Update app.py

app.py

@@ -35,7 +35,7 @@ if not os.path.exists(TOKENIZER_PATH):
 MAX_LEN = 128
 EMBED_DIM = 384
 LATENT_DIM = 384
-DROPOUT_RATE = 0.01
+DROP_RATE = 0.1 
 
 # ===============================
 # 1️⃣ 토크나이저 로딩
@@ -51,12 +51,13 @@ def pad_sentence(tokens):
     return tokens + [pad_id]*(MAX_LEN - len(tokens))
 
 class EncoderBlock(tf.keras.layers.Layer):
-    def __init__(self, embed_dim=EMBED_DIM, ff_dim=1152, seq_len=MAX_LEN):
+    def __init__(self, embed_dim=EMBED_DIM, ff_dim=1152, seq_len=MAX_LEN, drop_rate=DROP_RATE):
         super().__init__()
         self.embed_dim = embed_dim
         self.seq_len = seq_len
+        self.drop_rate = drop_rate
 
-        self.fc1 = layers.Dense(ff_dim)
+        self.fc1 = layers.Dense(ff_dim*2)
         self.fc2 = layers.Dense(embed_dim)
         self.fc3 = layers.Dense(ff_dim//2)
         self.fc4 = layers.Dense(embed_dim)
@@ -71,15 +72,15 @@ class EncoderBlock(tf.keras.layers.Layer):
         self.ln3 = layers.LayerNormalization(epsilon=1e-5)
         self.ln4 = layers.LayerNormalization(epsilon=1e-5)
 
-    def call(self, x, mask):
-        mask = mask
-        # x: (B, L, D)
+        self.dropout = layers.Dropout(drop_rate)
+
+    def call(self, x, mask, training=False):
         x_norm = self.ln(x)
 
         h = self.fc1(x_norm)
-        g, v = tf.split(h, 2, axis=-1) 
+        g, v = tf.split(h, 2, axis=-1)
         h = tf.nn.silu(g) * v
-        h = self.fc2(h)  
+        h = self.fc2(h)
 
         h = x + self.ln1(h)
 
@@ -92,12 +93,13 @@ class EncoderBlock(tf.keras.layers.Layer):
         v = self.token_mixer(v) * self.token_gate(v)
         v = tf.transpose(v, [0, 2, 1])
 
-        x_norm = attn + self.ln3(v)
-        x = self.fc3(x_norm)
+        x_norm2 = attn + self.ln3(v)
+        x = self.fc3(x_norm2)
         x = tf.nn.silu(x)
         x = self.fc4(x)
 
-        return x_norm + self.ln4(x)
+        x = self.dropout(x, training=training)
+        return x_norm2 + self.ln4(x)
 
 class L2NormLayer(layers.Layer):
     def __init__(self, axis=1, epsilon=1e-10, **kwargs):
@@ -110,34 +112,36 @@ 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=384, latent_dim=384, max_len=128, pad_id=pad_id):
+    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):
         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() for _ in range(2)]
+        self.blocks = [EncoderBlock(embed_dim=embed_dim, drop_rate=drop_rate) 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)  # tanh 제거
-        self.l2norm = L2NormLayer()  # 추가
+        self.latent = layers.Dense(latent_dim, activation=None)
+        self.l2norm = L2NormLayer()
+        self.drop_embed = layers.Dropout(drop_rate)
 
-    def call(self, x):
+    def call(self, x, training=False):
         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)
-        x = x_embed
+        h = x_embed
         for block in self.blocks:
-            x = block(x, mask)
-        x = self.ln_f(x)
+            h = block(h, mask, training=training)
+        h = self.ln_f(h)
 
-        scores = self.attn_pool(x)
+        scores = self.attn_pool(h)
         scores = tf.where(tf.equal(mask[..., tf.newaxis], 0), -1e9, scores)
         scores = tf.nn.softmax(scores, axis=1)
-        pooled = tf.reduce_sum(x * scores, axis=1)
+        pooled = tf.reduce_sum(h * scores, axis=1)
 
         latent = self.latent(pooled)
-        return self.l2norm(latent)  # L2 정규화 후 반환
-
+        return self.l2norm(latent)
 # 3️⃣ 모델 로드
 # ===============================
 encoder = SentenceEncoder(vocab_size=vocab_size)