V2.py

# TPU-compatible training script (수정본)
import os, requests, math
import numpy as np
import tensorflow as tf
from tensorflow.keras import layers, Model
import sentencepiece as spm

# =========================
# 설정 (필요하면 변경)
# =========================
TOKENIZER_PATH = "bpe.model"
DATA_PATH = "shuffled_corpus.txt"
MAX_LEN = 384
EMBED_DIM = 512
LATENT_DIM = 512
BATCH_SIZE = 768           # global batch size (Keras/TPU가 replica-wise로 나눠서 처리)
EPOCHS = 1
SHUFFLE_BUFFER = 200000
LEARNING_RATE = 1e-4
TEMPERATURE = 0.05
DROPOUT_AUG = 0.1
EMBED_DROPOUT = 0.1
SEED = 42

print('1')
tf.get_logger().setLevel("ERROR")
tf.random.set_seed(SEED)
np.random.seed(SEED)

# =========================
# TPU 초기화 / 분산전략 선택
# =========================
on_tpu = False
try:
    resolver = tf.distribute.cluster_resolver.TPUClusterResolver(tpu="local")
    tf.tpu.experimental.initialize_tpu_system(resolver)
    strategy = tf.distribute.TPUStrategy(resolver)
    print("✅ TPU 초기화 완료:", resolver.cluster_spec().as_dict())
    on_tpu = True
except Exception as e:
    print("⚠️ TPU 미사용, GPU/CPU로 진행:", e)
    strategy = tf.distribute.get_strategy()

# Mixed precision
from tensorflow.keras import mixed_precision
policy = mixed_precision.Policy("mixed_bfloat16" if on_tpu else "float32")
mixed_precision.set_global_policy(policy)
print("✅ Mixed precision:", policy)

# =========================
# 파일 다운로드 (필요시)
# =========================
def download_file(url, save_path):
    if os.path.exists(save_path):
        print(f"exists: {save_path}")
        return
    print(f"Downloading {save_path} ...")
    r = requests.get(url, stream=True)
    r.raise_for_status()
    with open(save_path, "wb") as f:
        for chunk in r.iter_content(8192*2):
            if not chunk:
                break
            f.write(chunk)
    print(f"✅ {save_path} saved")

# (필요하면 주석 해제/사용)
download_file(
   "https://huggingface.co/datasets/OpenLab-NLP/ko-corpus/resolve/main/bpe.model?download=true",
    TOKENIZER_PATH
)
download_file(
     "https://huggingface.co/datasets/OpenLab-NLP/ko-corpus/resolve/main/shuffled_corpus%20(1).txt?download=true",
    DATA_PATH 
)

# =========================
# Tokenizer 로드
# =========================
sp = spm.SentencePieceProcessor()
sp.load(TOKENIZER_PATH)
pad_id = sp.piece_to_id("<pad>")
if pad_id == -1:
    pad_id = 0
vocab_size = sp.get_piece_size()
print("vocab_size:", vocab_size, "pad_id:", pad_id)

# =========================
# 인코딩/데이터 파이프라인
# =========================
def encode_sentence_py(s: str):
    ids = sp.encode(s, out_type=int)[:MAX_LEN]
    if len(ids) < MAX_LEN:
        ids = ids + [pad_id] * (MAX_LEN - len(ids))
    else:
        ids = ids[:MAX_LEN]
    return np.array(ids, dtype=np.int32)

def tf_encode(line):
    def _encode_py(s_tensor):
        s = s_tensor.numpy().decode("utf-8")
        return encode_sentence_py(s)
    ids = tf.py_function(func=_encode_py, inp=[line], Tout=tf.int32)
    ids.set_shape([MAX_LEN])
    return ids

def token_dropout(tokens, drop_prob=DROPOUT_AUG):
    rnd = tf.random.uniform(tf.shape(tokens), 0, 1)
    keep_mask = rnd > drop_prob
    return tf.where(keep_mask, tokens, tf.cast(pad_id, tf.int32))

def make_views(tokens):
    v1 = token_dropout(tokens)
    v2 = token_dropout(tokens)
    return v1, v2

# 텍스트파일 기반 데이터셋
ds = tf.data.TextLineDataset(DATA_PATH)
ds = ds.map(lambda x: tf.strings.strip(x), num_parallel_calls=tf.data.AUTOTUNE)
ds = ds.filter(lambda x: tf.not_equal(x, ""))

ds = ds.map(tf_encode, num_parallel_calls=tf.data.AUTOTUNE)
ds = ds.shuffle(SHUFFLE_BUFFER, seed=SEED)
ds = ds.repeat()
ds = ds.map(lambda t: make_views(t), num_parallel_calls=tf.data.AUTOTUNE)
ds = ds.batch(BATCH_SIZE, drop_remainder=True)
# model.fit expects (inputs, labels), labels는 사용하지 않음(더미)
ds = ds.map(lambda v1, v2: ((v1, v2), tf.zeros([BATCH_SIZE], dtype=tf.float32)), num_parallel_calls=tf.data.AUTOTUNE)
ds = ds.prefetch(tf.data.AUTOTUNE)



class MixerBlock(layers.Layer):
    def __init__(self, seq_len, dim, token_mlp_dim, channel_mlp_dim, dropout=0.0):
        super().__init__()
        self.seq_len = seq_len
        self.dim = dim
        self.token_mlp_dim = token_mlp_dim
        self.channel_mlp_dim = channel_mlp_dim

        self.ln1 = layers.LayerNormalization(epsilon=1e-6, dtype=tf.float32)
        # token-mixing MLP: operate over tokens => apply Dense on transposed axis
        self.token_fc1 = layers.Dense(token_mlp_dim, activation='gelu', dtype=tf.float32)
        self.token_fc2 = layers.Dense(seq_len, dtype=tf.float32)

        self.ln2 = layers.LayerNormalization(epsilon=1e-6, dtype=tf.float32)
        # channel-mixing MLP: operate per-token over channels
        self.channel_fc1 = layers.Dense(channel_mlp_dim, activation='gelu', dtype=tf.float32)
        self.channel_fc2 = layers.Dense(dim, dtype=tf.float32)

        self.dropout = layers.Dropout(dropout)

    def call(self, x, training=None):
        # x: (B, L, D)
        B = tf.shape(x)[0]
        L = tf.shape(x)[1]
        D = tf.shape(x)[2]

        # Token-mixing
        y = self.ln1(x)                          # (B, L, D)
        y_t = tf.transpose(y, perm=[0,2,1])     # (B, D, L)
        y_t = self.token_fc1(y_t)               # (B, D, token_mlp_dim)
        y_t = self.token_fc2(y_t)               # (B, D, L)
        y = tf.transpose(y_t, perm=[0,2,1])     # (B, L, D)
        x = x + self.dropout(y, training=training)

        # Channel-mixing
        z = self.ln2(x)
        z = self.channel_fc1(z)
        z = self.channel_fc2(z)
        x = x + self.dropout(z, training=training)

        return x

class L2NormLayer(layers.Layer):
    def __init__(self, axis=1, epsilon=1e-10, **kwargs):
        super().__init__(**kwargs)
        self.axis = axis
        self.epsilon = epsilon
    def call(self, inputs):
        return tf.math.l2_normalize(inputs, axis=self.axis, epsilon=self.epsilon)

class SentenceEncoder(Model):
    def __init__(self, vocab_size, embed_dim=EMBED_DIM, latent_dim=LATENT_DIM, max_len=MAX_LEN, pad_id=pad_id, dropout_rate=EMBED_DROPOUT):
        super().__init__()
        self.pad_id = pad_id
        self.embed = layers.Embedding(vocab_size, embed_dim)
        self.pos_embed = layers.Embedding(input_dim=max_len, output_dim=embed_dim)
        self.dropout = layers.Dropout(dropout_rate)
        self.blocks = [MixerBlock(seq_len=MAX_LEN, dim=embed_dim, token_mlp_dim=256, channel_mlp_dim=embed_dim, dropout=0.1) for _ in range(3)]
        self.attn_pool = layers.Dense(1)
        
        self.ln_f = layers.LayerNormalization(epsilon=1e-5, dtype=tf.float32)
        
        self.latent = layers.Dense(latent_dim, activation=None)
        self.l2norm = L2NormLayer(axis=1)
        
        self.fc1 = layers.Dense(1152)
        self.fc2 = layers.Dense(embed_dim)

    def call(self, x, training=None):
        positions = tf.range(tf.shape(x)[1])[tf.newaxis, :]
        x_embed = self.embed(x) + self.pos_embed(positions)
        x_embed = self.dropout(x_embed, training=training)

        mask = tf.cast(tf.not_equal(x, self.pad_id), tf.float32)

        h = x_embed
        for block in self.blocks:
            h = block(h)

        v = h
        h = self.fc1(v)
        g, v_split = tf.split(h, 2, axis=-1)
        h = tf.nn.silu(g) * v_split
        h = self.fc2(h)
        h = self.ln_f(h)

        # 🔥 scores를 float32 강제
        scores = self.attn_pool(h)
        scores = tf.cast(scores, tf.float32)

        scores = tf.where(mask[..., tf.newaxis] == 0, tf.constant(-1e9, tf.float32), scores)
        scores = tf.nn.softmax(scores, axis=1)

        pooled = tf.reduce_sum(h * scores, axis=1)
        latent = self.latent(pooled)
        latent = self.l2norm(latent)

        # 🔥 출력만 float32
        return tf.cast(latent, tf.float32)

# contrastive wrapper: 두 뷰를 인코딩하고 (2B, D) concat 반환
def build_contrastive_model(vocab_size):
    encoder = SentenceEncoder(vocab_size=vocab_size)
    input1 = layers.Input(shape=(MAX_LEN,), dtype=tf.int32, name="view1")
    input2 = layers.Input(shape=(MAX_LEN,), dtype=tf.int32, name="view2")
    z1 = encoder(input1)
    z2 = encoder(input2)
    out = layers.Concatenate(axis=0)([z1, z2])  # (2B, D)
    return Model(inputs=[input1, input2], outputs=out), encoder

def nt_xent_loss(y_true, y_pred):
    # y_pred: (2N, D) normalized
    z = y_pred
    z = tf.cast(z, tf.float32)
    sim = tf.matmul(z, z, transpose_b=True)  # (2N, 2N)
    sim = sim / TEMPERATURE
    # large negative on diagonal to avoid trivial argmax
    diag = tf.eye(tf.shape(sim)[0])
    sim = sim - diag * 1e9
    N2 = tf.shape(sim)[0]
    N = N2 // 2
    # positive index for i: if i < N => i+N, else i-N
    labels_pos = tf.concat([tf.range(N, N2), tf.range(0, N)], axis=0)
    loss = tf.nn.sparse_softmax_cross_entropy_with_logits(labels=labels_pos, logits=sim)
    return tf.reduce_mean(loss)

# =========================
# 모델 생성 / 컴파일 (strategy.scope 안에서)
# =========================
with strategy.scope():
    model, encoder = build_contrastive_model(vocab_size)
    optimizer = tf.keras.optimizers.Adam(learning_rate=LEARNING_RATE)
    model.compile(optimizer=optimizer, loss=nt_xent_loss)
    model.summary()

# =========================
# steps_per_epoch 계산 (파일 라인 수 기반)
# =========================
try:
    with open(DATA_PATH, "r", encoding="utf-8") as f:
        num_lines = sum(1 for _ in f)
except Exception as e:
    print("Warning: 데이터 파일 라인 수 계산 실패:", e)
    num_lines = None

if num_lines:
    steps_per_epoch = max(1, num_lines // BATCH_SIZE)
else:
    # fallback (작게 잡음)
    steps_per_epoch = 1000

print("steps_per_epoch:", steps_per_epoch)


history = model.fit(ds, epochs=EPOCHS, steps_per_epoch=steps_per_epoch, verbose=1)

# encoder 가중치 저장
encoder.save_weights("encoder_fit.weights.h5")
print("Training finished and weights saved.")