Bugfixed save/load set transformers with inducing points

bayesflow/attention.py

@@ -191,8 +191,10 @@ def call(self, x, **kwargs):
             Output of shape (batch_size, set_size, input_dim)
         """
 
-        batch_size = x.shape[0]
-        h = self.mab0(tf.stack([self.I] * batch_size), x, **kwargs)
+        batch_size = tf.shape(x)[0]
+        I_expanded = self.I[None, ...]
+        I_tiled = tf.tile(I_expanded, [batch_size, 1, 1])
+        h = self.mab0(I_tiled, x, **kwargs)
         return self.mab1(x, h, **kwargs)
 
 
@@ -240,7 +242,7 @@ def __init__(
             summary_dim, attention_settings, num_dense_fc, dense_settings, use_layer_norm, **kwargs
         )
         init = tf.keras.initializers.GlorotUniform()
-        self.seed_vec = init(shape=(num_seeds, summary_dim))
+        self.seed_vec = tf.Variable(init(shape=(num_seeds, summary_dim)), name="seed_vec", trainable=True)
         self.fc = Sequential([Dense(**dense_settings) for _ in range(num_dense_fc)])
         self.fc.add(Dense(summary_dim))
 
@@ -258,7 +260,9 @@ def call(self, x, **kwargs):
             Output of shape (batch_size, num_seeds * summary_dim)
         """
 
-        batch_size = x.shape[0]
         out = self.fc(x)
-        out = self.mab(tf.stack([self.seed_vec] * batch_size), out, **kwargs)
-        return tf.reshape(out, (out.shape[0], -1))
+        batch_size = tf.shape(x)[0]
+        seed_expanded = self.seed_vec[None, ...]
+        seed_tiled = tf.tile(seed_expanded, [batch_size, 1, 1])
+        out = self.mab(seed_tiled, out, **kwargs)
+        return tf.reshape(out, (tf.shape(out)[0], -1))