pytorch memo

#1. torch.nn really?

• https://pytorch.org/tutorials/beginner/nn_tutorial.html
• np -> torch
• TensorDataset
• DataLoader batch 로 튀어나옴
• WrappedDataLoader 원래 있던거에서
  • __init__, __len__, __iter__ 만져서 수정
• nn.Sequential
• class Lambda(nn.Module):
  • __init__ , forward
• AdaptiveAvgPool2d -> global pool 동일

#2. andrej karpathy nn zero to hero

1. https://youtu.be/PaCmpygFfXo :makemore
   broadcasting 주의 ,
   P \= P.sum(1, keepdims=True) in place 로 해야 더 빠르다.

#2-1

conv understanding

• https://github.com/vdumoulin/conv_arithmetic/blob/master/README.md

---

ToDos

• https://pytorch.org/tutorials/beginner/nlp/deep_learning_tutorial.html
• U-net pytorch impl : https://amaarora.github.io/2020/09/13/unet.html
  • nn.ModuleList 그냥 list랑 뭐가 다를까?

---

dataloader > collate_fn

def collate_fn(batch):
	batch # type: list 임.
    batch를 풀어서 하거나, list 를 바꿔서 해야함.
    ```
- `list(zip(*batch))` --> trainspose임

- 한방에 batch를 처리하려면.. lambda 를 쓰는게 좋을 수도