealing aproach might
be needed.] (type: Int64, default: 0)
--minvar MINVAR minimum allowed variance required for
[original implementation uses 9.0 for Mouse
and 1.0 for Mouse s&p] (type: Float64,
default: -1.0)
--dataset DATASET Choose 0 for Mouse and 1 for Mouse s&p (type:
Int64, default: 0)
usage: <W2S> [--seed SEED] [--batchsize BATCHSIZE]
[--epochs EPOCHS] [--nf NF] [--lr LR]
[--channel CHANNEL] [--avg AVG] [--minvar MINVAR]
Fully Unsupervised Training on W2S datasets.
optional arguments:
--seed SEED random number seed: use a nonnegative int for
repeatable results (type: Int64, default: -1)
--batchsize BATCHSIZE
minibatch size (type: Int64, default: 32)
--epochs EPOCHS number of epochs for training (type: Int64,
default: 1000)
--nf NF number of filters for first conv layer of
encoder (type: Int64, default: 32)
--lr LR initial learning rate (type: Float64, default:
0.001)
--channel CHANNEL W2S dataset channel to train 0,1 or 2 (type:
Int64)
--avg AVG W2S 1 or 16 corresponds to Avg1 and Avg 16
respectively (type: Int64)
--minvar MINVAR minimum allowed variance [original
implementation uses 1.0 for Avg16 9.0 for
Avg1] (type: Float64)
Please check train.ipynb for details. It shows how to reproduce my results.
| Dataset | Fully Unsupervised | Paper Reported Fully Unsupervised | Unsupervised | Paper Reported Unsupervised |
|---|---|---|---|---|
| DenoiSeg Mouse | 34.11 dB | 34.06 dB | 34.19 dB | 34.13 dB |
| DenoiSeg Mouse s&p | 36.20 dB | 35.19 dB | __ | 36.21 dB |
| DenoiSeg Flywing | 24.79 dB | 24.92 dB | 25.12 dB | 25.02 dB |
| W2S Ch.0 Avg 1 | 34.36 dB | 34.24 dB | - | __ |
| W2S Ch.1 Avg 1 | 32.24 dB | 32.22 dB | - | __ |
| W2S Ch.2 Avg 1 | 35.31 dB | 35.24 dB | - | __ |
| W2S Ch.0 Avg 16 | 39.60 dB | 39.45 dB | - | __ |
| W2S Ch.1 Avg 16 | 38.46 dB | 38.41 dB | - | __ |
| W2S Ch.2 Avg 16 | 40.32 dB | 40.56 dB | - | __ |
@article{2020DivNoising,
title={DivNoising: Diversity Denoising with Fully Cnvolutional Variational Autoencoders},
author={Prakash, Mangal and Alexander Krull and Jug, Florian},
journal={arXiv preprint arXiv:2006.06072},
year={2020}
}