MLBEDSW-7316: Fix crash for networks with resource variables
- The problem was that networks with resource variables have
not been thought of. The major problem was the graph traversal
where these ops were not visited resulting in an empty subgraph
that resulted in the crash.
- Fixed the problem by attaching virtual tensors to the ops simulating
subgraph output. These tensors are only used to get the graph
traversal to work.
- Fixed serializing of attribute container and shared_name
- Fixed subgraph index for operator CallOnce
- All resource variable ops are pushed to the CPU
Change-Id: I815f9c81baf7a3fbb686e895980b462f58208b6e
Signed-off-by: Johan Alfven <johan.alfven@arm.com>
diff --git a/ethosu/vela/tflite_writer.py b/ethosu/vela/tflite_writer.py
index e527cd4..3298229 100644
--- a/ethosu/vela/tflite_writer.py
+++ b/ethosu/vela/tflite_writer.py
@@ -1,4 +1,4 @@
-# SPDX-FileCopyrightText: Copyright 2020-2022 Arm Limited and/or its affiliates <open-source-office@arm.com>
+# SPDX-FileCopyrightText: Copyright 2020-2023 Arm Limited and/or its affiliates <open-source-office@arm.com>
#
# SPDX-License-Identifier: Apache-2.0
#
@@ -321,6 +321,10 @@
attrs["dilation_w_factor"] = attrs["dilation"][2]
if "channel_multiplier" in attrs:
attrs["depth_multiplier"] = attrs["channel_multiplier"]
+ if "container" in attrs:
+ attrs["container"] = builder.CreateString(attrs["container"])
+ if "shared_name" in attrs:
+ attrs["shared_name"] = builder.CreateString(attrs["shared_name"])
attrs["fused_activation_function"] = op.activation.op_type if op.activation is not None else None
builtin_opt_offset, custom_opt_offset = opt_serializer.serialize(builder, attrs)
@@ -362,6 +366,13 @@
# to an op.
tensor_set = set(sg.original_inputs)
+ # Remove any virtual outputs since they are only used internally when
+ # traversing the graph.
+ for tens in sg.virtual_outputs:
+ tens.ops[0].outputs = []
+ if tens in sg.output_tensors:
+ sg.output_tensors.remove(tens)
+
# Add the tensors from all valid ops, as well as the tensors from placeholder ops
# This allows us to serialise tensors which arent attached to any specific ops,
# e.g. due to an empty graph containing no ops