#!/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__ = ["quantize", "dump_config"]
import warnings
from collections import defaultdict
import numpy as np
import keras
from onnx import ModelProto
try:
# Legacy support for keras 2.10
from keras.utils.generic_utils import get_registered_object, get_registered_name
except ImportError:
from keras.saving.object_registration import get_registered_object, get_registered_name
from .utils import apply_weights_to_model
from .transforms import sanitize
from .transforms.insert_layer import insert_in_config
from .transforms.transforms_utils import get_inbound_layers_config
from .calibrate import calibrate
from ..layers import (OutputQuantizer, WeightQuantizer, Dequantizer, quantization,
QuantizationParams, get_quantization_params, Attention, QuantizedConv2D,
QuantizedDepthwiseConv2D, QuantizedSeparableConv2D, QuantizedDense,
StatefulRecurrent, QuantizedActivation)
from ..layers.layers_base import (_GLOBAL_LAYER_TO_QLAYER, _GLOBAL_NO_OUTPUT_QUANTIZER,
_GLOBAL_ALIGNED_INPUTS)
from ..onnx_support.quantization.quantize import quantize as quantize_onnx
# List of Quantizer layer's that do not have a float layer representation
NO_FLOAT_CUSTOM_QLAYERS = [Dequantizer, OutputQuantizer, WeightQuantizer]
def get_quantized_layer(layer):
""" Returns the quantized version of the layer.
Args:
layer (keras.layers.Layer or dict): layer of interest
Returns:
keras.layer: quantized version of the layer if it exists, None otherwise.
"""
if isinstance(layer, keras.layers.Layer):
config = layer.get_config()
layer_class = layer.__class__
else:
config = layer['config']
layer_class = get_layer_class_from_name(layer["class_name"])
qlayer_class = _GLOBAL_LAYER_TO_QLAYER.get(layer_class.__name__, None)
# Special case for activations: avoid quantization if activation is not within the
# allowed activations
allowed_activations = QuantizedActivation.arg_constraints['activation']()
if qlayer_class == QuantizedActivation and config['activation'] not in allowed_activations:
return None
return qlayer_class
def is_quantized_layer(layer):
""" Returns True when the layer is a quantized layer.
Args:
layer (keras.layers.Layer or type): layer of interest
Returns:
bool: True when the layer is a quantized layer, False otherwise.
"""
if isinstance(layer, type):
return layer in _GLOBAL_LAYER_TO_QLAYER.values()
return layer.__class__ in _GLOBAL_LAYER_TO_QLAYER.values()
def get_layer_class_from_name(name):
""" Returns the class object of a registered keras layer.
Args:
name (str): the layer class name
Returns:
type: the class of the layer
"""
if hasattr(keras.layers, name):
return getattr(keras.layers, name)
return get_registered_object(name)
def _handle_not_quantizable_layers(model, config, skip_warning=False):
""" Includes a number of dequantizers such that the model is compatible.
Args:
model (keras.Model): model structure to check
config (dict): config where Dequantizer(s) will be placed
skip_warning (bool_optional): whether to skip warning provided by partial quantization.
Defaults to False.
Returns:
bool: whether the model was fully quantized.
"""
# Find where to insert the Dequantizer(s):
# A dequantizer will be added for all links that connect a quantized layer to a floating one.
num_no_quantizable_layers = 0
dequantizer_inbounds = defaultdict(list)
for layer in config['layers']:
layer_class = get_layer_class_from_name(layer['class_name'])
if layer_class == Dequantizer:
# If a model has a Dequantizer is because it is requantizing.
# In that case nothing has to be done (model already has the necessary dequantizers)
return False
elif not (is_quantized_layer(layer_class) or layer_class in NO_FLOAT_CUSTOM_QLAYERS):
# The current layer is float.
for inbound_layer in get_inbound_layers_config(layer, config):
inbound_layer_class = get_layer_class_from_name(inbound_layer['class_name'])
# A dequantizer will be added for each quantized inbound layer
if is_quantized_layer(inbound_layer_class):
inbound_layer_name = inbound_layer['config']['name']
if len(dequantizer_inbounds) == 0 and not skip_warning:
# Print warning for the first non-quantizable layer
warnings.warn(f"'{layer['config']['name']}' of type {layer['class_name']} "
"is not supported to quantize, a Dequantizer is added "
"before it and quantization will stop at this layer.")
dequantizer_inbounds[inbound_layer_name].append(layer['config']['name'])
num_no_quantizable_layers += 1
if len(config['layers']) == num_no_quantizable_layers:
raise RuntimeError(f"Impossible to quantize '{model.name}'. "
"At least one layer should be quantizable.")
if len(dequantizer_inbounds) == 0:
# Model was completely quantized.
return True
# Insert a Dequantizer on each target_layer -> outbound link
for target_layer_name, outbound_names in dequantizer_inbounds.items():
dequantizer = Dequantizer(name=f'{target_layer_name}/dequantizer')
insert_in_config(model, target_layer_name, dequantizer, config, outbound_names)
return False
def _prepare_output_quantizers(model):
""" Parse the model and prepare OutputQuantizer configurations for layers requiring them.
To ensure that an OutputQuantizer will be added to the latest possible layer in a 'block', the
model is parsed in reverse order. If a layer requires aligned inputs, the function will find the
preceding layer that can accept an OutputQuantizer and set it in the returned dictionary.
Args:
model (keras.Model): the model to parse
Returns:
dict: dictionary mapping layer names to an OutputQuantizer config.
"""
# Dictionary that will contain layers and their OutputQuantizer configurations
out_quantizer_configs = {}
# Get quantization parameters
qparams = get_quantization_params()
def set_output_quantizer(layer_names, next_layer):
""" Populates `out_quantizer_configs` with layer names and their OutputQuantizer. """
for name in layer_names:
current_layer = model.get_layer(name)
# Handle special cases where the OutputQuantizer must be per-tensor:
# - when current_layer has vector outputs,
# - when next_layer is an Attention layer and the layer is Query or Key
# (first and second inputs)
# - when the layer is a StatefulRecurrent layer
if isinstance(current_layer, Attention):
output_shape = current_layer.output_shape[0]
else:
output_shape = current_layer.output_shape
# remove batch_size dim
output_shape = output_shape[1:]
vector_outputs = np.prod(output_shape) == output_shape[-1]
query_or_key = (isinstance(current_layer, keras.layers.Dense)
and isinstance(next_layer, Attention)
and next_layer.inbound_nodes[0].inbound_layers.
index(current_layer) in [0, 1])
is_stateful_rec = isinstance(current_layer, StatefulRecurrent)
is_next_activation = get_quantized_layer(next_layer) == QuantizedActivation
per_tensor = (query_or_key or vector_outputs or is_stateful_rec or
is_next_activation or qparams.per_tensor_activations)
# If this is a new entry, set a default configuration
if name not in out_quantizer_configs:
axis = "per-tensor" if per_tensor else "per-axis"
if isinstance(current_layer, keras.layers.ReLU):
params = dict(bitwidth=qparams.activation_bits,
signed=qparams.activation_bits >= 8,
axis=axis)
elif isinstance(current_layer, keras.layers.Activation):
params = dict(bitwidth=qparams.activation_bits, axis=axis)
elif is_next_activation:
params = dict(bitwidth=QuantizedActivation.DEFAULT_INPUT_BITWIDTH, axis=axis)
else:
# StatefulRecurrent special: previous and self OutputQuantizer should be 16-bits
if is_stateful_rec or isinstance(next_layer, StatefulRecurrent):
bitwidth = 16
else:
bitwidth = qparams.output_bits
params = dict(bitwidth=bitwidth, axis=axis)
params['buffer_bitwidth'] = qparams.buffer_bits
out_quantizer_configs[name] = dict(output_quantizer=params)
# If the layer OutputQuantizer configuration is already set, simply check the axis:
# override the config if the outputs must be per-tensor
else:
current_axis = out_quantizer_configs[name]["output_quantizer"]["axis"]
per_tensor = per_tensor or current_axis == "per-tensor"
axis = "per-tensor" if per_tensor else "per-axis"
out_quantizer_configs[name]["output_quantizer"]["axis"] = axis
def cannot_have_output_quantizer(layer):
""" Returns True when the layer cannot have an OutputQuantizer. """
qlayer = get_quantized_layer(layer)
return (isinstance(layer, Dequantizer)
or qlayer is None
or qlayer in _GLOBAL_NO_OUTPUT_QUANTIZER)
def get_preceding_layer_names(layer):
""" Retrieve inbounds layers names where an OutputQuantizer can be set. """
previous_layers = []
inbounds = layer.inbound_nodes[0].inbound_layers
if not isinstance(inbounds, list):
inbounds = [inbounds]
for inbound in inbounds:
# Skip input layers
if isinstance(inbound, keras.layers.InputLayer):
continue
# When the given layer cannot have an OutputQuantizer, recursively call the function on
# this layer
if cannot_have_output_quantizer(inbound):
previous_layers.extend(get_preceding_layer_names(inbound))
else:
previous_layers.append(inbound.name)
return previous_layers
# Parse the layers in reverse order
for layer in model.layers[::-1]:
# Find layers that will need aligned inputs
if get_quantized_layer(layer) in _GLOBAL_ALIGNED_INPUTS:
# Retrieve the inbounds that can have an OutputQuantizer
previous_layers = get_preceding_layer_names(layer)
# Set an OutputQuantizer in their inbounds
set_output_quantizer(previous_layers, layer)
return out_quantizer_configs
def quantize_keras(model, q_config=None, qparams=QuantizationParams(), samples=None,
num_samples=1024, batch_size=None, epochs=1, quantize_until=None):
"""Quantizes a Keras model using the provided configuration or parameters.
Details on how this function behaves:
- `q_config` has priority over `qparams`, meaning that when a match is found in `q_config` the
given configuration will be used instead of `qparams`. This is useful to handle specific cases
(e.g per-tensor output quantizer).
- when no configuration is given, quantization parameters are deduced from `qparams` and
OutputQuantizers are automatically set on appropriate layers.
- `qparams` are only applied to 'float' Keras layers when they are first quantized. As a result,
when re-quantizing a model, one must provide a complete `q_config`. This is made easy with the
`dump_config` helper.
If not already present, a final Dequantizer will be added at the end of the Model.
The model will also be calibrated using the provided (or randomly generated inputs).
Args:
model (keras.Model): the model to quantize
q_config (dict, optional): quantization configuration as a dictionary mapping layer names to
their quantization configuration. Defaults to None.
qparams (QuantizationParams, optional): global quantization parameters. Defaults to
QuantizationParams().
samples (tf.Dataset, np.array or generator, optional): calibration samples. When no samples
are provided, random samples are generated. Defaults to None.
num_samples (int, optional): number of samples to use in the provided samples or number of
samples to generate. Defaults to 1024.
batch_size (int, optional): the batch size. Defaults to None.
epochs (int, optional): the number of epochs. Defaults to 1.
quantize_until (str, optional): name of the layer until which to quantize:
other layers after it will stay unchanged. Defaults to None.
Returns:
keras.Model: the quantized model
"""
q_config = q_config or dict()
if quantize_until and not any(ly.name == quantize_until for ly in model.layers):
raise ValueError(f"'{quantize_until}' is not a recognized layer in {model.name}")
# Handle input_weight_bits using another QuantizationParams where
# weight_bits = qparams.input_weight_bits, it will be set to False once the input layer has been
# quantized.
input_qparams = QuantizationParams(activation_bits=qparams.activation_bits,
per_tensor_activations=qparams.per_tensor_activations,
weight_bits=qparams.input_weight_bits,
output_bits=qparams.output_bits,
buffer_bits=qparams.buffer_bits)
def get_quantize_layer(layer, quantize_config=None):
"""Get quantize config from float layer:
- first, we get its quantized version,
- then, we return the quantized layer with config updated
"""
# Check if qlayer exists in custom layers and returns the float version of the layer if not
l_class = get_layer_class_from_name(layer["class_name"])
ql_class = get_quantized_layer(layer)
if ql_class is None:
ql_class = l_class
# Initialize quantized layer from the float config
qlayer = layer
# Instantiate quantized layer from configuration if there is one
if quantize_config:
qlayer['config']['quant_config'] = quantize_config
# Set the preset default configuration otherwise
elif qlayer['config']['name'] in out_quantizer_configs:
qlayer['config']['quant_config'] = out_quantizer_configs[qlayer['config']['name']]
# Retrieve the quantize config after initializing the quantized layer, in order to
# configure the specific parameters given by the QuantizationParams context.
qlayer['config'] = ql_class.from_config(qlayer['config']).get_config()
# Update layer class name by the quantized one
qlayer['class_name'] = get_registered_name(ql_class)
return qlayer
# Sanitize the model and make it quantization ready
model = sanitize(model)
# Determine where to set OutputQuantizers, the return dict will be used as a non-local
# variable in the _replace_layer function.
with quantization(qparams):
out_quantizer_configs = _prepare_output_quantizers(model)
# Quantize the model, modifying each layer config by its respective quantized version
input_layers = (QuantizedConv2D, QuantizedDepthwiseConv2D, QuantizedSeparableConv2D,
QuantizedDense)
qmodel_config = model.get_config()
quantized_layers = set()
for idx, layer in enumerate(qmodel_config['layers']):
# If some layer is quantized then this is requantization.
# Raise exception for this case if quantize_until is provided
layer_class = get_layer_class_from_name(layer['class_name'])
if is_quantized_layer(layer_class) and quantize_until:
raise ValueError("'quantize_until' is not supported when requantizing.")
# Retrieve quantize config from layer
match_conf = q_config.get(layer['config']['name'], None)
# Overwrite quantization context with input_qparams (if they are not None)
with quantization(input_qparams or qparams):
inbound_layers = get_inbound_layers_config(layer, qmodel_config)
# Quantization is only performed if the inbound layers were quantized
if all(x['config']['name'] in quantized_layers for x in inbound_layers):
qlayer = get_quantize_layer(layer, match_conf)
else:
qlayer = layer
# When the qlayer is an input layer that has been quantized, disable input_qparams
qlayer_class = get_layer_class_from_name(qlayer['class_name'])
if input_qparams and qlayer_class in input_layers:
input_qparams = None
# Skip input layers
if qlayer_class == keras.layers.InputLayer:
# Although InputLayer is not quantizable, layer is treated as one
# so its outbounds can be quantized.
quantized_layers.add(qlayer['config']['name'])
continue
# If it was not possible to quantize the layer, try to quantize the next one.
# This ensures that as many layers as possible are quantized.
if not is_quantized_layer(qlayer_class):
continue
# Finally, update model with quantize layer config
# Note at this point, we know the layer was quantized successfully
qmodel_config['layers'][idx] = qlayer
if quantize_until != layer["config"]["name"]:
# If quantize_until is provided, layer is quantized but is not added to
# quantized_layers list, preventing the quantization of layers after it.
# Note if layer is within a branch, quantization will end only for this branch
quantized_layers.add(qlayer['config']['name'])
# Insert the number of Dequantizers necessary for the model to be compatible
is_full_quantized = _handle_not_quantizable_layers(model,
qmodel_config,
skip_warning=quantize_until is not None)
# Build the model and transfer weights
qmodel = model.from_config(qmodel_config)
apply_weights_to_model(qmodel, {var.name: var for var in model.variables}, False)
# Convert model into a functional one.
# Note if model was completely quantized, we add a last dequantizer to produce a float output
y = qmodel.output
if is_full_quantized:
y = Dequantizer()(y)
qmodel = keras.Model(qmodel.input, y, name=model.name)
# Now that the model is quantized, proceed to calibration
calibrate(model, qmodel, samples=samples, num_samples=num_samples, batch_size=batch_size,
epochs=epochs)
return qmodel
[docs]def quantize(model, q_config=None, qparams=QuantizationParams(), samples=None, num_samples=1024,
batch_size=None, epochs=1, quantize_until=None):
"""Quantizes a Keras or ONNX model using the provided configuration or parameters.
Details on how this function behaves:
- `q_config` has priority over `qparams`, meaning that when a match is found in `q_config` the
given configuration will be used instead of `qparams`. This is useful to handle specific cases
(e.g per-tensor output quantizer). This is only used when quantizing Keras models.
- when no configuration is given, quantization parameters are deduced from `qparams` and
OutputQuantizers are automatically set on appropriate layers.
- `qparams` are only applied to 'float' Keras layers when they are first quantized. As a result,
when re-quantizing a model, one must provide a complete `q_config`. This is made easy with the
`dump_config` helper. Note the only configuration supported when quantizing ONNX models is
8-bit for weights and activations, but per_tensor_activations param will be taken into
account.
If not already present, a final Dequantizer will be added at the end of the Model.
The model will also be calibrated using the provided (or randomly generated inputs).
Args:
model (keras.Model or ModelProto): the model to quantize
q_config (dict, optional): quantization configuration as a dictionary mapping layer names to
their quantization configuration. Defaults to None.
qparams (QuantizationParams, optional): global quantization parameters. Defaults to
QuantizationParams().
samples (tf.Dataset, np.array or generator, optional): calibration samples. When no samples
are provided, random samples are generated. Defaults to None.
num_samples (int, optional): number of samples to use in the provided samples or number of
samples to generate. Defaults to 1024.
batch_size (int, optional): the batch size. Defaults to None.
epochs (int, optional): the number of epochs. This parameter must be 1 for ONNX models.
Defaults to 1.
quantize_until (str, optional): name of the layer/node until which to quantize:
other ones after it will stay unchanged. Defaults to None.
Returns:
keras.Model or ModelProto: the quantized model
"""
# Calibration with random samples will only provide meaningful results when quantizing
# per-tensor
if samples is None and not qparams.per_tensor_activations:
warnings.warn("Quantizing per-axis with random calibration samples is not accurate.\
Set QuantizationParams.per_tensor_activations=True when calibrating with \
random samples.")
if type(model) != ModelProto:
return quantize_keras(model=model,
q_config=q_config,
qparams=qparams,
samples=samples,
num_samples=num_samples,
batch_size=batch_size,
epochs=epochs,
quantize_until=quantize_until)
elif q_config:
raise ValueError("unsupported parameter q_config for ONNX models quantization")
elif epochs != 1:
raise ValueError("unsupported parameter epochs != 1 for ONNX models quantization")
return quantize_onnx(model=model,
qparams=qparams,
samples=samples,
num_samples=num_samples,
batch_size=batch_size,
quantize_until=quantize_until)
[docs]def dump_config(model):
"""Dump the quantization configuration of a quantized model, exporting the configuration for
each quantized layer.
Args:
model (keras.Model): a quantized model.
Returns:
dict: the configuration of the model.
"""
# Get the configuration of the model, iterating over each layer and updating on config.
config = {}
for layer in model.layers:
# Try to take the current quantized configuration
ly_config = layer.get_config().get('quant_config')
# Only append quantized configuration
if is_quantized_layer(layer) and ly_config:
config[layer.name] = ly_config
return config