candywal

I implemented a transformer from scratch, with different sampling methods, beam search and caching. Made it while working through https://www.arena.education/.

candywal

# %% This file should
import torch as t
from torch import nn
import torch.utils.data as data
from torchvision import datasets, transforms
from dataclasses import dataclass
from typing import Literal
from tqdm import tqdm
import matplotlib.pyplot as plt
from IPython import display

candywal

#lang typed/racket
(require typed/rackunit)

; Data definitions
(define-type ExprC (U NumC IdC AppC IfC LamTC StringC))
(struct NumC ([n : Real])#:transparent)
(struct IdC ([s : Symbol])#:transparent)  
(struct AppC ([fun : ExprC] [arg : (Listof ExprC)]) #:transparent)
(struct IfC ([test : ExprC] [then : ExprC] [else : ExprC]) #:transparent)
(struct LamTC ([args : (Listof Symbol)] [arg-types : (Listof ty)] [body : ExprC]) #:transparent)