Introduction to PyTorch
Simple Perceptron และการฝึกโมเดล

import torch

model = torch.nn.Sequential(
    torch.nn.Linear(2, 1),   # z = w·x + b
    torch.nn.Sigmoid(),      # ŷ = σ(z)
)

# ลอง peek ค่าเริ่มต้น:
print(model[0].weight)   # tensor([[...]])
print(model[0].bias)     # tensor([...])
y_hat = model(torch.randn(5, 2))  # ทาย 5 จุด

import torch

# 1) requires_grad=True → ให้ PyTorch ติดตาม
w = torch.tensor(1.0, requires_grad=True)

# 2) Forward pass — PyTorch สร้าง computational graph
x = torch.tensor(2.0)
z = w * x          # z = 2w
loss = (z - 3)**2  # loss = (2w - 3)²

# 3) Backward pass — คำนวณ dloss/dw อัตโนมัติ
loss.backward()
print(w.grad)      # tensor(4.)  ← dloss/dw = 4(2w-3)

# 4) Update — optimizer ใช้ gradient ปรับ w
opt = torch.optim.SGD([w], lr=0.1)
opt.step()          # w = 1.0 - 0.1*4.0 = 0.6

def plot_boundary(model):
    w = model[0].weight.data.ravel()
    b = model[0].bias.data.item()
    xs = torch.linspace(-3, 3, 2)
    ys = -(w[0] * xs + b) / w[1]
    plt.plot(xs.numpy(), ys.numpy(),
             "k-", lw=2)

plot_boundary(model)   # ลองก่อนเทรน!
plt.show()

# views.py
import threading, json, torch
from django.http import JsonResponse

_state = dict(epoch=0, loss=0, acc=0,
              w=[0.0, 0.0], b=0.0, running=False)

def _train():
    torch.manual_seed(42)
    X = torch.randn(200, 2)
    y = ((X[:,0]*1.5+X[:,1]-0.5)>0
         ).float().unsqueeze(1)
    model = torch.nn.Sequential(
        torch.nn.Linear(2,1), torch.nn.Sigmoid())
    loss_fn = torch.nn.BCELoss()
    opt = torch.optim.SGD(model.parameters(), lr=0.1)
    _state['running'] = True
    for ep in range(200):
        y_hat = model(X)
        loss = loss_fn(y_hat, y)
        opt.zero_grad(); loss.backward(); opt.step()
        _state.update(
            epoch=ep,
            loss=round(loss.item(),4),
            acc=round(((y_hat>0.5).float()==y
                ).float().mean().item(),4),
            w=[round(x,4) for x in
                model[0].weight.data.tolist()[0]],
            b=round(model[0].bias.data.item(),4))
    _state['running'] = False

def train_start(request):
    threading.Thread(target=_train, daemon=True).start()
    return JsonResponse({'ok': True})

def train_state(request):
    return JsonResponse(_state)

<canvas id="cv" width="400" height="400"></canvas>
<div id="info"></div>
<script>
const cv = document.getElementById('cv');
const ctx = cv.getContext('2d');
const CX = 200, S = 60;

const pts = [];
function randn(){let u=0,v=0;while(!u)u=Math.random();
  v=Math.random();return Math.sqrt(-2*Math.log(u))*Math.cos(6.283*v)}
for(let i=0;i<200;i++){
  const x=randn(),y=randn();
  pts.push({x,y,l:(x*1.5+y-0.5)>0?1:0})}

function draw(s){
  ctx.fillStyle='#1a1a2e'; ctx.fillRect(0,0,400,400);
  for(const p of pts){
    ctx.beginPath();
    ctx.arc(CX+p.x*S, CX-p.y*S, 3, 0, 6.283);
    ctx.fillStyle=p.l?'#4ade80':'#f87171'; ctx.fill();}
  if(s.w && s.w[1]!==0){
    ctx.beginPath();
    ctx.moveTo(0,  CX-(-(s.w[0]*(-3)+s.b)/s.w[1])*S);
    ctx.lineTo(400,CX-(-(s.w[0]*( 3)+s.b)/s.w[1])*S);
    ctx.strokeStyle='#fbbf24'; ctx.lineWidth=2; ctx.stroke();}
  document.getElementById('info').textContent=
    `epoch ${s.epoch} | loss ${s.loss} | acc ${(s.acc*100).toFixed(0)}%`;}

// polling loop — ยิง request ซ้ำทุก 50ms
async function poll(){
  const r = await fetch('/api/state/');
  const s = await r.json();
  draw(s);
  if(s.running) setTimeout(poll, 50);}

fetch('/api/train/').then(()=>poll());
</script>

# views.py
import json, torch
from django.http import StreamingHttpResponse

def train_stream(request):
    def generate():
        torch.manual_seed(42)
        X = torch.randn(200, 2)
        y = ((X[:,0]*1.5+X[:,1]-0.5)>0
             ).float().unsqueeze(1)
        model = torch.nn.Sequential(
            torch.nn.Linear(2,1), torch.nn.Sigmoid())
        loss_fn = torch.nn.BCELoss()
        opt = torch.optim.SGD(model.parameters(), lr=0.1)
        for ep in range(200):
            y_hat = model(X)
            loss = loss_fn(y_hat, y)
            opt.zero_grad(); loss.backward(); opt.step()
            yield "data: " + json.dumps(dict(
                epoch=ep,
                loss=round(loss.item(),4),
                acc=round(((y_hat>0.5).float()==y
                    ).float().mean().item(),4),
                w=[round(x,4) for x in
                    model[0].weight.data.tolist()[0]],
                b=round(model[0].bias.data.item(),4)
            )) + "\n\n"
    return StreamingHttpResponse(
        generate(),
        content_type='text/event-stream')

<canvas id="cv" width="400" height="400"></canvas>
<div id="info"></div>
<script>
const cv = document.getElementById('cv');
const ctx = cv.getContext('2d');
const CX = 200, S = 60;

// ข้อมูล (seed เดียวกับ backend)
const pts = [];
function randn(){let u=0,v=0;while(!u)u=Math.random();
  v=Math.random();return Math.sqrt(-2*Math.log(u))*Math.cos(6.283*v)}
for(let i=0;i<200;i++){
  const x=randn(),y=randn();
  pts.push({x,y,l:(x*1.5+y-0.5)>0?1:0})}

function draw(s){
  ctx.fillStyle='#1a1a2e'; ctx.fillRect(0,0,400,400);
  for(const p of pts){
    ctx.beginPath();
    ctx.arc(CX+p.x*S, CX-p.y*S, 3, 0, 6.283);
    ctx.fillStyle=p.l?'#4ade80':'#f87171'; ctx.fill();}
  if(s.w && s.w[1]!==0){
    ctx.beginPath();
    ctx.moveTo(0,  CX-(-(s.w[0]*(-3)+s.b)/s.w[1])*S);
    ctx.lineTo(400,CX-(-(s.w[0]*( 3)+s.b)/s.w[1])*S);
    ctx.strokeStyle='#fbbf24'; ctx.lineWidth=2; ctx.stroke();}
  document.getElementById('info').textContent=
    `epoch ${s.epoch} | loss ${s.loss} | acc ${(s.acc*100).toFixed(0)}%`;}

// EventSource = browser API สำหรับรับ SSE stream
const es = new EventSource('/api/train/');
es.onmessage = (e) => draw(JSON.parse(e.data));
es.onerror   = () => es.close();  // จบ stream → ปิด
</script>

$ uv run train.py
epoch  40 | loss 0.3940 | acc 94.50%
epoch  80 | loss 0.3074 | acc 97.50%
epoch 120 | loss 0.2647 | acc 100.00%
epoch 160 | loss 0.2379 | acc 99.50%
epoch 200 | loss 0.2190 | acc 99.00%
w = tensor([[2.1772, 1.4041]])
b = tensor([-0.7237])

def make_xor_data(n=200):
    """XOR: ช่อง 1,3 = 1, ช่อง 2,4 = 0
    → เส้นตรงเดียวแบ่งไม่ได้แน่นอน"""
    X = torch.rand(n, 2) * 4 - 2  # [-2, 2]
    y = (X[:, 0] * X[:, 1] > 0).float()
    return X, y.unsqueeze(1)

X_xor, y_xor = make_xor_data()

# scatter plot: สีเขียว=1, แดง=0
# จุดในเขตรวมสีเดียวกัน = XOR = ทแยงกัน
# → เส้นตรงเส้นเดียวไม่มีวันแบ่งถูกทั้งหมด