# Copyright (c) 2023, Apple Inc. All rights reserved.
#
# Use of this source code is governed by a BSD-3-clause license that can be
# found in the LICENSE.txt file or at https://opensource.org/licenses/BSD-3-Clause
import logging as _logging
from collections import OrderedDict as _OrderedDict
from enum import Enum as _Enum
from enum import unique as _unique
from typing import Any as _Any
from typing import Callable as _Callable
from typing import Dict as _Dict
from typing import List as _List
from typing import NewType as _NewType
from typing import Optional as _Optional
from typing import Union as _Union
import cattrs as _cattrs
import torch as _torch
import torch.ao.quantization as _aoquant
from attr import define as _define
from attr import field as _field
from attrs import validators as _validators
from coremltools.optimize.torch._utils.torch_utils import (
maybe_convert_str_to_dtype as _maybe_convert_str_to_dtype,
)
from coremltools.optimize.torch._utils.torch_utils import (
maybe_convert_str_to_mod_type as _maybe_convert_str_to_mod_type,
)
from coremltools.optimize.torch.optimization_config import (
ModuleOptimizationConfig as _ModuleOptimizationConfig,
)
from coremltools.optimize.torch.optimization_config import OptimizationConfig as _OptimizationConfig
from coremltools.optimize.torch.optimization_config import _structure_from_dict_hook_factory
_logger = _logging.getLogger(__name__)
[docs]@_unique
class ObserverType(_Enum):
"""
An enum indicating the type of observer. Allowed options are moving_average_min_max and mix_max.
"""
moving_average_min_max = "moving_average_min_max"
mix_max = "min_max"
@staticmethod
def get_observer(observer_type: "ObserverType", is_per_channel: bool) -> _Any:
_str_to_observer_map = {
"moving_average_min_max": _aoquant.MovingAverageMinMaxObserver,
"min_max": _aoquant.MinMaxObserver,
"moving_average_min_max_per_channel": _aoquant.MovingAveragePerChannelMinMaxObserver,
"min_max_per_channel": _aoquant.PerChannelMinMaxObserver,
}
observer_name = observer_type.value
if is_per_channel:
observer_name = f"{observer_name}_per_channel"
return _str_to_observer_map[observer_name]
[docs]@_unique
class QuantizationScheme(_Enum):
"""
An enum indicating the type of quantization to be performed. Allowed options are symmetric
and affine.
"""
symmetric = "symmetric"
affine = "affine"
@staticmethod
def get_qscheme(
quantizaton_scheme: "QuantizationScheme", is_per_channel: bool
) -> _torch.qscheme:
_str_to_qscheme_map = {
"symmetric": _torch.per_tensor_symmetric,
"affine": _torch.per_tensor_affine,
"symmetric_per_channel": _torch.per_channel_symmetric,
"affine_per_channel": _torch.per_channel_affine,
}
quantization_scheme_name = quantizaton_scheme.value
if is_per_channel:
quantization_scheme_name = f"{quantization_scheme_name}_per_channel"
return _str_to_qscheme_map[quantization_scheme_name]
_default_quantization_options = {
"weight_dtype": _torch.qint8,
"weight_per_channel": True,
"activation_dtype": _torch.quint8,
"observer": ObserverType.moving_average_min_max,
"quantization_scheme": QuantizationScheme.symmetric,
}
[docs]@_define
class ModuleLinearQuantizerConfig(_ModuleOptimizationConfig):
"""
Configuration class for specifying global and module level quantization options for linear quantization
algorithm implemented in :py:class:`LinearQuantizer`.
Linear quantization algorithm simulates the effects of quantization during training, by quantizing
and dequantizing the weights and/or activations during the model's forward pass. The forward and
backward pass computations are conducted in ``float`` dtype, however, these ``float`` values follow
the constraints imposed by ``int8`` and ``quint8`` dtypes. For more details, please refer to
`Quantization and Training of Neural Networks for Efficient Integer-Arithmetic-Only Inference
<https://arxiv.org/pdf/1712.05877.pdf>`_.
For most applications, the only parameters that need to be set are ``quantization_scheme`` and
``milestones``.
By default, ``quantization_scheme`` is set to :py:class:`QuantizationScheme.symmetric`, which means
all weights are quantized with zero point as zero, and activations are quantized with zero point as zero for
non-negative activations and 128 for all other activations. The weights are quantized using :py:class:`torch.qint8`
and activations are quantized using :py:class:`torch.quint8`.
Linear quantization algorithm inserts ``observers`` for each weight/activation tensor.
These observers collect statistics of these tensors' values, for example, the minimum and maximum values they can
take. These statistics are then used to compute the scale and zero point, which are in turn used for quantizing
the weights/activations. By default, ``moving_average_min_max`` observer is used. For more details, please
check `MinMaxObserver <https://pytorch.org/docs/stable/generated/torch.ao.quantization.observer.MinMaxObserver.html#torch.ao.quantization.observer.MinMaxObserver>`_.
The ``milestones`` parameter controls the flow of the quantization algorithm. The example below illustrates its
usage in more detail:
.. code-block:: python
model = define_model()
config = LinearQuantizerConfig(
global_config=ModuleLinearQuantizerConfig(
quantization_scheme="symmetric",
milestones=[0, 100, 300, 200],
)
)
quantizer = LinearQuantizer(model, config)
# prepare the model to insert FakeQuantize layers for QAT
model = quantizer.prepare()
# use quantizer in your PyTorch training loop
for inputs, labels in data:
output = model(inputs)
loss = loss_fn(output, labels)
loss.backward()
optimizer.step()
quantizer.step()
# In this example, from step 0 onwards, observers will collect statistics
# of the values of weights/activations. However, between steps 0 and 100,
# effects of quantization will not be simulated. At step 100, quantization
# simulation will begin and at step 300, observer statistics collection will
# stop. A batch norm layer computes mean and variance of input batch for normalizing
# it during training, and collects running estimates of its computed mean and variance,
# which are then used for normalization during evaluation. At step 200, batch norm
# statistics collection is frozen, and the batch norm layers switch to evaluation
# mode, thus more closely simulating the inference numerics during training time.
Args:
weight_dtype (:py:class:`torch.dtype`): The dtype to use for quantizing the weights. When dtype
is set to :py:class:`torch.float32`, the weights corresponding to that layer are not quantized.
Defaults to :py:class:`torch.qint8`.
weight_observer (:py:class:`ObserverType`): Type of observer to use for quantizing weights. Defaults
to ``moving_average_min_max``.
weight_per_channel (:obj:`bool`): When ``True``, weights are quantized per channel; otherwise, per tensor.
activation_dtype (:py:class:`torch.dtype`): The dtype to use for quantizing the activations. When dtype
is set to :py:class:`torch.float32`, the activations corresponding to that layer are not quantized.
Defaults to :py:class:`torch.quint8`.
activation_observer (:py:class:`ObserverType`): Type of observer to use for quantizing activations. Allowed
values are ``min_max`` and ``moving_average_min_max``. Defaults to ``moving_average_min_max``.
quantization_scheme: (:py:class:`QuantizationScheme`): Type of quantization configuration to use. When
this parameter is set to :py:class:`QuantizationScheme.symmetric`, all weights are
quantized with zero point as zero, and activations are quantized with zero point as zero for
non-negative activations and 128 for all other activations. When it is set to
:py:class:`QuantizationScheme.affine`, zero point can be set anywhere in the range of values allowed
for the quantized weight/activation. Defaults to :py:class:`QuantizationScheme.symmetric`.
milestones (:obj:`list` of :obj:`int`): A list of four integers indicating milestones to use during
quantization. The first milestone corresponds to enabling observers, the second to enabling fake
quantization simulation, the third to disabling observers, and the last to freezing batch norm statistics.
Defaults to ``None``, which means the ``step`` method of :py:class:`LinearQuantizer` will be a no-op and
all observers and quantization simulation will be turned on from the first step, batch norm layers always
operate in training mode, and mean and variance statistics collection is not frozen.
"""
weight_dtype: _torch.dtype = _field(
default=_default_quantization_options["weight_dtype"],
converter=_maybe_convert_str_to_dtype,
validator=[
_validators.instance_of(_torch.dtype),
_validators.in_([_torch.qint8, _torch.quint8, _torch.float32]),
],
)
weight_observer: ObserverType = _field(
default=_default_quantization_options["observer"],
converter=ObserverType,
validator=_validators.in_(ObserverType),
)
weight_per_channel: bool = _field(
default=_default_quantization_options["weight_per_channel"],
validator=_validators.instance_of(bool),
)
activation_dtype: _torch.dtype = _field(
default=_default_quantization_options["activation_dtype"],
converter=_maybe_convert_str_to_dtype,
validator=[
_validators.instance_of(_torch.dtype),
_validators.in_([_torch.quint8, _torch.float32]),
],
)
activation_observer: ObserverType = _field(
default=_default_quantization_options["observer"],
converter=ObserverType,
validator=_validators.in_(ObserverType),
)
quantization_scheme: QuantizationScheme = _field(
default=_default_quantization_options["quantization_scheme"],
converter=QuantizationScheme,
validator=_validators.in_(QuantizationScheme),
)
milestones: _Optional[_List[int]] = _field(
default=None,
validator=_validators.optional(
_validators.deep_iterable(
member_validator=_validators.instance_of(int),
iterable_validator=_validators.instance_of(list),
)
),
)
def __attrs_post_init__(self):
if self.weight_dtype == _torch.float32 and self.activation_dtype != _torch.float32:
raise ValueError(
f"Unsupported configuration: weight_dtype = {self.weight_dtype}, "
f"activation_dtype = {self.activation_dtype}. When weights are not quantized,"
f"activations cannot be quantized."
)
@milestones.validator
def _check_milestones(self, attribute, value):
if value is not None:
assert len(value) == 4, (
f"Received milestones = {value}. "
f"Milestones should be of length 4. "
f"Refer to docs for more information."
)
_ModuleTypeConfigType = _NewType(
"ModuleTypeConfigType",
_Dict[_Union[_Callable, str], _Optional[ModuleLinearQuantizerConfig]],
)
[docs]@_define
class LinearQuantizerConfig(_OptimizationConfig):
"""
Configuration class for specifying how different submodules of a model are
quantized by :py:class:`LinearQuantizer`.
In order to disable quantizing a layer or an operation, ``module_type_config`` or
``module_name_config`` corresponding to that operation can be set to ``None``.
For example:
.. code-block:: python
# The following config will enable weight only quantization for all layers:
config = LinearQuantizerConfig.from_dict(
{
"global_config": {
"activation_dtype": "float32",
}
}
)
# The following config will disable quantization for all linear layers and
# set quantization mode to weight only quantization for convolution layers:
config = LinearQuantizerConfig.from_dict(
{
"module_type_configs": {
"Linear": None,
"Conv2d": {
"activation_dtype": "float32",
},
}
}
)
# The following config will disable quantization for layers named conv1 and conv2:
config = LinearQuantizerConfig.from_dict(
{
"module_name_configs": {
"conv1": None,
"conv2": None,
}
}
)
Args:
global_config (:py:class:`ModuleLinearQuantizerConfig`): Config to be applied globally
to all supported modules. Missing values are chosen from the default config.
module_type_configs (:obj:`dict` of :obj:`str` to :py:class:`ModuleLinearQuantizerConfig`):
Module type level configs applied to a specific
module class, such as :py:class:`torch.nn.Linear`. The keys can be either strings
or module classes.
module_name_configs (:obj:`dict` of :obj:`str` to :py:class:`ModuleLinearQuantizerConfig`):
Module level configs applied to specific modules.
The name of the module must be a fully qualified name that can be used to fetch it
from the top level module using the ``module.get_submodule(target)`` method.
non_traceable_module_names (:obj:`list` of :obj:`str`):
Names of modules which cannot be traced using ``torch.fx``.
.. note::
The ``quantization_scheme`` parameter must be the same across all configs.
"""
global_config: _Optional[ModuleLinearQuantizerConfig] = _field(
default=None,
validator=_validators.optional(_validators.instance_of(ModuleLinearQuantizerConfig)),
)
module_type_configs: _ModuleTypeConfigType = _field(
factory=_OrderedDict,
validator=_validators.deep_mapping(
key_validator=_validators.instance_of((str, _Callable)),
value_validator=_validators.optional(
_validators.instance_of(ModuleLinearQuantizerConfig)
),
mapping_validator=_validators.instance_of(dict),
),
)
module_name_configs: _Dict[str, _Optional[ModuleLinearQuantizerConfig]] = _field(
factory=_OrderedDict,
validator=_validators.deep_mapping(
key_validator=_validators.instance_of(str),
value_validator=_validators.optional(
_validators.instance_of(ModuleLinearQuantizerConfig)
),
mapping_validator=_validators.instance_of(dict),
),
)
non_traceable_module_names: _List[str] = _field(
default=list(),
validator=_validators.deep_iterable(
member_validator=_validators.instance_of(str),
),
)
def __attrs_post_init__(self):
if (
self.global_config is None
and len(self.module_type_configs) == 0
and len(self.module_name_configs) == 0
):
self.global_config = ModuleLinearQuantizerConfig()
self.module_type_configs = {
_maybe_convert_str_to_mod_type(key): val
for key, val in self.module_type_configs.items()
}
self._validate_same_params(["quantization_scheme"])
[docs] @classmethod
def from_dict(cls, config_dict: _Dict[str, _Any]) -> "LinearQuantizerConfig":
super().from_dict(config_dict)
converter = _cattrs.Converter(forbid_extra_keys=True)
converter.register_structure_hook(
_ModuleTypeConfigType,
_structure_from_dict_hook_factory(ModuleLinearQuantizerConfig),
)
return converter.structure_attrs_fromdict(config_dict, cls)