try enumerated shapes

scripts/convert2coreml.py

@@ -28,14 +28,21 @@ def convert_encoder_to_tvm(model):
 def convert_decoder_to_tvm(model):
     model.eval()
 
-    tokens_shape = (1, 10)
+    tokens_shape = []
+    max_token = 448
+    # max_token = 448, max number of EnumeratedShapes supported by coreml = 128
+    segment = max_token//128 + 1
+    i = segment
+    while(i<max_token+segment):
+        tokens_shape.append([1, i])
+        i += segment
+
     audio_shape = (1, 1500, 384)
-    token_data = (1000*torch.rand(tokens_shape)).long()
+    token_data = (1000*torch.rand(tokens_shape[0])).long()
     audio_data = torch.rand(audio_shape)
     traced_model = torch.jit.trace(model, (token_data, audio_data))
 
-    token_flexible_shape = ct.Shape(shape=(1,
-                              ct.RangeDim(lower_bound=1, upper_bound=100000, default=1)))
+    token_flexible_shape = ct.EnumeratedShapes(shapes=tokens_shape, default=tokens_shape[0])
 
 
     model = ct.convert(