# 23: Super-resolution
或Deep Learning Foundations to Stable Diffusion
主要討論:深度學習模型在圖像處理(應用/改進)

## FID
首先他說了 notebook 23章 ~~ [~~(我找了很久)~~](https://github.com/fastai/course22p2)~~ FID (Fréchet Inception Distance) 的bug
> Fréchet Inception Distance 是什麼? 用來評估模型產生圖像品質和多樣性的重要指標/生成圖像與真實圖像在特徵空間中的相似度。
> 特徵提取>計算>Fréchet 距離計算
> value 越低越好, if ==0 即生成的和真實的圖像分布完全相同~ 完美~

BUG:
cos model: -0.5 to +0.5
sampling x2
** dependent variable is just "noise"

## U-Nets for super-resolution
然後就到Tiny Imagenet

fashion MNIST:最大的1training size 只有28x28
Tiny Imagenet: 64x64 [壞處: 很難找，只有standford 找到]

SHUTIL <-解壓用~

STEP1:
: create data set
class TinyDS:
    def __init__(self, path):
        self.path = Path(path)
        self.files = glob(str(path/'**/*.JPEG'), recursive=True) <-要小心recursive
    def __len__(self): return len(self.files)
    def __getitem__(self, i): return self.files[i],Path(self.files[i]).parent.parent.name
    
tds = TinyDS(path/'train')  <-

Label 是分開放在val
如上

> shift 提示大法~

### data augmentation
random resize crop
pick one area inside and zoom into it

可是tiny imagenet 太小/細了，太poor
=> blur

貓貓是很可愛啦，有貓就讚，但是我以為自己沒有戴眼鏡呢

所以padding around them
然後隨機從中找64x64

用random array
V
用transform batch passing in those transforms
(nn dot sequential: called one time in a row)
~~(我也覺得沒什麼神奇特別啦...)~~

get_dropmodel <=就用之前那個啦

> 在batch level 下，不是必然做augmentation
> 又用時間，又會濛

用AdamW mixed precision來train
攪leraning rate finder
trained it for 25epochs
59.3%

## 怎樣可以做得更好呢?

看論文，如何從60%>70%

用real ResNet ~~([我看了這條影片](https://youtu.be/o_3mboe1jYI?si=Oq3WPSWIOQRcHv7a))~~

1. no extra parameter 
2. 1x1 convolution (x2)
3. /2 size (=>32x32)but stride of 2



5 down sample!
3+2+2+1+1=9 res blocks

前

後
多過一倍
61.8%了

## More augmentation: trivial augment
[論文](https://arxiv.org/pdf/2103.10158)

在這個dataset 表現沒有太好
所以逐件逐件來 (/張?!/項?!)
(要做augmentation, 之後轉去tensor,之後normalize )

conv -> normalization->activation(optional)

64.9%

## 25 Super Resolution not Classification

independent variable will be scaled down to 32x32 pixel
dependent variable: original image

do random crop within padded image and random flips: need exactly the same cropping and flipping (both independent and dependent)

難版super resolution 可以把deleted 的pixels 換成更好的pixels

重點: dataset 要求簡單點，不用load labels
train 和validation 是沒分別~

TfmDS only appkued to the independent var

64x64->32x32 (2x2pixels by 2)->64
如果質數好32>64
更甚->128 etc.

## Denoising autoencoder (Review: ch8)


[Unet](https://arxiv.org/abs/1505.04597)!
2015開發的, 起初用作醫療影像
32x32 pixels in the lowest resolution
(-2魔法)

-> down sampling
^ up sampling

Question: Moudle list?! does not do anything?!

## Unet


8.6% 損失又真的是少了很多

intput: 

output: 

肉眼沒分別 ~~我眼睛業障~~


這個看得出來! 但這是t (target )

## Perceptual loss

拯救眼睛計劃~
傳統: pixel loss: 對比像數:模糊不清 (比較安全)
Perceptual loss: 印象/概念/大概的草稿

要求: classifier model
~~我眼鏡是有病吧?~~
scalling factor: 0.1x


trading

fast ai favourite trick:
gradually unfreezing pre trained networks

use actual weight for training and sampling

turn off require grad

trick: use random weight, set everything to unfrozen. do 20 epochs

==============================
然而，我還是沒成功~
不等了~