# -*- mode: python -*-
# =============================================================================
# @@-COPYRIGHT-START-@@
#
# Copyright (c) 2019, Qualcomm Innovation Center, Inc. All rights reserved.
#
# Redistribution and use in source and binary forms, with or without
# modification, are permitted provided that the following conditions are met:
#
# 1. Redistributions of source code must retain the above copyright notice,
# this list of conditions and the following disclaimer.
#
# 2. Redistributions in binary form must reproduce the above copyright notice,
# this list of conditions and the following disclaimer in the documentation
# and/or other materials provided with the distribution.
#
# 3. Neither the name of the copyright holder nor the names of its contributors
# may be used to endorse or promote products derived from this software
# without specific prior written permission.
#
# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
# AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
# IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
# ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE
# LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
# CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
# SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
# INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
# 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-@@
# =============================================================================
# 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
""" Code to perform bias correction for layers """
from typing import Callable, Tuple, List, Union, Dict
import copy
import torch
import torch.nn
import numpy as np
import aimet_common.libpymo as 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
"""
# first check if the model is on GPU or not
if utils.is_model_on_gpu(model):
batch = batch.cuda()
try:
with utils.in_eval_mode(model), torch.no_grad():
_ = model(batch)
except StopForwardException:
pass
def get_quantized_dequantized_weight(layer: torch.nn.Module) -> torch.Tensor:
"""
Gets quantized dequantized weights of a layer
:param layer: Conv/FC layer
:return: quantized dequantized weights
"""
weight_tensor = layer._module_to_wrap.weight
weight_quantizer = layer.param_quantizers['weight']
quant_dequant_weights = weight_quantizer.quantize_dequantize(weight_tensor, weight_quantizer.round_mode)
return quant_dequant_weights
def register_fwd_hook_for_layer(layer: torch.nn.Module, hook: Callable) -> torch.utils.hooks.RemovableHandle:
"""
register forward hook for given layer
:param layer: layer
:param hook: hook function
:return: hook handle
"""
hook_handle = layer.register_forward_hook(hook)
return hook_handle
def get_output_data(layer: torch.nn.Module, model: torch.nn.Module, images_in_one_batch: torch.Tensor) -> np.ndarray:
"""
Function to get output values of a layer
:param layer: layer
:param model: model
:param images_in_one_batch
:return: list of output of layer for all batches of images
"""
def _hook_to_collect_output_data(module, _, out_data):
"""
hook to collect output data
"""
out_data = utils.to_numpy(out_data)
orig_layer_out_data.append(out_data)
raise StopForwardException
hook_handles = list()
orig_layer_out_data = list()
# register forward hooks
hook_handles.append(register_fwd_hook_for_layer(layer, _hook_to_collect_output_data))
# forward pass for 1 batch for model
forward_pass(model, images_in_one_batch)
output_data = np.vstack(orig_layer_out_data)
# remove hook handles
for hook_handle in hook_handles:
hook_handle.remove()
return output_data
def call_empirical_mo_correct_bias(layer: torch.nn.Module, bias_correction: libpymo.BiasCorrection):
"""
:param layer: Layer to be corrected
:param bias_correction: BiasCorrection object to call pymo interface
"""
device = layer.bias.device
bias_tensor = libpymo.TensorParamBiasCorrection()
bias_tensor.data = layer.bias.detach().cpu().numpy()
bias_correction.correctBias(bias_tensor)
bias = torch.nn.Parameter(torch.Tensor(bias_tensor.data))
layer.bias.data = bias.to(device=device)
def call_analytical_mo_correct_bias(layer: torch.nn.Module, bn: Union[torch.nn.BatchNorm2d, None],
activation_type: Union[ActivationType, None]):
"""
:param layer: Layer to be corrected
:param bn: Input BN to layer
:param activation_type: Input activation to layer
"""
bias_correction = libpymo.BnBasedBiasCorrection()
# Passed wrapped layer since quantized network has to be corrected
device = layer._modules['_module_to_wrap'].bias.device
quant_dequant_weight = get_quantized_dequantized_weight(layer)
weight_tensor = layer._module_to_wrap.weight
# Transpose weights to C, N, H, W from N, C, H, W since axis are flipped for transposed conv
if isinstance(layer._module_to_wrap, torch.nn.ConvTranspose2d) and layer._module_to_wrap.groups == 1:
weight_tensor = weight_tensor.permute(1, 0, 2, 3)
quant_dequant_weight = quant_dequant_weight.permute(1, 0, 2, 3)
quant_dequant_weight = quant_dequant_weight.detach().cpu().numpy()
weight_tensor = weight_tensor.detach().cpu().numpy()
bias_tensor = libpymo.TensorParamBiasCorrection()
bias_tensor.data = layer._module_to_wrap.bias.detach().cpu().numpy()
# Assigning activation to No Acivation
activation = libpymo.ActivationType.noActivation
bn_params = libpymo.BnParamsBiasCorr()
if bn is None:
shape = weight_tensor.shape[1]
bn_params.gamma = np.ones(shape)
bn_params.beta = np.zeros(shape)
else:
bn_params.gamma = bn.get_module().weight.detach().cpu().numpy()
bn_params.beta = bn.get_module().bias.detach().cpu().numpy()
if activation_type == ActivationType.relu:
activation = libpymo.ActivationType.relu
# Relu6's type in connected graph is hardtanh
elif activation_type == ActivationType.relu6:
activation = libpymo.ActivationType.relu6
bias_correction.correctBias(bias_tensor, quant_dequant_weight, weight_tensor, bn_params, activation)
# Assigning the updated bias back to the layer
bias = torch.nn.Parameter(torch.Tensor(bias_tensor.data))
layer._module_to_wrap.bias.data = bias.to(device=device)
[docs]def correct_bias(model: torch.nn.Module, quant_params: qsim.QuantParams,
num_quant_samples: int, data_loader, num_bias_correct_samples: int,
conv_bn_dict: Union[Dict[torch.nn.Module, ConvBnInfoType], None] = None,
perform_only_empirical_bias_corr: bool = True,
layers_to_ignore: List[torch.nn.Module] = None):
"""
Corrects bias for each Conv layer of model (unless ignored). A combination of Analytical and Empirical Bias
Correction is used i.e. all the layers which can be corrected using Analytical Bias Correction are corrected
using Analytical Bias Correction and remaining layers are corrected using Empirical method.
Returns an in-place corrected floating point model
:param model: Model to be corrected
:param quant_params: Named tuple for quantization simulation for bias correction
:param num_quant_samples: number of samples of images to pass through quantization sim for bias correction.
:param data_loader: data loader for the model
:param num_bias_correct_samples: number of samples for Bias correction
:param conv_bn_dict: Dict of conv and bn with information related to activation. If None, the function calc it
:param perform_only_empirical_bias_corr: Default True. If true will perform only empirical Bias Corr for all layers
irrespective of the fact that layer is eligible for Analytical Bias Corr.
:param layers_to_ignore: list of layer names for which we need to skip bias correction.
"""
if layers_to_ignore is None:
layers_to_ignore = []
# Find batch size and shape of input tensor
batch_size, input_shape = utils.get_input_shape_batch_size(data_loader)
# Rounding up number of samples to batch size
n_batches_bias_correction = int(np.ceil(num_bias_correct_samples / batch_size))
n_batches_quantization = int(np.ceil(num_quant_samples / batch_size))
data_loader_n_samples_bias_corr = utils.IterFirstX(data_loader, n_batches_bias_correction)
data_loader_n_samples_quant = utils.IterFirstX(data_loader, n_batches_quantization)
# TODO: Remove wrapper function
# Create a wrapping function for data loader for quantization
def pass_data_through_model(model, early_stopping_iterations=None, use_cuda=False):
# pylint: disable=unused-argument
# forward pass for given number of batches for model
for (images_in_one_batch, *_) in data_loader_n_samples_quant:
forward_pass(model, images_in_one_batch)
ordered_conv_linear_nodes = get_ordered_lists_of_conv_fc(model, input_shape)
if conv_bn_dict is None:
conv_bn_dict = find_all_conv_bn_with_activation(model, input_shape)
# Create a copy of the model as reference model
model_copy = copy.deepcopy(model)
# Add bias for all the layers whose bias is None
for name, module in ordered_conv_linear_nodes:
if module.bias is None:
if isinstance(module, (torch.nn.Conv2d, torch.nn.ConvTranspose2d)):
output_size = module.out_channels
elif isinstance(module, torch.nn.Linear):
output_size = module.out_features
module.bias = torch.nn.Parameter(torch.zeros(output_size))
module.bias.data = module.bias.data.to(device=module.weight.device)
# Quantize full model
dummy_tensors = utils.create_rand_tensors_given_shapes(input_shape, utils.get_device(model))
q = qsim.QuantizationSimModel(model=model, quant_scheme=quant_params.quant_scheme,
rounding_mode=quant_params.round_mode,
default_output_bw=quant_params.act_bw,
default_param_bw=quant_params.weight_bw,
in_place=True,
dummy_input=dummy_tensors, config_file=quant_params.config_file)
# make sure model got updated in-place before we use it for bc updates
assert(q.model is model)
# updates to skip_output_activation and layers_to_ignore
for name, module in model.named_modules():
# Skip all layer's output quantization
if isinstance(module, QcQuantizeWrapper):
module.output_quantizers[0].enabled = False
q.compute_encodings(pass_data_through_model, None)
# For first conv layer, perform analytical bc if perform_only_empirical_bias_corr is set to False
# and layer is not marked to be ignored during bc.
if not perform_only_empirical_bias_corr:
module_name, module = ordered_conv_linear_nodes[0]
if module not in layers_to_ignore:
logger.info('Correcting layer %s using Analytical Bias Correction', module_name)
quantize_layer = utils.get_layer_by_name(model, module_name)
call_analytical_mo_correct_bias(quantize_layer, None, None)
logger.info('Corrected bias for the layer')
ordered_conv_linear_nodes.pop(0)
for module_name, module in ordered_conv_linear_nodes:
# Ignore all layers which are skipped by user
if module in layers_to_ignore:
continue
else:
# make sure module is in the model used by qsim.
assert(module in list(q.model.modules()))
# Analytical Bias Correction is only done for Conv layers
reference_layer = utils.get_layer_by_name(model_copy, module_name)
quantize_layer = utils.get_layer_by_name(model, module_name)
if module in conv_bn_dict.keys():
bn_layer_info = conv_bn_dict[module]
if perform_only_empirical_bias_corr or bn_layer_info is None or bn_layer_info.input_bn is None:
logger.info('Correcting layer %s using Empirical Bias Correction', module_name)
bias_correction = libpymo.BiasCorrection()
# Get output from quantized model and reference model
for images_in_one_batch, *_ in data_loader_n_samples_bias_corr:
reference_output_batch = get_output_data(reference_layer, model_copy, images_in_one_batch)
quantized_model_output_batch = get_output_data(quantize_layer, model, images_in_one_batch)
if isinstance(reference_layer, torch.nn.Linear):
extended_shape = np.concatenate((reference_output_batch.shape, np.array([1, 1])))
reference_output_batch = reference_output_batch.reshape(extended_shape)
quantized_model_output_batch = quantized_model_output_batch.reshape(extended_shape)
bias_correction.storePreActivationOutput(reference_output_batch)
bias_correction.storeQuantizedPreActivationOutput(quantized_model_output_batch)
call_empirical_mo_correct_bias(module, bias_correction)
else:
logger.info('Correcting layer %s using Analytical Bias Correction', module_name)
call_analytical_mo_correct_bias(quantize_layer, bn_layer_info.input_bn,
bn_layer_info.in_activation_type)
logger.info('Corrected bias for the layer')
SaveUtils.remove_quantization_wrappers(model)
logger.info('Completed bias correction')
def find_all_conv_bn_with_activation(model: torch.nn.Module, input_shape: Tuple) -> Dict:
"""
Uses searcher to find preceding and next bn layers for a conv/linear layer
:param model: PyTorch model
:param input_shape: shape of input to the model
:return: dictionary of conv/linear layers with associated bn op / activation info
"""
activation_types = ['Relu', 'Clip']
# initialize all patterns to be matched and associated call back functions
patterns_with_callbacks = []
layer_select_handler = ConvBnPatternHandler()
patterns_with_callbacks.append(PatternType(pattern=['BatchNormalization', 'Conv'],
action=layer_select_handler))
patterns_with_callbacks.append(PatternType(pattern=['BatchNormalization', 'ConvTranspose'],
action=layer_select_handler))
patterns_with_callbacks.append(PatternType(pattern=['Conv'],
action=layer_select_handler))
patterns_with_callbacks.append(PatternType(pattern=['Gemm'],
action=layer_select_handler))
for activation in activation_types:
patterns_with_callbacks.append(PatternType(pattern=['BatchNormalization', activation, 'Conv'],
action=layer_select_handler))
patterns_with_callbacks.append(PatternType(pattern=['BatchNormalization', activation, 'ConvTranspose'],
action=layer_select_handler))
device = utils.get_device(model)
connected_graph = ConnectedGraph(model, (torch.rand(input_shape).to(device),))
# create graph searcher instance with connected graph and patterns to search
graph_searcher = GraphSearcher(connected_graph, patterns_with_callbacks)
# get all conv/linear and bn info
graph_searcher.find_all_patterns_in_graph_apply_actions()
convs_bn_activation_dict = layer_select_handler.get_conv_linear_bn_info_dict()
return convs_bn_activation_dict