From 3cde6c5eb68ed8e6e279b24f4d0283d7caed99ce Mon Sep 17 00:00:00 2001
From: Lev Zlotnik <46742999+levzlotnik@users.noreply.github.com>
Date: Mon, 3 Jun 2019 15:14:10 +0300
Subject: [PATCH] [Breaking] PTQ: Removed special handling of clipping
overrides
* In PostTrainLinearQuantizer - moved 'clip_acts' and 'clip_n_stds'
to overrides, removed 'no_clip_layers' parameter from __init__
* The 'no_clip_layers' command line argument REMAINS, handled in
PostTrainLinearQuantizer.from_args()
* Removed old code from comments, fixed warnings in
test_post_train_quant.py
* Updated tests
* Update post-train quant sample YAML
---
distiller/quantization/range_linear.py | 29 +++++-----
.../resnet18_imagenet_post_train.yaml | 2 +-
tests/full_flow_tests.py | 6 +-
tests/test_post_train_quant.py | 57 ++++++++++++++++++-
4 files changed, 75 insertions(+), 19 deletions(-)
diff --git a/distiller/quantization/range_linear.py b/distiller/quantization/range_linear.py
index fb460c0..9545a05 100644
--- a/distiller/quantization/range_linear.py
+++ b/distiller/quantization/range_linear.py
@@ -768,9 +768,6 @@ class PostTrainLinearQuantizer(Quantizer):
overrides (:obj:`OrderedDict`, optional): Overrides the layers quantization settings.
mode (LinearQuantMode): Quantization mode to use (symmetric / asymmetric-signed / unsigned)
clip_acts (ClipMode): Activations clipping mode to use
- no_clip_layers (list): List of fully-qualified layer names for which activations clipping should not be done.
- A common practice is to not clip the activations of the last layer before softmax.
- Applicable only if clip_acts is True.
per_channel_wts (bool): Enable quantization of weights using separate quantization parameters per
output channel
model_activation_stats (str / dict / OrderedDict): Either a path to activation stats YAML file, or a dictionary
@@ -788,7 +785,7 @@ class PostTrainLinearQuantizer(Quantizer):
to half precision, regardless of bits_activations/parameters/accum.
"""
def __init__(self, model, bits_activations=8, bits_parameters=8, bits_accum=32,
- overrides=None, mode=LinearQuantMode.SYMMETRIC, clip_acts=ClipMode.NONE, no_clip_layers=None,
+ overrides=None, mode=LinearQuantMode.SYMMETRIC, clip_acts=ClipMode.NONE,
per_channel_wts=False, model_activation_stats=None, fp16=False, clip_n_stds=None,
scale_approx_mult_bits=None):
super(PostTrainLinearQuantizer, self).__init__(model, bits_activations=bits_activations,
@@ -817,32 +814,33 @@ class PostTrainLinearQuantizer(Quantizer):
'mode': str(mode).split('.')[1],
'clip_acts': _enum_to_str(clip_acts),
'clip_n_stds': clip_n_stds,
- 'no_clip_layers': no_clip_layers,
'per_channel_wts': per_channel_wts,
'fp16': fp16,
'scale_approx_mult_bits': scale_approx_mult_bits}}
def replace_param_layer(module, name, qbits_map, per_channel_wts=per_channel_wts,
- mode=mode, fp16=fp16, scale_approx_mult_bits=scale_approx_mult_bits):
+ mode=mode, fp16=fp16, scale_approx_mult_bits=scale_approx_mult_bits,
+ clip_acts=clip_acts, clip_n_stds=clip_n_stds):
if fp16:
return FP16Wrapper(module)
norm_name = distiller.utils.normalize_module_name(name)
- clip = self.clip_acts if norm_name not in self.no_clip_layers else ClipMode.NONE
+ clip_acts = verify_clip_mode(clip_acts)
return RangeLinearQuantParamLayerWrapper(module, qbits_map[name].acts, qbits_map[name].wts,
- num_bits_accum=self.bits_accum, mode=mode, clip_acts=clip,
+ num_bits_accum=self.bits_accum, mode=mode, clip_acts=clip_acts,
per_channel_wts=per_channel_wts,
activation_stats=self.model_activation_stats.get(norm_name, None),
clip_n_stds=clip_n_stds,
scale_approx_mult_bits=scale_approx_mult_bits)
def replace_non_param_layer(wrapper_type, module, name, qbits_map, fp16=fp16,
- scale_approx_mult_bits=scale_approx_mult_bits):
+ scale_approx_mult_bits=scale_approx_mult_bits,
+ clip_acts=clip_acts, clip_n_stds=clip_n_stds):
if fp16:
return FP16Wrapper(module)
norm_name = distiller.utils.normalize_module_name(name)
- clip = self.clip_acts if norm_name not in self.no_clip_layers else ClipMode.NONE
+ clip_acts = verify_clip_mode(clip_acts)
try:
- return wrapper_type(module, qbits_map[name].acts, mode=mode, clip_acts=clip,
+ return wrapper_type(module, qbits_map[name].acts, mode=mode, clip_acts=clip_acts,
activation_stats=self.model_activation_stats.get(norm_name, None),
clip_n_stds=clip_n_stds, scale_approx_mult_bits=scale_approx_mult_bits)
except NoStatsError:
@@ -858,7 +856,6 @@ class PostTrainLinearQuantizer(Quantizer):
stats=self.model_activation_stats.get(norm_name, None))
self.clip_acts = clip_acts
- self.no_clip_layers = no_clip_layers or []
self.clip_n_stds = clip_n_stds
self.model_activation_stats = model_activation_stats or {}
self.bits_accum = bits_accum
@@ -885,17 +882,21 @@ class PostTrainLinearQuantizer(Quantizer):
return distiller.config_component_from_file_by_class(model, args.qe_config_file,
'PostTrainLinearQuantizer')
else:
+ overrides = OrderedDict(
+ [(layer, OrderedDict([('clip_acts', 'NONE')]))
+ for layer in args.qe_no_clip_layers]
+ )
return cls(model,
bits_activations=args.qe_bits_acts,
bits_parameters=args.qe_bits_wts,
bits_accum=args.qe_bits_accum,
mode=args.qe_mode,
clip_acts=args.qe_clip_acts,
- no_clip_layers=args.qe_no_clip_layers,
per_channel_wts=args.qe_per_channel,
model_activation_stats=args.qe_stats_file,
clip_n_stds=args.qe_clip_n_stds,
- scale_approx_mult_bits=args.qe_scale_approx_bits)
+ scale_approx_mult_bits=args.qe_scale_approx_bits,
+ overrides=overrides)
###############################################################################
diff --git a/examples/quantization/post_train_quant/resnet18_imagenet_post_train.yaml b/examples/quantization/post_train_quant/resnet18_imagenet_post_train.yaml
index 0b6252b..0e07184 100644
--- a/examples/quantization/post_train_quant/resnet18_imagenet_post_train.yaml
+++ b/examples/quantization/post_train_quant/resnet18_imagenet_post_train.yaml
@@ -45,7 +45,6 @@ quantizers:
model_activation_stats: ../quantization/post_train_quant/stats/resnet18_quant_stats.yaml
per_channel_wts: True
clip_acts: AVG
- no_clip_layers: fc
overrides:
# First and last layers + element-wise add layers in 8-bits
conv1:
@@ -57,3 +56,4 @@ quantizers:
fc:
bits_weights: 8
bits_activations: 8
+ clip_acts: NONE
diff --git a/tests/full_flow_tests.py b/tests/full_flow_tests.py
index 4a27458..01f15c6 100755
--- a/tests/full_flow_tests.py
+++ b/tests/full_flow_tests.py
@@ -116,9 +116,9 @@ TestConfig = namedtuple('TestConfig', ['args', 'dataset', 'checker_fn', 'checker
test_configs = [
TestConfig('--arch simplenet_cifar --epochs 2', DS_CIFAR, accuracy_checker, [44.610, 92.080]),
- TestConfig('-a resnet20_cifar --resume {0} --quantize-eval --evaluate'.
- format(os.path.join(examples_root, 'ssl', 'checkpoints', 'checkpoint_trained_dense.pth.tar')),
- DS_CIFAR, accuracy_checker, [91.710, 99.610]),
+ TestConfig('-a resnet20_cifar --resume {0} --quantize-eval --evaluate --qe-clip-acts avg --qe-no-clip-layers {1}'.
+ format(os.path.join(examples_root, 'ssl', 'checkpoints', 'checkpoint_trained_dense.pth.tar'), 'fc'),
+ DS_CIFAR, accuracy_checker, [91.55, 99.63]),
TestConfig('-a preact_resnet20_cifar --epochs 2 --compress {0}'.
format(os.path.join('full_flow_tests', 'preact_resnet20_cifar_pact_test.yaml')),
DS_CIFAR, accuracy_checker, [54.590, 94.810]),
diff --git a/tests/test_post_train_quant.py b/tests/test_post_train_quant.py
index 4522e01..2514c21 100644
--- a/tests/test_post_train_quant.py
+++ b/tests/test_post_train_quant.py
@@ -16,10 +16,13 @@
import pytest
import torch
import torch.testing
+import torch.nn as nn
from collections import OrderedDict
from distiller.quantization import RangeLinearQuantParamLayerWrapper, LinearQuantMode, ClipMode, \
- RangeLinearQuantConcatWrapper, RangeLinearQuantEltwiseMultWrapper, RangeLinearQuantEltwiseAddWrapper
+ RangeLinearQuantConcatWrapper, RangeLinearQuantEltwiseMultWrapper, RangeLinearQuantEltwiseAddWrapper, \
+ PostTrainLinearQuantizer
+from distiller.data_loggers import QuantCalibrationStatsCollector, collector_context
import distiller.modules
@@ -352,3 +355,55 @@ def test_eltwise_add_layer_wrapper(inputs, eltwise_add_stats, mode, clip_acts, e
output = model(*inputs)
torch.testing.assert_allclose(output, expected_output)
+
+
+class DummyRNN(nn.Module):
+ def __init__(self, input_size, hidden_size, num_layers):
+ super(DummyRNN, self).__init__()
+ self.rnn = distiller.modules.DistillerLSTM(input_size, hidden_size, num_layers)
+ self.softmax = nn.Softmax(dim=-1)
+
+ def forward(self, x, h=None):
+ y, h = self.rnn(x, h)
+ y = self.softmax(y)
+ return y, h
+
+
+@pytest.fixture()
+def rnn_model():
+ return DummyRNN(20, 128, 1)
+
+
+@pytest.fixture()
+def rnn_model_stats(rnn_model):
+ collector = QuantCalibrationStatsCollector(rnn_model)
+ dummy_input = torch.randn(35, 50, 20)
+ with collector_context(collector):
+ y,h = rnn_model(dummy_input)
+ return collector.value()
+
+
+@pytest.mark.parametrize(
+ "overrides, e_clip_acts, e_n_stds",
+ [
+ (None, ClipMode.AVG, 0),
+
+ (distiller.utils.yaml_ordered_load("""
+ rnn.cells.0.eltwisemult_hidden:
+ clip_acts: NONE
+ """), ClipMode.NONE, 0),
+
+ (distiller.utils.yaml_ordered_load("""
+ rnn.cells.0.eltwisemult_hidden:
+ clip_acts: N_STD
+ clip_n_stds: 2
+ """), ClipMode.N_STD, 2)
+ ]
+)
+def test_override_no_clip(overrides, e_clip_acts, e_n_stds, rnn_model, rnn_model_stats):
+ quantizer = PostTrainLinearQuantizer(rnn_model, clip_acts="AVG", clip_n_stds=0, overrides=overrides,
+ model_activation_stats=rnn_model_stats)
+ quantizer.prepare_model()
+ assert isinstance(quantizer.model.rnn.cells[0].eltwisemult_hidden, RangeLinearQuantEltwiseMultWrapper)
+ assert quantizer.model.rnn.cells[0].eltwisemult_hidden.clip_acts == e_clip_acts
+ assert quantizer.model.rnn.cells[0].eltwisemult_hidden.clip_n_stds == e_n_stds
--
GitLab