from transformers import pipeline            # 🌐
clf = pipeline('text-classification')        # 不指定模型就用預設
print(clf('This course is amazing!'))        # → [{'label': 'POSITIVE', 'score': 0.9998...}]

def iou(a, b):                               # 💻 框的格式 (x1, y1, x2, y2)
    x1, y1 = max(a[0], b[0]), max(a[1], b[1])
    x2, y2 = min(a[2], b[2]), min(a[3], b[3])
    inter = max(0, x2 - x1) * max(0, y2 - y1)
    area = lambda r: (r[2] - r[0]) * (r[3] - r[1])
    return inter / (area(a) + area(b) - inter)

print(iou((0, 0, 10, 10), (5, 5, 15, 15)))   # → 0.14285714285714285（重疊 25 ÷ 聯集 175）
print(iou((0, 0, 10, 10), (0, 0, 10, 10)))   # → 1.0

from transformers import pipeline            # 🌐 物件偵測
from PIL import Image
import requests

url = 'http://images.cocodataset.org/val2017/000000039769.jpg'   # 兩隻貓和遙控器
img = Image.open(requests.get(url, stream=True).raw)
det = pipeline('object-detection', model='facebook/detr-resnet-50')
for r in det(img):
    print(r['label'], round(r['score'], 2), r['box'])

from ultralytics import YOLO                 # 🌐 pip install ultralytics
model = YOLO('yolov8n.pt')                   # n＝最小最快
model.train(data='coco8.yaml', epochs=3)     # 換成自己的資料設定檔
results = model('bus.jpg')                   # 預測

import torch, torch.nn as nn                  # 💻 迷你 GAN：學會產生「平均 4、標準差 1.25」的數字
torch.manual_seed(0)
G = nn.Sequential(nn.Linear(1, 32), nn.ReLU(), nn.Linear(32, 1))                 # 生成器
D = nn.Sequential(nn.Linear(1, 32), nn.ReLU(), nn.Linear(32, 1), nn.Sigmoid())   # 判別器
optG = torch.optim.Adam(G.parameters(), lr=1e-3)
optD = torch.optim.Adam(D.parameters(), lr=1e-3)
bce = nn.BCELoss()
ones, zeros = torch.ones(64, 1), torch.zeros(64, 1)

for step in range(6000):
    real = torch.randn(64, 1) * 1.25 + 4      # 真資料
    fake = G(torch.randn(64, 1))              # 假資料
    # 1. 訓練判別器：真的說 1，假的說 0
    lossD = bce(D(real), ones) + bce(D(fake.detach()), zeros)
    optD.zero_grad(); lossD.backward(); optD.step()
    # 2. 訓練生成器：讓判別器把假的說成 1
    lossG = bce(D(fake), ones)
    optG.zero_grad(); lossG.backward(); optG.step()

sample = G(torch.randn(1000, 1))
print(round(sample.mean().item(), 1), round(sample.std().item(), 1))   # → 3.9 1.1（目標 4 和 1.25）
# 試試把 6000 改成 3000，或 lr 改成 2e-3：結果會差很多。這就是「GAN 訓練不穩定」

from transformers import pipeline            # 🌐 BERT 填空：看它懂不懂上下文
fill = pipeline('fill-mask', model='bert-base-uncased')
for r in fill('The capital of France is [MASK].')[:3]:
    print(r['token_str'], round(r['score'], 3))           # → paris 排第一

from transformers import pipeline            # 🌐 編碼器-解碼器：英翻中
tr = pipeline('translation', model='Helsinki-NLP/opus-mt-en-zh')
print(tr('Deep learning is fun.')[0]['translation_text'])

from transformers import pipeline            # 🌐 Whisper 語音轉文字
asr = pipeline('automatic-speech-recognition', model='openai/whisper-small')
print(asr('my_voice.wav')['text'])           # 換成自己的錄音檔

import torch                                  # 🌐 HuBERT 當特徵抽取器
from transformers import AutoFeatureExtractor, HubertModel
fe = AutoFeatureExtractor.from_pretrained('facebook/hubert-base-ls960')
model = HubertModel.from_pretrained('facebook/hubert-base-ls960')
wave = torch.randn(16000)                    # 假裝是 1 秒的聲音
inputs = fe(wave.numpy(), sampling_rate=16000, return_tensors='pt')
feat = model(**inputs).last_hidden_state     # 形狀 (1, 49, 768)：49 個時間點 × 768 維
print(feat.mean(dim=1).shape)                # 平均成一個 768 維向量，接分類器 → torch.Size([1, 768])

import torch, torch.nn as nn                  # 💻 手寫 LoRA 層
class LoRALinear(nn.Module):
    def __init__(self, base, r=4):
        super().__init__()
        self.base = base
        for p in self.base.parameters(): p.requires_grad = False   # 凍結原本的權重
        self.A = nn.Parameter(torch.randn(r, base.in_features) * 0.01)
        self.B = nn.Parameter(torch.zeros(base.out_features, r))   # B 從 0 開始 → 一開始不改變模型
    def forward(self, x):
        return self.base(x) + x @ self.A.T @ self.B.T

layer = LoRALinear(nn.Linear(768, 768))
train = sum(p.numel() for p in layer.parameters() if p.requires_grad)
total = sum(p.numel() for p in layer.parameters())
print(train, total, f'{train / total:.1%}')  # → 6144 596736 1.0%

from peft import LoraConfig, get_peft_model   # 🌐 pip install peft
from transformers import AutoModelForSequenceClassification
model = AutoModelForSequenceClassification.from_pretrained('bert-base-uncased', num_labels=2)
model = get_peft_model(model, LoraConfig(r=8, target_modules=['query', 'value'], task_type='SEQ_CLS'))
model.print_trainable_parameters()           # 只有約 0.3% 要訓練