A Simple Method for Training GPT-2 To Generate Haiku Using the NanoGPT Repository

import csv
import re

with open('all_haiku.csv', 'r') as csv_file:
    reader = csv.reader(csv_file)
    i = 0
    txt_file = open('input.txt', 'w')
    for row in reader:
        strings = row[1:4]
        upd_strings = []
        for str in strings:
            if len(str) > 0:
                if str[0] == ' ':
                    while str[0] == ' ':
                        str = str[1:]
                if str[-1] == ' ':
                    while str[-1] == ' ':
                        str = str[:-1] 
                upd_strings.append(re.sub(r"[^a-zA-Z ]+", "", str)) 
            else:
                upd_strings.append('')
        # skip text and label 
        if i > 0: 
            txt_file.write(upd_strings[0] + ';' + upd_strings[1] + ';' + upd_strings[2] +'\n')
        i+=1

    print("Added", i-1, "strings of text")

import time

out_dir = 'out-haiku'
eval_interval = 5
eval_iters = 40
wandb_log = True # feel free to turn on
wandb_project = 'haiku'
wandb_run_name = 'test3'

dataset = 'haiku'
init_from = 'gpt2-large' # this is the largest GPT-2 model

# only save checkpoints if the validation loss improves
always_save_checkpoint = False

# the number of examples per iter:
# 1 batch_size * 32 grad_accum * 1024 tokens = 32,768 tokens/iter
# shakespeare has 301,966 tokens, so 1 epoch ~= 9.2 iters
# haiku has 2,301,966 tokens, so 1 epoch ~= 70 iters * 10 number of iterations
batch_size = 1
gradient_accumulation_steps = 32
max_iters = 1000

# finetune at constant LR
learning_rate = 1e-6
decay_lr = True
warmup_iters = 200#max_iters/10
lr_decay_iters = max_iters 
min_lr = learning_rate/10

compile=False

"""
Sample from a trained model
"""
import os
import pickle
from contextlib import nullcontext
import torch
import tiktoken
from model import GPTConfig, GPT

promt = "Full Moon is shining\n"

# -----------------------------------------------------------------------------
init_from = 'resume' # either 'resume' (from an out_dir) or a gpt2 variant (e.g. 'gpt2-xl')
out_dir = 'out-haiku_1k' # ignored if init_from is not 'resume'
start = "\n" # or "<|endoftext|>" or etc. Can also specify a file, use as: "FILE:prompt.txt"
num_samples = 10 # number of samples to draw
max_new_tokens = 20 # number of tokens generated in each sample
temperature = 0.9 # 1.0 = no change, < 1.0 = less random, > 1.0 = more random, in predictions
top_k = 200 # retain only the top_k most likely tokens, clamp others to have 0 probability
seed = 1381
device = 'cuda' # examples: 'cpu', 'cuda', 'cuda:0', 'cuda:1', etc.
dtype = 'bfloat16' # 'float32' or 'bfloat16' or 'float16'
compile = False # use PyTorch 2.0 to compile the model to be faster
exec(open('configurator.py').read()) # overrides from command line or config file
# -----------------------------------------------------------------------------

torch.manual_seed(seed)
torch.cuda.manual_seed(seed)
torch.backends.cuda.matmul.allow_tf32 = True # allow tf32 on matmul
torch.backends.cudnn.allow_tf32 = True # allow tf32 on cudnn
device_type = 'cuda' if 'cuda' in device else 'cpu' # for later use in torch.autocast
ptdtype = {'float32': torch.float32, 'bfloat16': torch.bfloat16, 'float16': torch.float16}[dtype]
ctx = nullcontext() if device_type == 'cpu' else torch.amp.autocast(device_type=device_type, dtype=ptdtype)

# model
if init_from == 'resume':
    # init from a model saved in a specific directory
    ckpt_path = os.path.join(out_dir, 'ckpt.pt')
    checkpoint = torch.load(ckpt_path, map_location=device)
    gptconf = GPTConfig(**checkpoint['model_args'])
    model = GPT(gptconf)
    state_dict = checkpoint['model']
    unwanted_prefix = '_orig_mod.'
    for k,v in list(state_dict.items()):
        if k.startswith(unwanted_prefix):
            state_dict[k[len(unwanted_prefix):]] = state_dict.pop(k)
    model.load_state_dict(state_dict)
elif init_from.startswith('gpt2'):
    # init from a given GPT-2 model
    model = GPT.from_pretrained(init_from, dict(dropout=0.0))

model.eval()
model.to(device)
if compile:
    model = torch.compile(model) # requires PyTorch 2.0 (optional)

# look for the meta pickle in case it is available in the dataset folder
load_meta = False
if init_from == 'resume' and 'config' in checkpoint and 'dataset' in checkpoint['config']: # older checkpoints might not have these...
    meta_path = os.path.join('data', checkpoint['config']['dataset'], 'meta.pkl')
    load_meta = os.path.exists(meta_path)
if load_meta:
    print(f"Loading meta from {meta_path}...")
    with open(meta_path, 'rb') as f:
        meta = pickle.load(f)
    # TODO want to make this more general to arbitrary encoder/decoder schemes
    stoi, itos = meta['stoi'], meta['itos']
    encode = lambda s: [stoi[c] for c in s]
    decode = lambda l: ''.join([itos[i] for i in l])
else:
    # ok let's assume gpt-2 encodings by default
    print("No meta.pkl found, assuming GPT-2 encodings...")
    enc = tiktoken.get_encoding("gpt2")
    encode = lambda s: enc.encode(s, allowed_special={"<|endoftext|>"})
    decode = lambda l: enc.decode(l)

# encode the beginning of the prompt
#if start.startswith('FILE:'):
#    with open(start[5:], 'r', encoding='utf-8') as f:
#        start = f.read()
start_ids = encode(promt)
x = (torch.tensor(start_ids, dtype=torch.long, device=device)[None, ...])

print('---------------')
# run generation
with torch.no_grad():
    with ctx:
        for k in range(num_samples):
            y = model.generate(x, max_new_tokens, temperature=temperature, top_k=top_k)
            print(decode(y[0].tolist()))
            print('---------------')