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/openlem3-retrieval-qa/resolve/main/encoder_fit.weights.h5?download=true",
        MODEL_PATH
    )

if not os.path.exists(TOKENIZER_PATH):
    download_file(
        "https://huggingface.co/OpenLab-NLP/openlem3-retrieval-qa/resolve/main/bpe.model?download=true",
        TOKENIZER_PATH
    )

MAX_LEN = 384
TOP_K = 3
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
DROPOUT_AUG = 0.1
EMBED_DROPOUT = 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 HyperConv1D(layers.Layer):
    def __init__(self, d_model, k=7, mem_size=64, hyper_dim=128, dropout=0.0):
        super().__init__()
        assert k % 2 == 1
        self.k = k
        self.d_model = d_model
        self.mem_size = mem_size

        # Input projection
        self.input_proj = layers.Dense(d_model, name="input_proj")

        # Local depthwise conv
        self.local_conv = layers.DepthwiseConv1D(kernel_size=k, padding='same', activation='silu')
        self.local_proj = layers.Dense(d_model, name="local_proj")

        # Hypernetwork: global -> scale vector
        self.hyper = tf.keras.Sequential([
            layers.Dense(hyper_dim, activation='gelu'),
            layers.Dense(d_model)
        ], name="hyper")

        # Associative memory
        self.mem_keys = self.add_weight((mem_size, d_model), initializer='glorot_uniform', trainable=True)
        self.mem_vals = self.add_weight((mem_size, d_model), initializer='glorot_uniform', trainable=True)
        self.mem_proj = layers.Dense(d_model)

        self.norm = layers.LayerNormalization()
        self.attn_pool = layers.Dense(1)

    def call(self, x):
        x_in = x
        x_dtype = x.dtype  # 입력 dtype 기억

        # 1) input projection
        x_proj = self.input_proj(x)
        # memory와 연산 위해 dtype 통일
        mem_dtype = self.mem_keys.dtype
        x_proj = tf.cast(x_proj, mem_dtype)

        # 2) local conv
        out_local = self.local_conv(x_proj)
        # hypernetwork scaling
        global_z = self.attn_pool(x_proj)
        global_z = tf.nn.softmax(global_z, axis=1)
        global_z = tf.reduce_sum(x_proj * global_z, axis=1)

        scale = tf.expand_dims(tf.nn.sigmoid(self.hyper(global_z)), 1)
        out_local = out_local * scale
        out_local = self.local_proj(out_local)


        # 3) associative memory
        sims = tf.matmul(x_proj, self.mem_keys, transpose_b=True) / tf.math.sqrt(tf.cast(self.d_model, mem_dtype))
        attn = tf.nn.softmax(sims, axis=-1)
        mem_read = tf.matmul(attn, self.mem_vals)
        mem_read = self.mem_proj(mem_read)

        # 4) fuse & residual
        out = out_local + mem_read
        out = self.norm(x_proj + out)
        out = tf.nn.silu(out)

        # 최종 출력 dtype 원래 입력 dtype으로 캐스트
        return tf.cast(out, x_dtype)

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 = [HyperConv1D(d_model=embed_dim, k=7, mem_size=128, hyper_dim=256) for _ in range(4)]
        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)

# 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)

def tokenize(texts):
    token_ids = []
    for t in texts:
        ids = sp.encode(t, out_type=int)[:MAX_LEN]
        if len(ids) < MAX_LEN:
            ids += [pad_id]*(MAX_LEN-len(ids))
        token_ids.append(ids)
    return np.array(token_ids, dtype=np.int32)


def search_and_answer(query, docs_text):
    docs = [d.strip() for d in docs_text.split("\n") if d.strip()]
    if not docs:
        return [], "문서를 한 줄씩 입력하세요."

    q_ids = tokenize([query])
    d_ids = tokenize(docs)

    q_emb = encoder(q_ids, training=False).numpy()
    d_embs = encoder(d_ids, training=False).numpy()

    scores = np.dot(q_emb, d_embs.T)[0]
    top_k_idx = scores.argsort()[::-1][:min(TOP_K, len(docs))]
    top_docs = [(docs[i], float(scores[i])) for i in top_k_idx]

    answer = docs[top_k_idx[0]]
    return top_docs, answer

with gr.Blocks() as demo:
    gr.Markdown("## OpenLEM2 Retrieval-QA 데모 (사용자 문서 입력 가능)")

    with gr.Row():
        query_input = gr.Textbox(label="질문/쿼리", placeholder="예: 서울 날씨 어때?")
    
    docs_input = gr.Textbox(label="문서 리스트 (한 줄씩)", placeholder="문서를 한 줄씩 입력하세요.", lines=10)

    with gr.Row():
        top_docs_out = gr.Dataframe(headers=["Document", "Score"])

        answer_out = gr.Textbox(label="답변")

    run_btn = gr.Button("검색/QA 실행")
    run_btn.click(fn=search_and_answer, inputs=[query_input, docs_input], outputs=[top_docs_out, answer_out])

demo.launch()