[MLBEDSW-3813] Fix LSTM operator pass through
Fixed pass through of LSTM operator.
Change-Id: I23140c69ab6cdc83f6bb8129256b4cc6a7c5ffac
Signed-off-by: Fredrik Svedberg <fredrik.svedberg@arm.com>
diff --git a/ethosu/vela/operation.py b/ethosu/vela/operation.py
index 16431be..e4d11be 100644
--- a/ethosu/vela/operation.py
+++ b/ethosu/vela/operation.py
@@ -407,6 +407,7 @@
"attrs",
"inputs",
"outputs",
+ "intermediates",
"flops",
"scheduled_pass",
"run_on_npu",
@@ -427,6 +428,7 @@
self.attrs: Dict[str, Any] = {}
self.inputs: List[Tensor] = []
self.outputs: List[Tensor] = []
+ self.intermediates: List[Tensor] = []
self.flops = 0
self.run_on_npu = True
# Fused activation function. If not none: operator code.
@@ -453,6 +455,7 @@
res.attrs = dict(self.attrs)
res.inputs = list(self.inputs)
res.outputs = list(self.outputs)
+ res.intermediates = list(self.intermediates)
res.flops = self.flops
res.run_on_npu = self.run_on_npu
res.activation = None if self.activation is None else self.activation.clone()
diff --git a/ethosu/vela/tensor.py b/ethosu/vela/tensor.py
index b7d4307..97885d0 100644
--- a/ethosu/vela/tensor.py
+++ b/ethosu/vela/tensor.py
@@ -341,6 +341,7 @@
"bandwidth_shape",
"dtype",
"name",
+ "is_variable",
"ops",
"consumer_list",
"values",
@@ -378,6 +379,7 @@
self.bandwidth_shape = shape
self.dtype = dtype
self.name = name
+ self.is_variable = False
self.equivalence_id: UUID = uuid.uuid4()
self.ops: List[Operation] = []
diff --git a/ethosu/vela/tflite_reader.py b/ethosu/vela/tflite_reader.py
index ae99c33..daea1bf 100644
--- a/ethosu/vela/tflite_reader.py
+++ b/ethosu/vela/tflite_reader.py
@@ -1,4 +1,4 @@
-# Copyright (C) 2020 Arm Limited or its affiliates. All rights reserved.
+# Copyright (C) 2020-2021 Arm Limited or its affiliates. All rights reserved.
#
# SPDX-License-Identifier: Apache-2.0
#
@@ -109,6 +109,7 @@
dtype = datatype_map[tens_dtype]
tens = Tensor(shape, dtype, name)
quant = tens_data.Quantization()
+ tens.is_variable = tens_data.IsVariable()
tens.quantization = QuantizationParameters()
if quant is not None:
@@ -144,6 +145,10 @@
op_type, opt_serializer, custom_code = self.graph.operator_codes[op_data.OpcodeIndex()]
inputs = [self.tensors[idx] if idx != -1 else None for idx in op_data.InputsAsNumpy()]
outputs = [self.tensors[idx] if idx != -1 else None for idx in op_data.OutputsAsNumpy()]
+ intermediates = []
+ if op_data.IntermediatesLength():
+ intermediates = [self.tensors[idx] if idx != -1 else None for idx in op_data.IntermediatesAsNumpy()]
+
name = "unknown_op_name"
if len(outputs):
name = outputs[0].name
@@ -151,6 +156,7 @@
op.op_index = op_index
op.inputs = inputs
op.outputs = outputs
+ op.intermediates = intermediates
for out in op.outputs:
out.ops = [op]
diff --git a/ethosu/vela/tflite_writer.py b/ethosu/vela/tflite_writer.py
index e190a74..687b887 100644
--- a/ethosu/vela/tflite_writer.py
+++ b/ethosu/vela/tflite_writer.py
@@ -1,4 +1,4 @@
-# Copyright (C) 2020 Arm Limited or its affiliates. All rights reserved.
+# Copyright (C) 2020-2021 Arm Limited or its affiliates. All rights reserved.
#
# SPDX-License-Identifier: Apache-2.0
#
@@ -203,6 +203,7 @@
def serialise_quantization_parameters(self, quant):
builder = self.builder
+ qp = None
min = None
max = None
scale = None
@@ -217,16 +218,18 @@
if quant.zero_point is not None:
zero_point = self.write_long_vector(make_vector(quant.zero_point))
- QuantizationParameters.QuantizationParametersStart(builder)
- if min is not None:
- QuantizationParameters.QuantizationParametersAddMin(builder, min)
- if max is not None:
- QuantizationParameters.QuantizationParametersAddMax(builder, max)
- if scale is not None:
- QuantizationParameters.QuantizationParametersAddScale(builder, scale)
- if zero_point is not None:
- QuantizationParameters.QuantizationParametersAddZeroPoint(builder, zero_point)
- return QuantizationParameters.QuantizationParametersEnd(builder)
+ QuantizationParameters.QuantizationParametersStart(builder)
+ if min is not None:
+ QuantizationParameters.QuantizationParametersAddMin(builder, min)
+ if max is not None:
+ QuantizationParameters.QuantizationParametersAddMax(builder, max)
+ if scale is not None:
+ QuantizationParameters.QuantizationParametersAddScale(builder, scale)
+ if zero_point is not None:
+ QuantizationParameters.QuantizationParametersAddZeroPoint(builder, zero_point)
+ qp = QuantizationParameters.QuantizationParametersEnd(builder)
+
+ return qp
def serialise_tensor(self, tens):
builder = self.builder
@@ -258,7 +261,9 @@
# Empty buffers should be kept unique for TensorFlow Lite Micro
Tensor.TensorAddBuffer(builder, buf_id)
Tensor.TensorAddName(builder, name)
- Tensor.TensorAddQuantization(builder, quant)
+ if quant is not None:
+ Tensor.TensorAddQuantization(builder, quant)
+ Tensor.TensorAddIsVariable(builder, tens.is_variable)
res = Tensor.TensorEnd(builder)
return res
@@ -266,10 +271,15 @@
def serialise_operator(self, op):
builder = self.builder
- inputs_offset = self.write_int_vector([self.tensor_map[tens] for tens in op.inputs if tens in self.tensor_map])
+ inputs_offset = self.write_int_vector(
+ [self.tensor_map[tens] if tens in self.tensor_map else -1 for tens in op.inputs]
+ )
outputs_offset = self.write_int_vector(
[self.tensor_map[tens] for tens in op.outputs if tens in self.tensor_map]
)
+ intermediates_offset = self.write_int_vector(
+ [self.tensor_map[tens] for tens in op.intermediates if tens in self.tensor_map]
+ )
if op.type == Op.Custom:
op_idx, tflop, opt_serializer = self.operator_code_map[op.type][op.attrs.get("custom_code", "")]
@@ -300,6 +310,7 @@
Operator.OperatorAddOpcodeIndex(builder, op_idx)
Operator.OperatorAddInputs(builder, inputs_offset)
Operator.OperatorAddOutputs(builder, outputs_offset)
+ Operator.OperatorAddIntermediates(builder, intermediates_offset)
if builtin_opt_offset is not None:
Operator.OperatorAddBuiltinOptionsType(builder, opt_serializer.builtin_opt_type)
@@ -328,7 +339,7 @@
# This allows us to serialise tensors which arent attached to any specific ops,
# e.g. due to an empty graph containing no ops
for op in all_ops + placeholder_ops:
- for tens in op.inputs + op.outputs:
+ for tens in op.inputs + op.outputs + op.intermediates:
if tens is not None:
tensor_set.add(tens)