app.py

import numpy as np
import tensorflow as tf
from tensorflow.keras import layers
import sentencepiece as spm
import gradio as gr
import requests
import os

# ----------------------
# 파일 다운로드 유틸
# ----------------------
def download_file(url, 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):
            f.write(chunk)
    print(f"✅ {save_path} 저장됨")

MODEL_PATH = "encoder.weights.h5"
TOKENIZER_PATH = "bpe.model"

if not os.path.exists(MODEL_PATH):
    download_file(
        "https://huggingface.co/OpenLab-NLP/openlem2/resolve/main/encoder_fit.weights.h5?download=true",
        MODEL_PATH
    )

if not os.path.exists(TOKENIZER_PATH):
    download_file(
        "https://huggingface.co/OpenLab-NLP/openlem2/resolve/main/bpe.model?download=true",
        TOKENIZER_PATH
    )

MAX_LEN = 384
EMBED_DIM = 512
LATENT_DIM = 512
DROP_RATE = 0.1 

# ===============================
# 1️⃣ 토크나이저 로딩
# ===============================
sp = spm.SentencePieceProcessor(TOKENIZER_PATH)
pad_id = sp.piece_to_id("<pad>") if sp.piece_to_id("<pad>") != -1 else 0
vocab_size = sp.get_piece_size()

def encode_sentence(sentence, max_len=MAX_LEN):
    return sp.encode(sentence, out_type=int)[:max_len]

def pad_sentence(tokens):
    return tokens + [pad_id]*(MAX_LEN - len(tokens))


class DynamicConv(layers.Layer):
    def __init__(self, d_model, k=7):
        super().__init__()
        assert k % 2 == 1
        self.k = k
        self.dense = layers.Dense(d_model, activation='silu')
        self.proj = layers.Dense(d_model)
        self.generator = layers.Dense(k, dtype='float32')
    def call(self, x):
        x_in = x
        x = tf.cast(x, tf.float32)

        B = tf.shape(x)[0]
        L = tf.shape(x)[1]
        D = tf.shape(x)[2]

        kernels = self.generator(self.dense(x))
        kernels = tf.nn.softmax(kernels, axis=-1)

        pad = (self.k - 1) // 2
        x_pad = tf.pad(x, [[0,0],[pad,pad],[0,0]])

        x_pad_4d = tf.expand_dims(x_pad, axis=1)
        patches = tf.image.extract_patches(
            images=x_pad_4d,
            sizes=[1,1,self.k,1],
            strides=[1,1,1,1],
            rates=[1,1,1,1],
            padding='VALID'
        )
        patches = tf.reshape(patches, [B, L, self.k, D])

        kernels_exp = tf.expand_dims(kernels, axis=-1)
        out = tf.reduce_sum(patches * kernels_exp, axis=2)
        out = self.proj(out)

        # 🔥 원래 dtype으로 돌려줌
        return tf.cast(out, x_in.dtype)

class EncoderBlock(tf.keras.layers.Layer):
    def __init__(self, embed_dim=EMBED_DIM, ff_dim=1152, seq_len=MAX_LEN, num_conv_layers=2):
        super().__init__()
        self.embed_dim = embed_dim
        self.seq_len = seq_len

        # MLP / FFN
        self.fc1 = layers.Dense(ff_dim)
        self.fc2 = layers.Dense(embed_dim)
        self.blocks = [DynamicConv(d_model=embed_dim, k=7) for _ in range(num_conv_layers)]
        # LayerNorm
        self.ln = layers.LayerNormalization(epsilon=1e-5)   # 입력 정규화
        self.ln1 = layers.LayerNormalization(epsilon=1e-5)  # Conv residual
        self.ln2 = layers.LayerNormalization(epsilon=1e-5)  # FFN residual

    def call(self, x, mask=None):
        # 입력 정규화
        x_norm = self.ln(x)

        # DynamicConv 여러 층 통과
        out = x_norm
        for block in self.blocks: out = block(out)
        # Conv residual 연결
        x = x_norm + self.ln1(out)

        # FFN / GLU
        v = out
        h = self.fc1(v)
        g, v_split = tf.split(h, 2, axis=-1)
        h = tf.nn.silu(g) * v_split
        h = self.fc2(h)

        # FFN residual 연결
        x = x + self.ln2(h)

        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(tf.keras.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 = [EncoderBlock() for _ in range(2)]
        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)

    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, training=training)

        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)

# 3️⃣ 모델 로드
# ===============================
encoder = SentenceEncoder(vocab_size=vocab_size)
encoder(np.zeros((1, MAX_LEN), dtype=np.int32))  # 모델 빌드
encoder.load_weights(MODEL_PATH)

# ===============================
# 4️⃣ 벡터화 함수
# ===============================
def get_sentence_vector(sentence):
    tokens = pad_sentence(encode_sentence(sentence))
    vec = encoder(np.array([tokens])).numpy()[0]
    return vec / np.linalg.norm(vec)

# ===============================
# 5️⃣ 가장 비슷한 문장 찾기
# ===============================
def find_most_similar(query, s1, s2, s3):
    candidates = [s1, s2, s3]
    candidate_vectors = np.stack([get_sentence_vector(c) for c in candidates]).astype(np.float32)
    query_vector = get_sentence_vector(query)
    
    sims = candidate_vectors @ query_vector  # cosine similarity
    top_idx = np.argmax(sims)
    
    return {
        "가장 비슷한 문장": candidates[top_idx],
        "유사도": float(sims[top_idx])
    }

# ===============================
# 6️⃣ Gradio UI
# ===============================
with gr.Blocks() as demo:
    gr.Markdown("## 🔍 문장 유사도 검색기 (쿼리 1개 + 후보 3개)")
    with gr.Row():
        query_input = gr.Textbox(label="검색할 문장 (Query)", placeholder="여기에 입력")
    with gr.Row():
        s1_input = gr.Textbox(label="검색 후보 1")
        s2_input = gr.Textbox(label="검색 후보 2")
        s3_input = gr.Textbox(label="검색 후보 3")
    output = gr.JSON(label="결과")
    
    search_btn = gr.Button("가장 비슷한 문장 찾기")
    search_btn.click(
        fn=find_most_similar,
        inputs=[query_input, s1_input, s2_input, s3_input],
        outputs=output
    )

demo.launch()