def test_log_areas(self):
logger.info(" ")
logger.info("Testing test_log_areas()")
# Test the UTILS logger
utils_logger = AimetLogger.get_area_logger(AimetLogger.LogAreas.Utils)
utils_logger.debug("Testing Debug")
utils_logger.info("Testing Info")
utils_logger.warning("Testing Warning")
utils_logger.error("Testing Error")
utils_logger.critical("Testing Critical")
utils_logger.critical("**************************************** \n")
# Test the QUANT logger
quant_logger = AimetLogger.get_area_logger(AimetLogger.LogAreas.Quant)
quant_logger.debug("Testing Debug")
quant_logger.info("Testing Info")
quant_logger.warning("Testing Warning")
quant_logger.error("Testing Error")
quant_logger.critical("Testing Critical")
quant_logger.critical("**************************************** \n")
# Test the SVD logger
svd_logger = AimetLogger.get_area_logger(AimetLogger.LogAreas.Svd)
svd_logger.debug("Testing Debug")
svd_logger.info("Testing Info")
svd_logger.warning("Testing Warning")
svd_logger.error("Testing Error")
svd_logger.critical("Testing Critical")
svd_logger.critical("**************************************** \n")
def __init__(self, graph, op_map=None, ops_to_ignore=None, strict=True):
"""
Constructor
:param graph: The graph to search
:param op_map: The map of operations used to identify op sequences as "one op".
The default op_map used is defined in op_deps.py. Please refer to
that format for passing a custom op_map.
:param ops_to_ignore: List of ops to ignore
:param strict: If strict mode is set to True queries will only return the last ops
at the end of well known "op layers" as defined by the op_map. When False,
queries will return ops at the end of well known layers and, in addition,
all ops which are not "known".
Eg If you have a list of ops in a graph like: Conv2D, BiasAdd, WeirdOp
Strict mode will return ["BiasAdd"] since it knows that Conv2D+BiasAdd are
one logical "layer". When strict mode is disabled it will return ["BiasAdd", "WeirdOp"]
:param debug: Whether to enable debug messages or not.
"""
self._log = AimetLogger.get_area_logger(AimetLogger.LogAreas.Utils)
self._graph = graph
self._strict = strict
if op_map:
self._op_map = op_map
else:
self._op_map = _OP_MAP
if ops_to_ignore:
self._ops_to_ignore = ops_to_ignore
else:
self._ops_to_ignore = []
self._trained_vars = graph.get_collection(
tf.compat.v1.GraphKeys.TRAINABLE_VARIABLES)
def test_setting_log_level_for_all_areas(self):
logger.info("*** test_setting_log_level_for_all_areas() ***\n")
svd_logger = AimetLogger.get_area_logger(AimetLogger.LogAreas.Svd)
quant_logger = AimetLogger.get_area_logger(AimetLogger.LogAreas.Quant)
util_logger = AimetLogger.get_area_logger(AimetLogger.LogAreas.Utils)
test_logger = AimetLogger.get_area_logger(AimetLogger.LogAreas.Test)
# The default logging level for all Log Areas defined in default_logging_config.json is used.
logger.info(
"Log at the default log level for all Log Areas defined in default_logging_config.json"
)
svd_logger.debug("Testing Debug")
svd_logger.info("Testing Info")
quant_logger.warning("Testing Warning")
quant_logger.error("Testing Error")
util_logger.critical("Testing Critical")
util_logger.info("Testing Info")
test_logger.critical("Testing Critical")
test_logger.critical("****************************************\n")
# Change the default log level for all areas
# Only CRITICAL level logs will be logged.
logger.info("Change the logging level for all Log Areas to WARNING")
AimetLogger.set_level_for_all_areas(logging.WARNING)
svd_logger.debug("Testing Debug")
svd_logger.info("Testing Info")
quant_logger.warning("Testing Warning")
quant_logger.error("Testing Error")
util_logger.critical("Testing Critical")
util_logger.info("Testing Info")
test_logger.critical("Testing Critical")
test_logger.critical("****************************************\n")
def __init__(self,
model=None,
quant_mode='tf_enhanced',
round_mode='nearest',
use_cuda=True):
"""
:param model: The input model to add quantization ops to
:param quant_mode: Indicates which quantization algorithm should be used, either
'tf' or 'tf_enhanced'. Defaults to 'tf_enhanced'.
:param round_mode: The round scheme to used. One of: 'nearest' or 'stochastic'. Default
is 'nearest'.
:param use_cuda: Indicates on which hardware which the quantization algorithm should run. Currently
defaults to GPU (True).
:raises: ValueError: An error occurred processing one of the input parameters.
"""
if quant_mode not in ('tf_enhanced', 'tf'):
raise ValueError(
'Parameter quantization mode is not a valid selection. Valid selections are tf, '
'tf_enhanced')
if round_mode not in ('nearest', 'stochastic'):
raise ValueError(
'Parameter round mode is not a valid selection. Valid selections are nearest or '
'stochastic')
self._model = model
self._use_cuda = use_cuda
if isinstance(quant_mode, str):
if quant_mode == 'tf':
quant_mode = QuantScheme.post_training_tf
elif quant_mode == 'tf_enhanced':
quant_mode = QuantScheme.post_training_tf_enhanced
self._quant_mode = quant_mode
self._round_mode = round_mode
self._logger = AimetLogger.get_area_logger(AimetLogger.LogAreas.Quant)
def test_setting_log_level(self):
logger.info("*** Testing test_setting_log_level() *** \n")
svd_logger = AimetLogger.get_area_logger(AimetLogger.LogAreas.Svd)
# The default logging level for SVD defined in default_logging_config.json is used.
logger.info(
"Log at the default log level for SVD defined in default_logging_config.json"
)
svd_logger.debug("Testing Debug")
svd_logger.info("Testing Info")
svd_logger.warning("Testing Warning")
svd_logger.error("Testing Error")
svd_logger.critical("Testing Critical")
svd_logger.critical("****************************************\n")
# Change the default log level for SVD.
# Only CRITICAL level logs will be logged.
logger.info("Change SVD area's logging level to Critical")
AimetLogger.set_area_logger_level(AimetLogger.LogAreas.Svd,
logging.CRITICAL)
svd_logger.debug("Testing Debug")
svd_logger.info("Testing Info")
svd_logger.warning("Testing Warning")
svd_logger.error("Testing Error")
svd_logger.critical("Testing Critical")
svd_logger.critical("****************************************\n")
# Change the default log level for SVD.
# All logs will be logged.
logger.info("Change SVD area's logging level to Critical")
AimetLogger.set_area_logger_level(AimetLogger.LogAreas.Svd,
logging.DEBUG)
svd_logger.debug("Testing Debug")
svd_logger.info("Testing Info")
svd_logger.warning("Testing Warning")
svd_logger.error("Testing Error")
svd_logger.critical("Testing Critical")
svd_logger.critical("****************************************\n")
# CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
# ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
# POSSIBILITY OF SUCH DAMAGE.
#
# SPDX-License-Identifier: BSD-3-Clause
#
# @@-COPYRIGHT-END-@@
# =============================================================================
"""Utility functions for test/train data cache implementation"""
import os
import shutil
from aimet_common.utils import AimetLogger
logger = AimetLogger.get_area_logger(AimetLogger.LogAreas.Utils)
def is_cache_env_set():
"""
check if Cache Environment variable is set or not
:param: None
:return: TRUE in case DEPENDENCY_DATA_PATH environment variable is set, False otherwise
"""
return 'DEPENDENCY_DATA_PATH' in os.environ
def is_mnist_cache_present():
"""
check if MNIST data is present in the cache
import libpymo
from aimet_common.graph_pattern_matcher import PatternType
from aimet_common.graph_searcher import GraphSearcher
from aimet_torch import utils
from aimet_torch import quantsim as qsim
from aimet_torch.meta.connectedgraph import ConnectedGraph
from aimet_torch.quantsim import QcQuantizeWrapper
from aimet_torch.save_utils import SaveUtils
from aimet_common.utils import AimetLogger
from aimet_common.bias_correction import ConvBnInfoType, ConvBnPatternHandler
from aimet_common.defs import ActivationType
from aimet_torch.utils import get_ordered_lists_of_conv_fc
logger = AimetLogger.get_area_logger(AimetLogger.LogAreas.Quant)
class StopForwardException(Exception):
""" Dummy exception to early-terminate forward-pass """
def forward_pass(model: torch.nn.Module, batch: torch.Tensor):
"""
forward pass depending model allocation on CPU / GPU till StopForwardException
:param model: model
:param batch: batch
:return: Nothing
"""
model.eval()
# first check if the model is on GPU or not
#
# SPDX-License-Identifier: BSD-3-Clause
#
# @@-COPYRIGHT-END-@@
# =============================================================================
# TODO Need to exclude this file for PyLint checking. We get the following error that needs to be investigated:
# RecursionError: maximum recursion depth exceeded while calling a Python object
# pylint: skip-file
""" Sub-sample data for weight reconstruction for channel pruning feature """
import numpy as np
from aimet_common.utils import AimetLogger
logger = AimetLogger.get_area_logger(AimetLogger.LogAreas.Test)
class InputMatchSearch:
""" Utilities to find a set of input pixels corresponding to an output pixel for weight reconstruction """
@staticmethod
def _check_and_update_pixel_sampled_from_output_data(
input_data_shape: tuple, layer_attributes: tuple, pixel: tuple):
"""
Function gets shape, layer, pixel indices (height, width) and first check if given height and width satisfy
O = (I - F + 2P) / S + 1 formula, and then update them using stride and
padding values of given layer.
:param input_data_shape: input data shape (Cin, Hin, Win)
:param layer_attributes: (kernel_size, stride, padding)
:param pixel: (height, width)
from typing import Tuple, List, Union, Dict
import numpy as np
import tensorflow as tf
import libpymo
from aimet_tensorflow.common.connectedgraph import ConnectedGraph
from aimet_tensorflow.common.operation import Op
from aimet_tensorflow.batch_norm_fold import fold_all_batch_norms
from aimet_tensorflow.utils.graph_saver import save_and_load_graph
from aimet_tensorflow.utils.op.conv import WeightTensorUtils, BiasUtils
import aimet_tensorflow.utils.op.relu as ReluUtils
from aimet_tensorflow.utils.op.fusedbatchnorm import BNUtils
from aimet_common.utils import AimetLogger
logger = AimetLogger.get_area_logger(AimetLogger.LogAreas.CrosslayerEqualization)
ScaleFactor = Union[np.ndarray, Tuple[np.ndarray]]
ClsSet = Union[Tuple[tf.Operation, tf.Operation],
Tuple[tf.Operation, tf.Operation, tf.Operation]]
class GraphSearchUtils:
""" Implements graph search utils required by CLE feature"""
def __init__(self, model: tf.Graph, start_op_names: Union[str, List[str]], output_op_names: Union[str, List[str]]):
if isinstance(start_op_names, str):
start_op_names = [start_op_names]
# @@-COPYRIGHT-END-@@
#
# =============================================================================
""" Winnow the API provided input channels from the modules in a model. """
import copy
from typing import List, Tuple, Dict
import torch
from aimet_common.utils import AimetLogger, ModelApi
from aimet_common.winnow.mask_propagation_winnower import MaskPropagationWinnower as AimetCommonMaskPropagationWinnower
from aimet_common.winnow.mask_propagator import MaskPropagator
from aimet_torch.meta.connectedgraph import ConnectedGraph
from aimet_torch.utils import get_layer_name, has_hooks
from aimet_torch.winnow.module_reducer import ModuleReducer
logger = AimetLogger.get_area_logger(AimetLogger.LogAreas.Winnow)
class MaskPropagationWinnower(AimetCommonMaskPropagationWinnower):
""" The MaskPropagationWinnower class implements winnowing based on propagating masks corresponding to each
module's input channels identified to be winnowed. """
def __init__(self,
model: torch.nn.Module,
input_shape: Tuple,
list_of_modules_to_winnow: List[Tuple[torch.nn.Module,
List]] = None,
reshape=True,
in_place=False,
verbose=False):
"""
MaskPropagationWinnower object initialization.
# POSSIBILITY OF SUCH DAMAGE.
#
# SPDX-License-Identifier: BSD-3-Clause
#
# @@-COPYRIGHT-END-@@
# =============================================================================
""" Utilities for working with ModelOptimization C++ library """
from typing import List
import torch
import libpymo as pymo
from aimet_torch.layer_database import Layer
from aimet_common.defs import CostMetric
from aimet_common.utils import AimetLogger
logger = AimetLogger.get_area_logger(AimetLogger.LogAreas.Svd)
class PymoSvdUtils:
""" Utilities for working with SVD ModelOptimization C++ library """
@classmethod
def configure_layers_in_pymo_svd(cls, layers: List[Layer],
cost_metric: CostMetric, svd_lib_ref):
"""
Configure layers with the pymo svd library
:param layers: List of layers to configure
:param cost_metric: Cost metric to use
:param svd_lib_ref: Reference to pymo instance
:return:
"""
import os
import libpymo as pymo
from aimet_common.bokeh_plots import DataTable
from aimet_common.bokeh_plots import LinePlot
from aimet_common.bokeh_plots import ProgressBar
from aimet_common.utils import AimetLogger
from aimet_common.curve_fit import MonotonicIncreasingCurveFit
from aimet_common.defs import CostMetric, LayerCompRatioPair, GreedyCompressionRatioSelectionStats, \
TarCompressionRatioSelectionStats, LayerCompRatioEvalScore, EvalFunction
from aimet_common.pruner import Pruner
from aimet_common import cost_calculator as cc
from aimet_common.layer_database import Layer, LayerDatabase
from aimet_common.comp_ratio_rounder import CompRatioRounder
logger = AimetLogger.get_area_logger(AimetLogger.LogAreas.CompRatioSelect)
class CompRatioSelectAlgo(metaclass=abc.ABCMeta):
"""
Abstract interface for all compression-ratio selection algorithms
"""
def __init__(self, layer_db: LayerDatabase,
cost_calculator: cc.CostCalculator, cost_metric: CostMetric,
comp_ratio_rounding_algo: Optional[CompRatioRounder]):
"""
Constructor
:param layer_db: Database of compressible layers
"""
# pylint: disable=too-many-arguments
from aimet_common.utils import AimetLogger
from aimet_common.pruner import Pruner
from aimet_common.channel_pruner import select_channels_to_prune
from aimet_common.cost_calculator import CostCalculator, Cost
from aimet_common.winnow.winnow_utils import update_winnowed_channels
from aimet_tensorflow.utils.graph_saver import save_and_load_graph
from aimet_tensorflow.utils.common import is_op_compressible, get_ordered_ops
from aimet_tensorflow.layer_database import Layer, LayerDatabase
from aimet_tensorflow.utils.op.conv import WeightTensorUtils
from aimet_tensorflow.winnow import winnow
from aimet_tensorflow.channel_pruning.data_subsampler import DataSubSampler
from aimet_tensorflow.channel_pruning.weight_reconstruction import WeightReconstructor
from aimet_tensorflow.common.graph_eval import initialize_uninitialized_vars
logger = AimetLogger.get_area_logger(AimetLogger.LogAreas.ChannelPruning)
class InputChannelPruner(Pruner):
"""
Pruner for Channel Pruning method
"""
def __init__(self, input_op_names: List[str], output_op_names: List[str],
data_set: tf.data.Dataset, batch_size: int,
num_reconstruction_samples: int,
allow_custom_downsample_ops: bool):
"""
Input Channel Pruner with given dataset, input shape, number of batches and samples per image.
:param input_op_names: list of input op names
:param output_op_names: List of output op names of the model, used to help ConnectedGraph determine valid ops
# POSSIBILITY OF SUCH DAMAGE.
#
# SPDX-License-Identifier: BSD-3-Clause
#
# @@-COPYRIGHT-END-@@
# =============================================================================
""" Module identifier class """
from abc import ABC, abstractmethod
from typing import List, Set
import tensorflow as tf
from aimet_common.utils import AimetLogger
from aimet_tensorflow.common import sub_graph_matcher
from aimet_tensorflow.common.sub_graph_matcher import ModuleIdentifierOpInfo
logger = AimetLogger.get_area_logger(AimetLogger.LogAreas.ConnectedGraph)
class ModuleIdentifier(ABC):
""" Module identifier class for identifying submodules from groups of tf Operations """
def __init__(self, graph: tf.Graph):
""" Initializer for ModuleIdentifier """
self._graph = graph
@abstractmethod
def get_op_info(self, op: tf.Operation) -> ModuleIdentifierOpInfo:
"""
Given a tf op in the graph, return OpInfo class containing:
- opname: Name that op/product graph should use to represent whatever module
this tf op belongs to
- type: Module type that should be stored in the op/product graph for this module
def __init__(self,
graph,
checkpoint,
output_file='./quantize/q_graph',
quant_mode='tf_enhanced',
round_mode='nearest',
op_map=None,
ops_to_ignore=None,
gpu=True,
debug=False,
skip_output=False,
skip_bias=True):
"""
:param graph: The input meta graph to add quantization ops to
:param checkpoint: The checkpoint file for the given graph
:param output_file: The file path for saving the compressed tensorflow graph
:param quant_mode: Indicates which quantization algorithm should be used, either
'tf' or 'tf_enhanced'. Defaults to 'tf_enhanced'.
:param round_mode: The round scheme to used. One of: 'nearest' or 'stochastic'. Default
is 'nearest'.
:param op_map: A map representing the op sequences to identify and quantize. See op_defs.py
for an example of the formatting required.
:param ops_to_ignore: A list of op names to ignore when selecting quantization ops
:param gpu: Indicates on which hardware which the quantization algorithm should run. Currently
defaults to CPU (False). To use GPU pass True (currently broken).
:param debug: Indicates whether debug information should be printed or not. Defaults to False.
:param skip_output: If output quantization is to be turned off. Default to False
:param skip_bias: If bias quantization is to be turned off. Default to True
:returns: An object which can be used to perform quantization on a tensorflow graph
:raises: ValueError: An error occurred processing one of the input parameters.
"""
# pylint: disable=too-many-arguments
self._log = AimetLogger.get_area_logger(AimetLogger.LogAreas.Quant)
self._debug = debug
self._default_meta_graph = graph
self._default_checkpoint = checkpoint
self._output_file = output_file
self._output_dir = os.path.dirname(output_file)
self._skip_output = skip_output
self._skip_bias = skip_bias
if not os.path.exists(self._output_dir):
os.makedirs(self._output_dir)
self._log.info('Saving quantized model as: %s', output_file)
if op_map:
self._op_map = op_map
else:
self._op_map = op_defs.default_op_map
if not ops_to_ignore:
ops_to_ignore = []
self._ops_to_ignore = ops_to_ignore
if quant_mode not in _QUANT_MODES:
raise ValueError('Invalid quantization mode: ' + quant_mode)
self._quant_mode = _QUANT_MODES[quant_mode]
self._quant_mode_str = quant_mode.upper()
if round_mode not in _ROUNDING_MODES:
raise ValueError('Invalid rounding mode: ' + round_mode)
self._round_mode_str = round_mode.upper()
self._comp_mode = libpymo.ComputationMode.COMP_MODE_GPU if gpu else libpymo.ComputationMode.COMP_MODE_CPU
self._gpu = gpu
self._quant_act_ops = []
self._activation_encodings = {
'quant_mode': quant_mode,
'encodings': {},
'activation_bitwidth': 8
}
self._is_train_variable = None
self._input_tensor_names = None
# Todo: Need to look at these attributes and see how to handle them better
# Very likely these attributes don't need to be object attributes
self._saver = None
self._bw_acts = None
self._bw_params = None
self._forward_callback = None
self._sess = None
self._iterations = None