---
tags: Python
---
# 使用YOLO辨識K線圖
* [安裝GIT](https://git-scm.com/)
* [安裝Pytorch](https://pytorch.org/get-started/locally/)
* [安裝CUDA(選擇性)](https://developer.nvidia.com/zh-cn/cuda-downloads?target_os=Windows&target_arch=x86_64&target_version=10&target_type=exelocal)
* pip install [mplfinance](https://github.com/matplotlib/mplfinance)
* `git clone https://github.com/ultralytics/yolov5`
* `cd yolov5`
* `pip install -r requirements.txt`
* 新增`yolov5/train.yaml`

train: train/
val: train/
# number of classes
nc: 2
# class names
names: ['sell','buy']

## 自動產生資料集
### 引入套件

import os
import sqlite3
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
import matplotlib.patches as patches
import matplotlib.dates as mpl_dates
from mplfinance.original_flavor import candlestick_ohlc

### 建立資料夾

os.makedirs('yolov5/train/images', exist_ok=True)
os.makedirs('yolov5/train/labels', exist_ok=True)

### 讀取資料庫
取得資料庫請參考[分析系統建置/步驟一](https://hackmd.io/dBWiFw-qQIiwGqx5QGrVyQ?view)

symbol = 2330
with sqlite3.connect('flaskr.db') as con:
    sql = f'SELECT Date,Open,High,Low,Close FROM TW{symbol};'
    df = pd.read_sql(sql, con)
df['Date'] = pd.to_datetime(df['Date']).apply(mpl_dates.date2num)

### 技術指標與交叉點

ROLL = 1 # 框框半徑
# 技術指標
DATE,O,H,L,C = df['Date'],df['Open'],df['High'],df['Low'],df['Close']
FAST = C.rolling(12).mean().values
SLOW = C.rolling(26).mean().values
# 交叉點
rects = []
pos = FAST[ROLL]>SLOW[ROLL]
for i in range(ROLL,len(SLOW)):
    curPos = FAST[i]>SLOW[i]
    if curPos != pos:
        arange = range(i-ROLL,i+ROLL)
        x = DATE[i-ROLL]
        y = min(*FAST[arange],*SLOW[arange])
        w = 2*ROLL
        h = max(*FAST[arange],*SLOW[arange])-y
        label = 1 if curPos else 0
        rects.append((i,label,x,y,w,h))
        pos = curPos

### 繪圖

STRIDE = 30 # 時窗
DOHLC = df[['Date', 'Open', 'High', 'Low', 'Close']].values
for j in range(0,len(df),STRIDE)[1:]:
    range_rects = [r for r in rects if j<r[0] and r[0]<j+STRIDE]
    # 跳過沒有標記的
    if not range_rects:
        continue
    # 畫蠟燭線、技術折線，圖片尺寸=fig.get_size_inches()*fig.dpi
    fig, ax = plt.subplots()
    candlestick_ohlc(ax, DOHLC[j:j+STRIDE,:], width=0.6, colorup='green', colordown='red', alpha=1)
    ax.plot(DATE[j:j+STRIDE],FAST[j:j+STRIDE])
    ax.plot(DATE[j:j+STRIDE],SLOW[j:j+STRIDE])
    # 畫方框
    for i,label,x,y,w,h in range_rects:
        patch = patches.Rectangle((x,y), w, h, linewidth=1, edgecolor='r', facecolor='none')
        ax.add_patch(patch)
    # 標題刻度
    ax.xaxis.set_major_formatter(mpl_dates.DateFormatter('%Y-%m-%d'))
    fig.autofmt_xdate()
    plt.axis('off')
    fig.tight_layout()
    # 儲存圖片、標籤
    date_str = mpl_dates.num2date(x).strftime('%Y-%m-%SLOW')
    plt.savefig(f'yolov5/train/images/{symbol}-{date_str}.jpg')

### YOLO圖片格式

# 將左下至右上的數值，轉換成由左上至右下的圖片的pixel絕對座標，同時正規化成[0,1]
def yolo_formatter(x,y,w,h,fig,ax):
	width, height = fig.canvas.get_width_height()
	left, bottom = ax.transData.transform([x,y])
	right, top = ax.transData.transform([x+w,y+h])
	w = round((right-left)/width,2)
	h = round((top-bottom)/height,2)
	cx = round((right+left)/2/width,2)
	cy = round(1-(top+bottom)/2/height,2)
	return cx,cy,w,h

### 產生標籤檔

# 寫入標籤文字檔
with open(f'yolov5/train/labels/{symbol}-{date_str}.txt','w') as f:
    text = ''
    for i,label,x,y,w,h in range_rects:
        cx,cy,w,h = yolo_formatter(x,y,w,h,fig,ax)
        text += f'{label} {cx} {cy} {w} {h}\n'
    f.write(text)
# 最後記得清空繪圖記憶體
plt.close()

檔案`2330-2018-10-01.txt`內容：

1 0.24 0.27 0.04 0.03
0 0.36 0.33 0.04 0.06
1 0.84 0.5 0.04 0.01

## 訓練、偵測
### 模型選擇
### 超參數設定

lr0: 0.01  # initial learning rate (SGD=1E-2, Adam=1E-3)
lrf: 0.2  # final OneCycleLR learning rate (lr0 * lrf)
momentum: 0.937  # SGD momentum/Adam beta1
weight_decay: 0.0005  # optimizer weight decay 5e-4
warmup_epochs: 3.0  # warmup epochs (fractions ok)
warmup_momentum: 0.8  # warmup initial momentum
warmup_bias_lr: 0.1  # warmup initial bias lr
box: 0.05  # box loss gain
cls: 0.5  # cls loss gain
cls_pw: 1.0  # cls BCELoss positive_weight
obj: 1.0  # obj loss gain (scale with pixels)
obj_pw: 1.0  # obj BCELoss positive_weight
iou_t: 0.20  # IoU training threshold
anchor_t: 4.0  # anchor-multiple threshold
# anchors: 3  # anchors per output layer (0 to ignore)
fl_gamma: 0.0  # focal loss gamma (efficientDet default gamma=1.5)
hsv_h: 0.015  # image HSV-Hue augmentation (fraction)
hsv_s: 0.7  # image HSV-Saturation augmentation (fraction)
hsv_v: 0.4  # image HSV-Value augmentation (fraction)
degrees: 0.0  # image rotation (+/- deg)
translate: 0.1  # image translation (+/- fraction)
scale: 0.5  # image scale (+/- gain)
shear: 0.0  # image shear (+/- deg)
perspective: 0.0  # image perspective (+/- fraction), range 0-0.001
flipud: 0.0  # image flip up-down (probability)
fliplr: 0.5  # image flip left-right (probability)
mosaic: 1.0  # image mosaic (probability)
mixup: 0.0  # image mixup (probability)

### 訓練
* **資料數量**：每個類別≥1.5k圖像
* **每類實例**：每類總計≥10,000個實例（帶標籤的對象）
* **圖像種類**：必須代表所部署的環境。對於現實世界中的例，建議使用一天中不同時間，不同季節，不同天氣，不同照明，不同角度，不同來源（網路圖片、本地收集、不同相機）等圖像。
* **標籤一致性**：所有圖像中所有類的所有實例都必須標記，部分標籤將不起作用。
* **標籤精度**：標籤必須緊密包圍每個對象，對象與其邊界框之間不應存在任何空間，任何物體都不應缺少標籤。
* **背景圖像**：背景圖像是沒有對象的圖像，這些圖像被添加到數據集中以減少誤報（FP），建議使用大約0-10％的背景圖片來幫助減少FP（COCO提供1000張背景圖片作為參考，佔總數的1％）。
* **Epochs**：起手300，若未過度擬和，可訓練至600,1200以上
* **Image size**：預設`--img 640`
* **Batch size**：應避免使用較小batch，否則會產生不準確的統計數據
* 使用[Google Cloud訓練](https://github.com/ultralytics/yolov5/wiki/GCP-Quickstart)
* 使用[AWS訓練](https://github.com/ultralytics/yolov5/wiki/AWS-Quickstart)

官方指令
`python train.py --img 640 --batch 16 --epochs 5 --data coco128.yaml --weights yolov5s.pt`
個人指令(workers核心、evolve視覺化超參數調校)
`python train.py --batch 16 --epochs 5 --data train.yaml --workers 0 --evolve`
繪製訓練結果

import sys
sys.path.append('.')
from yolov5.utils.plots import plot_results
plot_results(save_dir='yolov5/runs/train/exp')

* `hyp.yaml`
* `opt.yaml`
* `results.png`
* `test_batch*.jpg`
### 其他偵測指令
使用攝像頭(0代表webcam)
`python detect.py --source 0 --weights yolov5s.pt --conf 0.25`
使用glob批次偵測
`python detect.py --source path/*.jpg`