Pytorch tutorial

class: center, middle, inverse, title-slide

.title[
# Pytorch tutorial
]
.subtitle[
## My research
]
.author[
### Minsu Kim
]
.date[
### 2026.01.13
]

---

.contents-list {
  font-size: 30px;
  font-family: 'Trebuchet MS', sans-serif;
  line-height: 1.5;
}

.main-text {
  font-size: 30px;
  font-family: 'Trebuchet MS', sans-serif;
  line-height: 1.5;
}

/* 메인 bullet 크기 */
.remark-slide-content ul {
  font-size: 20px;
}

/* subbullet 크기 (더 작게) */
.remark-slide-content ul ul {
  font-size: 18px;
}

/* sub-subbullet이 있다면 더 작게 */
.remark-slide-content ul ul ul {
  font-size: 15px;
}

.remark-slide-number {
  font-size: 16px;
  bottom: 40px;
  right: 10px;
}

.remark-slide-content:not(.title-slide)::before {
  content: "";
  position: absolute;
  bottom: 8px;
  right: 10px;
  width: 80px;
  height: 30px;
  background: url('lab_logo.jpg') no-repeat center;
  background-size: contain;
}
</style>

# Contents

---
# Introduction

### SGD Optimization

First, we create tensors with `requires_grad=True` to track gradients.
`$$\theta = \{x, y\}$$`
```python
x = torch.tensor(0.0, requires_grad=True)
y = torch.tensor(0.0, requires_grad=True)
```

Next, we initialize the optimizer (SGD) with parameters and learning rate.
`$$\theta_{t+1} = \theta_t - \eta \nabla \mathcal{L}$$`
```python
optimizer = torch.optim.SGD([x, y], lr=0.1)
```

In the training loop, we first reset the gradients.
`$$\nabla \mathcal{L} \leftarrow 0$$`
```python
optimizer.zero_grad()
```

We compute the loss with noise and peform backpropagation.
`$$\mathcal{L} = f(x, y, \epsilon)$$`
```python
loss = loss_function(x, y, torch.randn(1) * 0.1)
loss.backward()
```

Finally, we update the parameters using the step function.
```python
optimizer.step() #<<
```