#!/usr/bin/env python
# ******************************************************************************
# Copyright 2022 Brainchip Holdings Ltd.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ******************************************************************************
__all__ = ["Add", "QuantizedAdd", "QuantizedConcatenate"]
import keras
import tensorflow as tf
from .layers_base import (register_quantize_target, register_no_output_quantizer,
register_aligned_inputs, apply_buffer_bitwidth, QuantizedLayer)
from .quantizers import OutputQuantizer
from .recorders import TensorRecorder
from ..tensors import FixedPoint
[docs]
@keras.saving.register_keras_serializable()
class Add(keras.layers.Layer):
"""Wrapper class of `keras.layers.Add` that allows to average inputs.
We only support a tuple of two inputs with the same shape.
Args:
average (bool, optional): if `True`, compute the average across all inputs.
Defaults to `False`.
activation (bool, optional): If `True`, apply an activation function after
the addition. Defaults to `False`.
"""
def __init__(self, *args, average=False, activation=False, **kwargs):
super().__init__(*args, **kwargs)
self.average = average
self.activation = activation
def build(self, input_shape):
if not isinstance(input_shape, (list, tuple)) or len(input_shape) != 2:
raise ValueError(f"{self.__class__.__name__} expects two input tensors")
super().build(input_shape)
def call(self, inputs):
a, b = inputs
output = tf.add(a, b)
if self.average:
output /= 2
if self.activation:
output = tf.nn.relu(output)
return output
def get_config(self):
config = super().get_config()
config["average"] = self.average
config["activation"] = self.activation
return config
# Note that keras.layers.Add is not registered as a quantized target because the class name 'Add' is
# enough and shared between Keras and QuantizeML
[docs]
@register_quantize_target(Add)
@register_aligned_inputs
@keras.saving.register_keras_serializable()
class QuantizedAdd(QuantizedLayer, Add):
"""Sums two inputs and quantize the output.
The two inputs must be provided as a list or tuple of FixedPoint or Tensors.
The outputs are quantized according to the specified quantization configuration.
Args:
quant_config (dict, optional): the serialized quantization configuration. Defaults to None.
"""
def __init__(self, *args, quant_config=None, **kwargs):
super().__init__(*args, quant_config=quant_config, **kwargs)
self.buffer_bitwidth = apply_buffer_bitwidth(self.quant_config, signed=True)
out_quant_cfg = self.quant_config.get("output_quantizer", False)
if out_quant_cfg:
self.out_quantizer = OutputQuantizer(
name="output_quantizer", **out_quant_cfg)
else:
self.out_quantizer = None
# Add objects that will store the shift values.
self.a_shift = TensorRecorder(self.name + "/a_shift")
self.b_shift = TensorRecorder(self.name + "/b_shift")
def _quantized_relu(self, inputs):
# Inputs are always fixed points (as they come after an Add or after an output quantizer)
# we can then represent the zero as a fixed point
zero = FixedPoint(tf.constant(0.), inputs.value_bits, inputs.frac_bits)
# ReLU inside the Add is always unbounded so we just remove negative values
return tf.math.maximum(inputs, zero)
def call(self, inputs):
a, b = inputs
if not (isinstance(a, FixedPoint) and isinstance(b, FixedPoint)):
# If any of the inputs is not a FixedPoint, raise an error
raise TypeError(f"QuantizedAdd only accepts FixedPoint\
inputs. Receives {(type(a), type(b))} inputs.")
# Align intermediate inputs before adding them
a, shift_ab = a.align(b, self.buffer_bitwidth)
b, shift_ba = b.align(a, self.buffer_bitwidth)
outputs = tf.add(a, b)
if self.average:
outputs = outputs >> 1
# Compute shifts
self.a_shift(shift_ab)
self.b_shift(shift_ba)
# Rescale outputs
if self.out_quantizer is not None:
outputs = self.out_quantizer(outputs)
# Apply activation
if self.activation:
outputs = self._quantized_relu(outputs)
return outputs
[docs]
@register_quantize_target(keras.layers.Concatenate)
@register_no_output_quantizer
@register_aligned_inputs
@keras.saving.register_keras_serializable()
class QuantizedConcatenate(QuantizedLayer, keras.layers.Concatenate):
""" A Concatenate layer that operates on quantized inputs
"""
def call(self, inputs):
a, b = inputs
if not (isinstance(a, FixedPoint) and isinstance(b, FixedPoint)):
# If any of the inputs is not a FixedPoint, raise an error
raise TypeError(f"QuantizedConcatenate only accepts FixedPoint\
inputs. Receives {(type(a), type(b))} inputs.")
return tf.concat([a, b], axis=self.axis)