【Python】GPTブームに乗じて簡易的な機械学習AIを完全自作してみた

突然なのですが、実は私は(到底信じてもらえないでしょうが)ChatGPTが流行る前から機械学習をやりたい思いはあったのですが、当時はプログラミングの知識がほとんどなく、結局Webアプリケーションに興味移ってしまったためしばらくやってませんでした。。。

ただ最近GPT絡みでAIブームが来ているので(?)簡易的な機械学習と言語モデル生成をやってみます。

趣味の範疇なので悪しからず。

コード

1.あらかじめ指定した文字列を整理(?)するやつ

import random

# function to generate n-grams from text
def generate_ngrams(text, n):
    ngrams = []
    words = text.split()
    for i in range(len(words)-n+1):
        ngrams.append(' '.join(words[i:i+n]))
    return ngrams

# function to generate a language model from text
def generate_language_model(text, n):
    ngrams = generate_ngrams(text, n)
    model = {}
    for i in range(len(ngrams)-1):
        if ngrams[i] not in model:
            model[ngrams[i]] = []
        model[ngrams[i]].append(ngrams[i+1])
    return model

# function to generate text using language model
def generate_text(model, n, length):
    current_ngram = random.choice(list(model.keys()))
    text = current_ngram
    for i in range(length):
        if current_ngram not in model:
            break
        possible_words = model[current_ngram]
        next_word = random.choice(possible_words)
        text += ' ' + next_word.split()[n-1]
        current_ngram = ' '.join(text.split()[-n:])
    return text

# example usage
text = 'The quick brown fox jumps over the lazy dog'
n = 2
model = generate_language_model(text, n)
generated_text = generate_text(model, n, 10)
print(generated_text)

2.標準入力をもとにJSONでモデル生成するやつ

import random
import json
# function to generate n-grams from text
def generate_ngrams(text, n):
    ngrams = []
    words = text.split()
    for i in range(len(words)-n+1):
        ngrams.append(' '.join(words[i:i+n]))
    return ngrams

# function to generate a language model from text
def generate_language_model(text, n):
    ngrams = generate_ngrams(text, n)
    model = {}
    for i in range(len(ngrams)-1):
        if ngrams[i] not in model:
            model[ngrams[i]] = []
        model[ngrams[i]].append(ngrams[i+1])
    return model

# function to generate text using language model
def generate_text(model, n, length):
    current_ngram = random.choice(list(model.keys()))
    text = current_ngram
    for i in range(length):
        if current_ngram not in model:
            break
        possible_words = model[current_ngram]
        next_word = random.choice(possible_words)
        text += ' ' + next_word.split()[n-1]
        current_ngram = ' '.join(text.split()[-n:])
    return text

# example usage
text1 = str(input())
text2 = str(input())
text = text1 + text2
n = 2
model = generate_language_model(text, n)
generated_text = generate_text(model, n, 10)
print(generated_text)

with open('model.json', 'w') as f:
    json.dump(model, f)

生成したモデル

モデル1

{
    "The quick": [
        "quick brown",
        "quick brown"
    ],
    "quick brown": [
        "brown fox",
        "brown fox",
        "brown fox",
        "brown fox"
    ],
    "brown fox": [
        "fox jumps",
        "fox jumps",
        "fox jumps",
        "fox jumps"
    ],
    "fox jumps": [
        "jumps over",
        "jumps over",
        "jumps over",
        "jumps over"
    ],
    "jumps over": [
        "over the",
        "over the",
        "over the",
        "over the"
    ],
    "over the": [
        "the lazy",
        "the lazy",
        "the lazy",
        "the lazy"
    ],
    "the lazy": [
        "lazy dog",
        "lazy dogThe",
        "lazy dogThe",
        "lazy dog"
    ],
    "lazy dog": [
        "dog The"
    ],
    "dog The": [
        "The quick"
    ],
    "lazy dogThe": [
        "dogThe quick",
        "dogThe quick"
    ],
    "dogThe quick": [
        "quick brown",
        "quick brown"
    ]
}

モデル2

{
    "1 2": [
        "2 3",
        "2 3"
    ],
    "2 3": [
        "3 4",
        "3 4",
        "3 4",
        "3 4"
    ],
    "3 4": [
        "4 5",
        "4 51",
        "4 51",
        "4 5"
    ],
    "4 5": [
        "5 1"
    ],
    "5 1": [
        "1 2"
    ],
    "4 51": [
        "51 2",
        "51 2"
    ],
    "51 2": [
        "2 3",
        "2 3"
    ]
}

モデル3（標準入力から生成するタイプ）

{
    "22233 3332231313131": [
        "3332231313131 daiwduai"
    ]
}

ノリと勢いでやりましたが案外いけますね

ﾀﾉｼｲ