# MuseGAN Paper
###### tags: `Meeting` `museGAN`
###### editor: Yu-Chun-Hung
###### paper source : [Convolutional Generative Adversarial Networks with Binary Neurons for Polyphonic Music Generation](https://salu133445.github.io/bmusegan/pdf/bmusegan-ismir2018-paper.pdf)
---
## abstract

1. deep convolutional GAN
* generate music by binary valued time-pitch matrices
2. **hard thresholding(HT) or Bernoulli sampling (BS)**
* hard thresholding : 界線值
* Bernoulli sampling : 2項分布
3. the generator and the discriminator are pretrained.
* using binary neurons instead of HT or BS indeed leads to better
4. deterministic binary neurons perform better than stochastic(=random) ones
## Introduction

1. data: multi-track piano-roll, regardless of tempo
* challenging : 大量的possible active notes per time step and the involvement of multiple instruments
* 不像旋律或和弦進行，可以看作一個序列音符/和弦事件，並由RNN建模
* ()RNN是擅長學習音樂的時間依賴性，CNNs) are usually considered better at learning **local patterns**
* **five-level Likert scale** : 音樂評分機制
2. 有幾種方法可以改進這項先前的工作。
* 二值化的方法很容易導致過度碎片化的音符
* binarization of the output of the generator G in GAN is done only at test time not at training time (see Section 2.1 for a brief introduction of GAN).
* the input to the discriminator D in GAN 有效地減少了模型空間

## Background

u $\rightarrow$ ceil function
* DBN: deterministic binary neuron
$DBN(x) = u(\sigma(x) - 0.5)$
* SBN: stochastic binary neuron
$SBN(x) = u(\sigma(x) - v), v \sim U[0, 1]$

DBN is non-differentable, computing exact gradient for SBN is intractable

## Proposed Model

Refiner $R$: refine real-valued $\hat{x} = G(z)$ into binary type $\tilde{x}$, between generator and discriminator

pretrain G and D, and then traind R with D (fixing G)

* refiner : real-value to binary.
* (a) raw predictions : 經過generator產生出來的，沒有經過任何處理
* (b) pretrained (+BS) (*c*) pretrained (+HT) (d) proposed (+SBNs) (d) proposed (+DBNs) : 用(a)做二值化

分2階段訓練:
* 第一階段為 Generator 與 Discriminstor 之 Pre-train
* 第二階段為 Refiner 與 Discriminator 訓練及調整

## Analysis

Qualified Note Rate (QN): 計算符合長度的音符，若太小則表示太分散
Polyphonicity (PP): 在一時間內同時有多音之比例
Tonal Distance (TD): 調性的距離
* training strategies
* joint: 第一階段先訓練 G, R，第二階段再加上 D 一起訓練
* end-to-end: 第一階段 G, R, D 三個共同訓練

## 變奏
https://towardsdatascience.com/ai-music-generation-lead-sheet-composition-and-arrangement-b984208f8519

https://towardsdatascience.com/ai-music-generation-ii-lead-sheet-variation-1884133e4f1

https://github.com/liuhaumin/LeadsheetVAE

#Run inference from a pretrained model
./scripts/run_inference.sh "./exp/default/" **"0"**
:第0個gpu