minimind 0.3.3

MiniMind: Lightweight and flexible AI generation library

from minimind import GPMGenerator, Sampler, SimpleTokenizer

def main(): print("MiniMind GPMGenerator 테스트 시작!")

import csv
csv_path = "MLdata.csv"  # 네 데이터셋 경로로 변경

pairs = []
with open(csv_path, encoding='utf-8') as f:
    reader = csv.DictReader(f)
    for row in reader:
        pairs.append((row['input_text'].strip(), row['output_text'].strip()))


# 샘플러 생성 (top-k 예시)
sampler = Sampler(method='top_k', k=3)
tokenizer = SimpleTokenizer()

# 생성기 초기화 시 sampler 연결
gpm = GPMGenerator(sampler=sampler, tokenizer=tokenizer)
gpm.fit(pairs[:327])

# 생성 테스트
prompt = "안녕하세요"
response = gpm.chat(prompt, max_tokens=10)

print("입력 프롬프트:", prompt)
print("생성된 텍스트:", response)

prompt = "오늘 날씨 어때?"
response = gpm.chat(prompt, max_tokens=10)

print("입력 프롬프트:", prompt)
print("생성된 텍스트:", response)

prompt = "지금 뭐해?"
response = gpm.chat(prompt, max_tokens=10)

print("입력 프롬프트:", prompt)
print("생성된 텍스트:", response)

if name == "main": main()

import autograd.numpy as anp from autograd import grad import numpy as np from minimind.neuralnet import token_embedding, positional_embedding, attention, dense, tanh, softmax

class SimpleModel: def init(self, vocab_size, seq_len, embed_dim, hidden_dim): rng = np.random.default_rng(42) self.vocab_size = vocab_size self.seq_len = seq_len self.embed_dim = embed_dim self.hidden_dim = hidden_dim

    self.params = {
        "W_embed": rng.normal(0, 0.1, (vocab_size, embed_dim)),
        "W_pos": rng.normal(0, 0.1, (seq_len, embed_dim)),
        "W_dense": rng.normal(0, 0.1, (embed_dim, hidden_dim)),
        "b_dense": np.zeros(hidden_dim),
        "W_out": rng.normal(0, 0.1, (hidden_dim, vocab_size)),
        "b_out": np.zeros(vocab_size)
    }

def forward(self, X, params):
    emb_tokens = token_embedding(X, params["W_embed"])  # (batch, seq_len, embed_dim)
    emb_positions = positional_embedding(self.seq_len, params["W_pos"])
    emb = emb_tokens + emb_positions
    Q, K, V = emb, emb, emb
    attn_out, _ = attention(Q, K, V)
    attn_out_flat = attn_out.reshape(-1, self.embed_dim)
    hidden = tanh(dense(attn_out_flat, params["W_dense"], params["b_dense"]))
    logits = dense(hidden, params["W_out"], params["b_out"])  # (batch*seq_len, vocab_size)
    logits = logits.reshape(-1, self.seq_len, self.vocab_size)
    return logits

def loss(self, params, X, Y):
    logits = self.forward(X, params)
    probs = softmax(logits, axis=2)
    batch_size = X.shape[0]
    loss = 0.0
    count = 0
    for i in range(batch_size):
        for t in range(self.seq_len):
            loss -= anp.log(probs[i, t, Y[i, t]] + 1e-12)
            count += 1
    return loss / count

def train(model, X_train, Y_train, epochs=20, batch_size=8, lr=0.01): loss_grad = grad(model.loss) N = X_train.shape[0]

for epoch in range(epochs):
    perm = np.random.permutation(N)
    total_loss = 0
    for i in range(0, N, batch_size):
        idx = perm[i:i+batch_size]
        X_batch, Y_batch = X_train[idx], Y_train[idx]
        grads = loss_grad(model.params, X_batch, Y_batch)

        # 파라미터 업데이트 (SGD)
        for k in model.params:
            model.params[k] -= lr * grads[k]

        batch_loss = model.loss(model.params, X_batch, Y_batch)
        total_loss += batch_loss * len(idx)

    avg_loss = total_loss / N
    print(f"Epoch {epoch+1}/{epochs}, Loss: {avg_loss:.4f}")

if name == "main": # 더미 데이터 생성 (단어 인덱스 시퀀스) vocab_size = 50 seq_len = 4 batch_size = 64

np.random.seed(123)
X_train = np.random.randint(0, vocab_size, (batch_size*10, seq_len))
Y_train = np.random.randint(0, vocab_size, (batch_size*10, seq_len))  # 정답도 토큰 인덱스

model = SimpleModel(vocab_size, seq_len, embed_dim=8, hidden_dim=16)
train(model, X_train, Y_train, epochs=20, batch_size=batch_size, lr=0.01)

from minimind.dl import MLPTrainer, NeuralGenerator import numpy as np

num_samples = 100

X_train = np.random.randn(num_samples, 10) y_train = np.random.randint(0, 3, size=num_samples)

model = NeuralGenerator(vocab_size=1000, embed_dim=64, hidden_layer_sizes=(128,64)) trainer = MLPTrainer(model, learning_rate=0.001)

trainer.fit(X_train, y_train, epochs=50)

from minimind.ml import SAPGenerator from minimind import SimpleTokenizer def main(): print("MiniMind SAPGenerator 테스트 시작!")

import csv
csv_path = "MLdata.csv"  # 네 데이터셋 경로로 변경

pairs = []
with open(csv_path, encoding='utf-8') as f:
    reader = csv.DictReader(f)
    for row in reader:
        pairs.append((row['input_text'].strip(), row['output_text'].strip()))


# SAPGenerator 인스턴스 생성 및 학습
tokenizer = SimpleTokenizer()
sap_gen = SAPGenerator(tokenizer=tokenizer)
sap_gen.fit(pairs[:200])

# 생성 테스트
prompt = "오늘은 뭐 할까?"
print(f"입력: {prompt}")
generated = sap_gen.chat(prompt, max_tokens=10)
print(f"생성 결과: {generated}")

if name == "main": main()

test_sampling.py

import numpy as np from minimind import top_k_sampling, top_p_sampling, temperature_sampling, Sampler

def dummy_probs(size=100): probs = np.random.rand(size) return probs / probs.sum()

def test_sampling_functions(): probs = dummy_probs()

print("top_k_sampling:", top_k_sampling(probs, k=5))
print("top_p_sampling:", top_p_sampling(probs, p=0.8))
print("temperature_sampling (temp=0.5):", temperature_sampling(probs, temperature=0.5))
print("temperature_sampling (temp=2.0):", temperature_sampling(probs, temperature=2.0))

def test_sampler_class(): probs = dummy_probs() sampler = Sampler(method='top_p', p=0.9) print("Sampler top_p:", sampler.sample(probs))

sampler.method = 'top_k'
sampler.k = 3
print("Sampler top_k:", sampler.sample(probs))

sampler.method = 'temperature'
sampler.temperature = 0.7
print("Sampler temperature:", sampler.sample(probs))

if name == "main": test_sampling_functions() test_sampler_class()

from minimind import SimpleTokenizer

tokenizer = SimpleTokenizer()

text = "Hello, 안녕하세요! Let's test the tokenizer 123." tokens = tokenizer.tokenize(text) print("토큰:", tokens)

reconstructed = tokenizer.detokenize(tokens) print("복원된 문장:", reconstructed)

import os import numpy as np from minimind import set_seed, save_json, load_json, save_model_weights, load_model_weights, simple_logger

if name == "main": # 테스트 함수들

def test_set_seed():
    set_seed(123)
    a = np.random.rand(3)
    set_seed(123)
    b = np.random.rand(3)
    assert np.allclose(a, b), "set_seed 실패!"
    print("set_seed 테스트 통과!")

def test_save_load_json():
    data = {'name': 'MiniMind', 'version': 1.0}
    filepath = 'test.json'
    save_json(data, filepath)
    loaded = load_json(filepath)
    assert data == loaded, "JSON 저장/로드 실패!"
    os.remove(filepath)
    print("save_json & load_json 테스트 통과!")

def test_save_load_weights_multi_format():
    weights = {
        'W1': np.array([1, 2, 3]),
        'b1': np.array([0.1, 0.2, 0.3])
    }
    for fmt in ['npz', 'joblib', 'json']:
        filepath = f"weights_test.{fmt}"
        save_model_weights(weights, filepath, format=fmt)
        loaded = load_model_weights(filepath, format=fmt)
        for k in weights:
            assert np.allclose(weights[k], loaded[k]), f"{fmt} {k} 가중치 저장/로드 실패!"
        os.remove(filepath)
    print("멀티 포맷 가중치 저장/로드 테스트 통과!")

def test_logger():
    simple_logger("테스트 로그 메시지")

# 실행 테스트 모음
test_set_seed()
test_save_load_json()
test_save_load_weights_multi_format()
test_logger()

from minimind.dl import CNN_MNIST, load_mnist, predict import autograd.numpy as anp

if name == "main": X_train, X_val, Y_train, Y_val = load_mnist() model = CNN_MNIST() model.fit(X_train, Y_train, X_val, Y_val, epochs=1, batch_size=128, lr=0.01)

# 샘플 디코딩
idx = 0
pred = predict(model, X_val[idx:idx+1])[0]
label = anp.argmax(Y_val[idx])
print(f"[예측 결과] 예측: {pred}, 실제: {label}")

from minimind.dl import VAE, init_params from Conv2D import load_mnist

if name == "main": X_train, X_val, _, _ = load_mnist() X_train = X_train.reshape(-1, 28, 28, 1) X_val = X_val.reshape(-1, 28, 28, 1)

params = init_params(latent_dim=30, init_std=0.05)
model = VAE((28, 28, 1), 30, params)

model.fit(params, X_train, X_val=X_val, epochs=10, batch_size=64, lr=0.001)