Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

Pruning masks applying #994

Merged
merged 3 commits into from
Nov 3, 2021
+23,383 −23,631
Merged

Pruning masks applying #994

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
678aabd
Pruning masks applying
daniil-lyakhov Oct 29, 2021
879a21f
Enable backward compatibility
daniil-lyakhov Nov 2, 2021
ed5dece
Backward compat test
daniil-lyakhov Nov 2, 2021
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
Prev Previous commit
Backward compat test
  • Loading branch information
@daniil-lyakhov
daniil-lyakhov committed Nov 2, 2021
commit ed5dece6ec966b3f06ac0fb47e341ea4b70eb575
25 changes: 19 additions & 6 deletions nncf/torch/utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -13,6 +13,7 @@
from collections import OrderedDict
from typing import Dict, Any

import warnings
import numpy as np
import random
import torch
Expand Down Expand Up @@ -298,19 +299,30 @@ def default_distributed_unwrapper(model: nn.Module):
return model


def maybe_convert_legacy_names_in_model_state(state_dict_to_load: Dict[str, Any]):
def maybe_convert_legacy_names_in_model_state(state_dict_to_load: Dict[str, Any]) -> None:
"""
Convert legacy layer names in compressed model state dict in case such names exist.

:param state_dict_to_load: State dict to convert.
"""
legacy_bn_names = [name for name in state_dict_to_load if 'BatchNorm2d' in name]
for name in legacy_bn_names:
tensor = state_dict_to_load.pop(name)
new_name = name.replace('BatchNorm2d', 'NNCFBatchNorm')
state_dict_to_load[new_name] = tensor

if legacy_bn_names:
nncf_logger.warning('Legacy Batch Norm layer names was detected in checkpoint model state dict.'
' All occurrences of `BatchNorm2d` in nodes names was replaced by `NNCFBatchNorm`')
warnings.warn('Legacy Batch Norm layer names was detected in checkpoint model state dict.'
' All occurrences of `BatchNorm2d` in nodes names was replaced by `NNCFBatchNorm`',
category=DeprecationWarning)


def maybe_convert_legacy_names_in_compress_state(compression_state: Dict[str, Any]):
def maybe_convert_legacy_names_in_compress_state(compression_state: Dict[str, Any]) -> None:
"""
Convert legacy layer names in compression state in case such names exist.

:param compression_state: Compression state to convert.
"""
if not compression_state or BaseController.BUILDER_STATE not in compression_state:
return

Expand All @@ -327,5 +339,6 @@ def maybe_convert_legacy_names_in_compress_state(compression_state: Dict[str, An
point['qip']['target_node_name'] = name.replace('BatchNorm2d', 'NNCFBatchNorm')

if legacy_bn_names:
nncf_logger.warning('Legacy Batch Norm layer names was detected in quantization setup target point names.'
' All occurrences of `BatchNorm2d` in nodes names was replaced by `NNCFBatchNorm`')
warnings.warn('Legacy Batch Norm layer names was detected in quantization setup target point names.'
' All occurrences of `BatchNorm2d` in nodes names was replaced by `NNCFBatchNorm`',
category=DeprecationWarning)
99 changes: 99 additions & 0 deletions tests/torch/test_backward_compat.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -22,6 +22,7 @@
from examples.torch.common.execution import prepare_model_for_execution
from examples.torch.common.model_loader import load_model
from examples.torch.common.sample_config import SampleConfig
from nncf.api.compression import CompressionStage
from nncf.torch import register_default_init_args
from nncf.torch.checkpoint_loading import load_state
from nncf.common.graph.definitions import MODEL_INPUT_OP_NAME
Expand Down Expand Up @@ -180,3 +181,101 @@ def test_can_compress_with_config_and_resume_of_old_checkpoint():
})
register_bn_adaptation_init_args(config)
create_compressed_model_and_algo_for_test(model, config, compression_state=old_style_sd)


# BN Wrapping backward compatibility test


class ConvBNLayer(torch.nn.Module):
def __init__(self):
super().__init__()
self.conv = torch.nn.Conv2d(3, 9, (3, 3))
self.bn = torch.nn.BatchNorm2d(9)
self.conv1 = torch.nn.Conv2d(9, 3, (3, 3))
self.bn1 = torch.nn.BatchNorm2d(3)

def forward(self, x):
x = self.bn(self.conv(x))
return self.bn1(self.conv1(x))


sd_without_nncf_bn_wrapping = {
'nncf_module.conv.weight': torch.empty([9, 3, 3, 3]),
'nncf_module.conv.bias': torch.empty([9]),
'nncf_module.conv.nncf_padding_value': torch.empty([1]),
'nncf_module.conv.pre_ops.0.op._num_bits': torch.empty([1]),
'nncf_module.conv.pre_ops.0.op.signed_tensor': torch.empty([1]),
'nncf_module.conv.pre_ops.0.op.enabled': torch.empty([1]),
'nncf_module.conv.pre_ops.0.op.scale': torch.empty([9, 1, 1, 1]),
'nncf_module.bn.weight': torch.empty([9]),
'nncf_module.bn.bias': torch.empty([9]),
'nncf_module.bn.running_mean': torch.empty([9]),
'nncf_module.bn.running_var': torch.empty([9]),
'nncf_module.bn.num_batches_tracked': torch.empty([]),
'nncf_module.conv1.weight': torch.empty([3, 9, 3, 3]),
'nncf_module.conv1.bias': torch.empty([3]),
'nncf_module.conv1.nncf_padding_value': torch.empty([1]),
'nncf_module.conv1.pre_ops.0.op._num_bits': torch.empty([1]),
'nncf_module.conv1.pre_ops.0.op.signed_tensor': torch.empty([1]),
'nncf_module.conv1.pre_ops.0.op.enabled': torch.empty([1]),
'nncf_module.conv1.pre_ops.0.op.scale': torch.empty([3, 1, 1, 1]),
'nncf_module.bn1.weight': torch.empty([3]),
'nncf_module.bn1.bias': torch.empty([3]),
'nncf_module.bn1.running_mean': torch.empty([3]),
'nncf_module.bn1.running_var': torch.empty([3]),
'nncf_module.bn1.num_batches_tracked': torch.empty([]),
'external_quantizers./nncf_model_input_0|OUTPUT._num_bits': torch.empty([1]),
'external_quantizers./nncf_model_input_0|OUTPUT.signed_tensor': torch.empty([1]),
'external_quantizers./nncf_model_input_0|OUTPUT.enabled': torch.empty([1]),
'external_quantizers./nncf_model_input_0|OUTPUT.scale': torch.empty([1]),
# Old bn layer names: |||||||||||
'external_quantizers.ConvBNLayer/BatchNorm2d[bn]/batch_norm_0|OUTPUT._num_bits': torch.empty([1]),
'external_quantizers.ConvBNLayer/BatchNorm2d[bn]/batch_norm_0|OUTPUT.signed_tensor': torch.empty([1]),
'external_quantizers.ConvBNLayer/BatchNorm2d[bn]/batch_norm_0|OUTPUT.enabled': torch.empty([1]),
'external_quantizers.ConvBNLayer/BatchNorm2d[bn]/batch_norm_0|OUTPUT.scale': torch.empty([1])
}

compression_state_without_bn_wrapping = {
'builder_state':
{'quantization':
{'quantizer_setup':
{'quantization_points':
{1: {'qip': {'target_node_name': '/nncf_model_input_0', 'input_port_id': None},
'qip_class': 'ActivationQuantizationInsertionPoint',
'qconfig':
{'num_bits': 8, 'mode': 'symmetric', 'signedness_to_force': None, 'per_channel': False},
'directly_quantized_operator_node_names': ['ConvBNLayer/NNCFConv2d[conv]/conv2d_0']},
# Old bn layer name: |||||||||||
2: {'qip': {'target_node_name': 'ConvBNLayer/BatchNorm2d[bn]/batch_norm_0', 'input_port_id': None},
'qip_class': 'ActivationQuantizationInsertionPoint',
'qconfig':
{'num_bits': 8, 'mode': 'symmetric', 'signedness_to_force': None, 'per_channel': False},
'directly_quantized_operator_node_names': ['ConvBNLayer/NNCFConv2d[conv1]/conv2d_0']},
4: {'qip': {'target_node_name': 'ConvBNLayer/NNCFConv2d[conv]/conv2d_0'},
'qip_class': 'WeightQuantizationInsertionPoint',
'qconfig':
{'num_bits': 8, 'mode': 'symmetric', 'signedness_to_force': True, 'per_channel': True},
'directly_quantized_operator_node_names': ['ConvBNLayer/NNCFConv2d[conv]/conv2d_0']},
5: {'qip': {'target_node_name': 'ConvBNLayer/NNCFConv2d[conv1]/conv2d_0'},
'qip_class': 'WeightQuantizationInsertionPoint',
'qconfig':
{'num_bits': 8, 'mode': 'symmetric', 'signedness_to_force': True, 'per_channel': True},
'directly_quantized_operator_node_names': ['ConvBNLayer/NNCFConv2d[conv1]/conv2d_0']}},
'unified_scale_groups': {}, 'shared_input_operation_set_groups': {0: [1, 4], 1: [2, 5]}},
'build_time_metric_infos': {'aq_potential_num': 3, 'wq_potential_num': 4}}},
'ctrl_state': {'quantization': {'loss_state': None, 'scheduler_state': {'current_step': -1, 'current_epoch': -1},
'compression_stage': CompressionStage.FULLY_COMPRESSED}}}


def test_quantization_ckpt_without_wrapped_bn_loading():
model = ConvBNLayer()
config = get_basic_quantization_config(input_info={
"sample_size": [1, 3, 100, 100]
})
register_bn_adaptation_init_args(config)
with pytest.deprecated_call():
compressed_model, _ = \
create_compressed_model_and_algo_for_test(model, config,
compression_state=compression_state_without_bn_wrapping)
with pytest.deprecated_call():
_ = load_state(compressed_model, sd_without_nncf_bn_wrapping, is_resume=True)