Update app.py

app.py

@@ -22,7 +22,7 @@ TOKENIZER_PATH = "bpe.model"
 
 if not os.path.exists(MODEL_PATH):
     download_file(
-        "https://huggingface.co/OpenLab-NLP/openlem-prototype/resolve/main/encoder_simcse.weights.h5?download=true",
+        "https://huggingface.co/OpenLab-NLP/openlem-prototype/resolve/main/encoder.weights.h5?download=true",
         MODEL_PATH
     )
 
@@ -50,56 +50,93 @@ def encode_sentence(sentence, max_len=MAX_LEN):
 def pad_sentence(tokens):
     return tokens + [pad_id]*(MAX_LEN - len(tokens))
 
+class DynamicConv(layers.Layer):
+    def __init__(self, k=7):
+        super().__init__()
+        assert k % 2 == 1, "kernel size should be odd for symmetric padding"
+        self.k = k
+        # generator는 각 토큰에 대해 k개의 로짓을 뱉음 -> softmax로 가중치화
+        self.generator = layers.Dense(k)
+
+    def call(self, x):
+        # x: (B, L, D)
+        B = tf.shape(x)[0]
+        L = tf.shape(x)[1]
+        D = tf.shape(x)[2]
+
+        # (B, L, k) logits -> softmax -> (B, L, k)
+        kernels = self.generator(x)
+        kernels = tf.nn.softmax(kernels, axis=-1)
+
+        # padding (same)
+        pad = (self.k - 1) // 2
+        x_pad = tf.pad(x, [[0, 0], [pad, pad], [0, 0]])  # (B, L+2pad, D)
+
+        # extract patches using tf.image.extract_patches:
+        # make 4D: (B, H=1, W=L+2pad, C=D)
+        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'
+        )  # (B, 1, L, k*D)
+
+        # reshape -> (B, L, k, D)
+        patches = tf.reshape(patches, [B, 1, L, self.k * D])
+        patches = tf.squeeze(patches, axis=1)
+        patches = tf.reshape(patches, [B, L, self.k, D])
+
+        # kernels: (B, L, k) -> (B, L, k, 1)
+        kernels_exp = tf.expand_dims(kernels, axis=-1)
+
+        # weighted sum over kernel dim -> (B, L, D)
+        out = tf.reduce_sum(patches * kernels_exp, axis=2)
+
+        return out
+
 class EncoderBlock(tf.keras.layers.Layer):
-    def __init__(self, embed_dim=EMBED_DIM, ff_dim=1152, seq_len=MAX_LEN, drop_rate=DROP_RATE):
+    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
-        self.drop_rate = drop_rate
 
-        self.fc1 = layers.Dense(ff_dim*2)
+        # MLP / FFN
+        self.fc1 = layers.Dense(ff_dim)
         self.fc2 = layers.Dense(embed_dim)
-        self.fc3 = layers.Dense(ff_dim//2)
-        self.fc4 = layers.Dense(embed_dim)
-
-        self.attn = layers.Dense(1)
-        self.token_mixer = layers.Dense(seq_len)
-        self.token_gate = layers.Dense(seq_len, activation='sigmoid')
 
-        self.ln = layers.LayerNormalization(epsilon=1e-5)
-        self.ln1 = layers.LayerNormalization(epsilon=1e-5)
-        self.ln2 = layers.LayerNormalization(epsilon=1e-5)
-        self.ln3 = layers.LayerNormalization(epsilon=1e-5)
-        self.ln4 = layers.LayerNormalization(epsilon=1e-5)
+        # DynamicConv 블록 여러 개 쌓기
+        self.blocks = [DynamicConv(k=7) for _ in range(num_conv_layers)]
 
-        self.dropout = layers.Dropout(drop_rate)
+        # 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, training=False):
+    def call(self, x, mask=None):
+        # 입력 정규화
         x_norm = self.ln(x)
 
-        h = self.fc1(x_norm)
-        g, v = tf.split(h, 2, axis=-1)
-        h = tf.nn.silu(g) * v
-        h = self.fc2(h)
+        # DynamicConv 여러 층 통과
+        out = x_norm
+        for block in self.blocks:
+            out = block(out)
 
-        h = x + self.ln1(h)
+        # Conv residual 연결
+        x = x_norm + self.ln1(out)
 
-        scores = self.attn(h)
-        scores = tf.where(tf.equal(mask[..., tf.newaxis], 0), -1e9, scores)
-        scores = tf.nn.softmax(scores, axis=1)
-        attn = h + self.ln2(h * scores)
-
-        v = tf.transpose(attn, [0, 2, 1])
-        v = self.token_mixer(v) * self.token_gate(v)
-        v = tf.transpose(v, [0, 2, 1])
+        # 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)
 
-        x_norm2 = attn + self.ln3(v)
-        x = self.fc3(x_norm2)
-        x = tf.nn.silu(x)
-        x = self.fc4(x)
+        # FFN residual 연결
+        x = x + self.ln2(h)
 
-        x = self.dropout(x, training=training)
-        return x_norm2 + self.ln4(x)
+        return x
 
 class L2NormLayer(layers.Layer):
     def __init__(self, axis=1, epsilon=1e-10, **kwargs):
@@ -112,36 +149,34 @@ class L2NormLayer(layers.Layer):
         return {"axis": self.axis, "epsilon": self.epsilon, **super().get_config()}
 
 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, drop_rate=DROP_RATE):
+    def __init__(self, vocab_size, embed_dim=384, latent_dim=384, max_len=128, pad_id=pad_id):
         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.blocks = [EncoderBlock(embed_dim=embed_dim, drop_rate=drop_rate) for _ in range(2)]
+        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()
-        self.drop_embed = layers.Dropout(drop_rate)
+        self.latent = layers.Dense(latent_dim, activation=None)  # tanh 제거
+        self.l2norm = L2NormLayer()  # 추가
 
-    def call(self, x, training=False):
+    def call(self, x):
         positions = tf.range(tf.shape(x)[1])[tf.newaxis, :]
         x_embed = self.embed(x) + self.pos_embed(positions)
-        x_embed = self.drop_embed(x_embed, training=training)
-
         mask = tf.cast(tf.not_equal(x, self.pad_id), tf.float32)
-        h = x_embed
+        x = x_embed
         for block in self.blocks:
-            h = block(h, mask, training=training)
-        h = self.ln_f(h)
+            x = block(x, mask)
+        x = self.ln_f(x)
 
-        scores = self.attn_pool(h)
+        scores = self.attn_pool(x)
         scores = tf.where(tf.equal(mask[..., tf.newaxis], 0), -1e9, scores)
         scores = tf.nn.softmax(scores, axis=1)
-        pooled = tf.reduce_sum(h * scores, axis=1)
+        pooled = tf.reduce_sum(x * scores, axis=1)
 
         latent = self.latent(pooled)
-        return self.l2norm(latent)
+        return self.l2norm(latent)  # L2 정규화 후 반환
+
 # 3️⃣ 모델 로드
 # ===============================
 encoder = SentenceEncoder(vocab_size=vocab_size)