TF: add onnx_export option and onnx_comp_floor_div (#1453)

returnn/tensor/_dim_extra.py

@@ -1182,8 +1182,12 @@ def _bin_op_tf(a, b):
             elif kind == "mul":
                 return a * b
             elif kind in ("floordiv", "truediv"):  # truediv assumes there is no remainder
+                if util.is_onnx_export_global():
+                    return tf_util.onnx_compat_floor_div(a, b)
                 return a // b
             elif kind == "ceildiv":
+                if util.is_onnx_export_global():
+                    return -tf_util.onnx_compat_floor_div(-a, b)
                 return -(-a // b)
             else:
                 raise ValueError("unknown op kind %r" % op.kind)

returnn/tf/frontend_low_level/_backend.py

@@ -8,7 +8,7 @@
 import tensorflow as tf
 
 import returnn.tf.compat as tf_compat
-from returnn.util.basic import NotSpecified
+from returnn.util.basic import NotSpecified, is_onnx_export_global
 from returnn.tensor import Tensor, Dim
 from returnn.tf.util import basic as tf_util
 
@@ -132,14 +132,17 @@ def combine_raw(a: tf.Tensor, kind: str, b: tf.Tensor) -> tf.Tensor:
         :return: a `kind` b
         """
         assert a.shape.ndims == b.shape.ndims or a.shape.ndims == 0 or b.shape.ndims == 0
-        kind = {
-            "sub": "subtract",
-            "mul": "multiply",
-        }.get(kind, kind)
-        op = getattr(tf, kind, None)  # e.g. tf.add
-        # In tf v2, some ops like floordiv or mod exist in the tf.math namespace instead
-        if op is None:
-            op = getattr(tf.math, kind)
+        if kind == "floordiv" and is_onnx_export_global():
+            op = tf_util.onnx_compat_floor_div
+        else:
+            kind = {
+                "sub": "subtract",
+                "mul": "multiply",
+            }.get(kind, kind)
+            op = getattr(tf, kind, None)  # e.g. tf.add
+            # In tf v2, some ops like floordiv or mod exist in the tf.math namespace instead
+            if op is None:
+                op = getattr(tf.math, kind)
         with tf_util.same_control_flow_ctx([a, b]):
             return op(a, b)
 

returnn/tf/util/basic.py

@@ -7820,3 +7820,14 @@ def is_axis_from_description_recurrent(axis, network, data):
             if axis == single_step_dim:
                 return True
     return False
+
+
+def onnx_compat_floor_div(a: tf.Tensor, b: tf.Tensor) -> tf.Tensor:
+    """
+    :param a:
+    :param b:
+    :return: for onnx export compatible floor_divide
+    """
+    # https://github.com/onnx/tensorflow-onnx/issues/2174
+    abs_a, abs_b = tf.abs(a), tf.abs(b)
+    return tf.where(a * b >= 0, a // b, -abs_a // abs_b - tf.cast(abs_a % abs_b != 0, dtype=a.dtype))

returnn/util/basic.py

@@ -3627,6 +3627,18 @@ def get_global_inf_value() -> float:
     return config.float("inf_value", _default_global_inf_value)
 
 
+def is_onnx_export_global() -> bool:
+    """
+    :return: False by default. If 'onnx_export' is set in the config, that value is used.
+    """
+    from returnn.config import get_global_config
+
+    config = get_global_config(raise_exception=False)
+    if not config:
+        return False
+    return config.bool("onnx_export", False)
+
+
 # See :func:`maybe_restart_returnn_with_atfork_patch` below for why you might want to use this.
 _c_code_patch_atfork = """
 #define _GNU_SOURCE