Fix wrong shapes in loop body inputs if shape invariances are set in …

…TF (#2203) * fix wrong shapes in loop body inputs if shape invariances are set in TF * fix and enable test for TF2 --------- Signed-off-by: f-salvetti <[email protected]>
onnx · Jul 26, 2023 · 6dda2bb · 6dda2bb
1 parent d0ba20e
commit 6dda2bb
Show file tree

Hide file tree

Showing 2 changed files with 5 additions and 6 deletions.
diff --git a/tests/test_loops.py b/tests/test_loops.py
@@ -7,7 +7,7 @@
 import tensorflow as tf
 
 from backend_test_base import Tf2OnnxBackendTestBase
-from common import unittest_main, check_tf_min_version, check_tf_max_version, \
+from common import unittest_main, check_tf_min_version, \
     check_onnxruntime_min_version, check_tfjs_max_version, skip_tflite
 from tf2onnx.tf_loader import is_tf2
 
@@ -286,15 +286,13 @@ def func(x, y):
         self.run_test_case(func, feed_dict, input_names_with_port, output_names_with_port, rtol=1e-5)
 
     @check_tf_min_version("1.9")
-    @check_tf_max_version("1.15")
+    @skip_tflite("infinite loop with tflite")
     def test_simple_while_loop_var_shape(self):
         # test for while_loop with variant shape variables
-        # may not meet ONNX Loop spec
-        # Note: this is not working on tf2 itself.
         def func(i):
             const = tf.constant(np.array([2], dtype=np.int32))
             c = lambda i: tf.reduce_all(tf.shape(i) < 10)
-            b = lambda i: tf.concat([i, const], 0)
+            b = lambda i: [tf.concat([i, const], 0)]
             r = tf.while_loop(c, b, [i], shape_invariants=[tf.TensorShape([None])])
             return tf.identity(r, name="output")
         input_names_with_port = ["input_1:0"]

diff --git a/tf2onnx/onnx_opset/controlflow.py b/tf2onnx/onnx_opset/controlflow.py
@@ -571,7 +571,8 @@ def wire_while_body(parent_g, g, loop_node, body_input_to_state_var, cond_input_
     g.set_dtype(func_inputs[0], onnx_pb.TensorProto.INT64)
     g.inputs = [g.get_node_by_output(inp) for inp in func_inputs]
 
-    for p, c in zip(loop_node.input, func_inputs):
+    # we should use outputs shape, not inputs, since there may be shape invariants
+    for p, c in zip(loop_node.output, func_inputs[2:]):
         g.copy_shape(p, c)
 
     for i, node in enumerate(g.inputs):