Post Training Quantization
Overview
Converting a trained float32
model into an int8
model through quantization after training can reduce the model size and improve the inference performance. In MindSpore Lite, this function is integrated into the model conversion tool converter_lite
. The quantized model can be transformed by configuring a quantization profile
.
MindSpore Lite quantization after training is classified into two types:
Weight quantization: quantizes a weight of a model and compresses only the model size.
float32
inference is still performed during inference.Full quantization: quantizes the weight and activation value of a model. The
int
operation is performed during inference to improve the model inference speed and reduce power consumption.
Configuration Parameter
Post training quantization can be enabled by configuring configFile
through Conversion Tool. The configuration file adopts the style of INI
, For quantization, configurable parameters include common quantization parameter [common_quant_param]
, fixed bit weight quantization parameter [weight_quant_param]
, mixed bit weight quantization parameter [mixed_bit_weight_quant_param]
,full quantization parameter [full_quant_param]
, data preprocess parameter [data_preprocess_param]
and dynamic quantization parameter [dynamic_quant_param]
.
Common Quantization Parameter
common quantization parameters are the basic settings for post training quantization, mainly including quant_type
, bit_num
, min_quant_weight_size
, and min_quant_weight_channel
. The detailed description of the parameters is as follows:
Parameter |
Attribute |
Function Description |
Parameter Type |
Default Value |
Value Range |
---|---|---|---|---|---|
|
Mandatory |
The quantization type. When set to WEIGHT_QUANT, weight quantization is enabled; when set to FULL_QUANT, full quantization is enabled; when set to DYNAMIC_QUANT, dynamic quantization is enabled. |
String |
- |
WEIGHT_QUANT, |
|
Optional |
The number of quantized bits. Currently, weight quantization supports 0-16bit quantization. When it is set to 1-16bit, it is fixed-bit quantization. When it is set to 0bit, mixed-bit quantization is enabled. Full quantization and Dynamic quantization supports 8bit quantization. |
Integer |
8 |
WEIGHT_QUANT:[0,16] |
|
Optional |
Set the threshold of the weight size for quantization. If the number of weights is greater than this value, the weight will be quantized. |
Integer |
0 |
[0, 65535] |
|
Optional |
Set the threshold of the number of weight channels for quantization. If the number of weight channels is greater than this value, the weight will be quantized. |
Integer |
16 |
[0, 65535] |
|
Optional |
Set the name of the operator that does not need to be quantified, and use |
String |
- |
- |
|
Optional |
Set the folder path where the quantized debug information file is saved. |
String |
- |
- |
|
Optional |
The enable switch of compression code for weight quantization. |
Boolean |
True |
True, False |
min_quant_weight_size
andmin_quant_weight_channel
are only valid for weight quantization.Recommendation: When the accuracy of full quantization is not satisfied, you can set
debug_info_save_path
to turn on the Debug mode to get the relevant statistical report, and setskip_quant_node
for operators that are not suitable for quantization to not quantize them.
The common quantization parameter configuration is as follows:
[common_quant_param]
# Supports WEIGHT_QUANT or FULL_QUANT
quant_type=WEIGHT_QUANT
# Weight quantization support the number of bits [0,16], Set to 0 is mixed bit quantization, otherwise it is fixed bit quantization
# Full quantization support 8bit
bit_num=8
# Layers with size of weights exceeds threshold `min_quant_weight_size` will be quantized.
min_quant_weight_size=0
# Layers with channel size of weights exceeds threshold `min_quant_weight_channel` will be quantized.
min_quant_weight_channel=16
# Set the name of the operator that skips the quantization, and use `,` to split between multiple operators.
skip_quant_node=node_name1,node_name2,node_name3
# Set the folder path where the quantization debug information file is saved.
debug_info_save_path=/home/workspace/mindspore/debug_info_save_path
# Enable tensor compression for weight quantization. If parameter bit_num not equal to 8 or 16, it can not be set to false.
enable_encode = true
Fixed Bit Weight Quantization Parameters
The fixed bit weight quantization parameters include dequant_strategy
, per_channel
, bias_correction
. The detailed description of the parameters is as follows:
Parameter |
Attribute |
Function Description |
Parameter Type |
Default Value |
Value Range |
---|---|---|---|---|---|
dequant_strategy |
Optional |
Weight Quantization mode |
String |
- |
ON_THE_FLY. After this parameter is enabled, Ascend online antiquantization mode is activated. |
per_channel |
Optional |
Select PerChannel or PerLayer quantization type. |
Boolean |
True |
True or False. Set to false to enable Perlayer quantization. |
bias_correction |
Optional |
Indicate whether to correct the quantization error. |
Boolean |
True |
True or False. After this parameter is enabled, the accuracy of the quantization model can be improved. |
[weight_quant_param]
dequant_strategy=ON_THE_FLY
# If set to true, it will enable PerChannel quantization, or set to false to enable PerLayer quantization.
per_channel=True
# Whether to correct the quantization error. Recommended to set to true.
bias_correction=False
Mixed Bit Weight Quantization Parameter
The mixed bit weight quantization parameters include init_scale
. When enable the mixed bit weight quantization, the optimal number of bits will be automatically searched for different layers. The detailed description of the parameters is as follows:
Parameter |
Attribute |
Function Description |
Parameter Type |
Default Value |
Value Range |
---|---|---|---|---|---|
init_scale |
Optional |
Initialize the scale. The larger the value, the greater the compression rate, but it will also cause varying degrees of accuracy loss. |
Float |
0.02 |
(0 , 1) |
auto_tune |
Optional |
Automatically search for the init_scale parameter. After setting, it will automatically search for a set of |
Boolean |
False |
True,False |
The mixed bit quantization parameter configuration is as follows:
[mixed_bit_weight_quant_param]
init_scale=0.02
auto_tune=false
Full Quantization Parameters
The full quantization parameters mainly include activation_quant_method
, bias_correction
and target_device
. The detailed description of the parameters is as follows:
Parameter |
Attribute |
Function Description |
Parameter Type |
Default Value |
Value Range |
---|---|---|---|---|---|
activation_quant_method |
Optional |
Activation quantization algorithm |
String |
MAX_MIN |
KL, MAX_MIN, or RemovalOutlier. |
bias_correction |
Optional |
Indicate whether to correct the quantization error. |
Boolean |
True |
True or False. After this parameter is enabled, the accuracy of the converted model can be improved. You are advised to set this parameter to true. |
per_channel |
Optional |
Select PerChannel or PerLayer quantization type. |
Boolean |
True |
True or False. Set to false to enable Perlayer quantization. |
target_device |
Optional |
Full quantization supports multiple hardware backends. After setting the specific hardware, the converted quantization model can execute the proprietry hardware quantization operator library. If not setting, universal quantization lib will be called. |
String |
- |
NVGPU: The quantized model can perform quantitative inference on the NVIDIA GPU. |
Data Preprocessing
Full quantization needs to provide 100-500 calibration data sets for pre-inference, which is used to calculate the quantization parameters of full quantization activation values. If there are multiple input Tensors, the calibration dataset for each input Tensor needs to be saved in a separate folder.
For the BIN calibration dataset, the .bin
file stores the input data buffer, and the format of the input data needs to be consistent with the format of the input data during inference. For 4D data, the default is NHWC
. If the command parameter inputDataFormat
of the converter tool is configured, the format of the input Buffer needs to be consistent.
For the image calibration dataset, post training quantization provides data preprocessing functions such as channel conversion, normalization, resize, and center crop.
Parameter |
Attribute |
Function Description |
Parameter Type |
Default Value |
Value Range |
---|---|---|---|---|---|
calibrate_path |
Mandatory |
The directory where the calibration dataset is stored; if the model has multiple inputs, please fill in the directory where the corresponding data is located one by one, and separate the directory paths with |
String |
- |
input_name_1:/mnt/image/input_1_dir,input_name_2:input_2_dir |
calibrate_size |
Mandatory |
Calibration data size |
Integer |
- |
[1, 65535] |
input_type |
Mandatory |
Correction data file format type |
String |
- |
IMAGE, BIN |
image_to_format |
Optional |
Image format conversion |
String |
- |
RGB, GRAY, BGR |
normalize_mean |
Optional |
Normalized mean |
Vector |
- |
Channel 3: [mean_1, mean_2, mean_3] |
normalize_std |
Optional |
Normalized standard deviation |
Vector |
- |
Channel 3: [std_1, std_2, std_3] |
resize_width |
Optional |
Resize width |
Integer |
- |
[1, 65535] |
resize_height |
Optional |
Resize height |
Integer |
- |
[1, 65535] |
resize_method |
Optional |
Resize algorithm |
String |
- |
LINEAR, NEAREST, CUBIC |
center_crop_width |
Optional |
Center crop width |
Integer |
- |
[1, 65535] |
center_crop_height |
Optional |
Center crop height |
Integer |
- |
[1, 65535] |
The data preprocessing parameter configuration is as follows:
[data_preprocess_param]
# Calibration dataset path, the format is input_name_1:input_1_dir,input_name_2:input_2_dir
# Full quantification must provide correction dataset
calibrate_path=input_name_1:/mnt/image/input_1_dir,input_name_2:input_2_dir
# Calibration data size
calibrate_size=100
# Input type supports IMAGE or BIN
# When set to IMAGE, the image data will be read
# When set to BIN, the `.bin` binary file will be read
input_type=IMAGE
# The output format of the preprocessed image
# Supports RGB or GRAY or BGR
image_to_format=RGB
# Image normalization
# dst = (src - mean) / std
normalize_mean=[127.5, 127.5, 127.5]
normalize_std=[127.5, 127.5, 127.5]
# Image resize
resize_width=224
resize_height=224
# Resize method supports LINEAR or NEAREST or CUBIC
resize_method=LINEAR
# Image center crop
center_crop_width=224
center_crop_height=224
Dynamic Quantization Parameters
The dynamic quantization parameter quant_strategy
sets the dynamic quantizaiton strategy. The detailed description of the parameter is as follows:
Parameter |
Attribute |
Function Description |
Parameter Type |
Default Value |
Value Range |
---|---|---|---|---|---|
quant_strategy |
Optional |
the dynamic quantizaiton strategy |
String |
ALWC |
ALWC: Enable activation perlayer and weight perchannel quantization; |
The dynamic quantization parameter configuration is as follows::
[dynamic_quant_param]
# If set to ALWC, it will enable activation perlayer and weight perchannel quantization. If set to ACWL, it will enable activation perchannel and weight perlayer quantization. Default value is ALWC.
quant_strategy=ACWL
Weight Quantization
Weight quantization supports mixed bit quantization, as well as fixed bit quantization between 1 and 16. The lower the number of bits, the greater the model compression rate, but the accuracy loss is usually larger. The following describes how to use weight quantization and its effects.
Mixed Bit Weight Quantization
Currently, weight quantization supports mixed bit quantization. According to the distribution of model parameters and the initial value of init_scale
set by the user, the number of bits that is most suitable for the current layer will be automatically searched out. When the bit_num
of the configuration parameter is set to 0, mixed bit quantization will be enabled.
The general form of the mixed bit weight requantization command is:
./converter_lite --fmk=ModelType --modelFile=ModelFilePath --outputFile=ConvertedModelPath --configFile=/mindspore/lite/tools/converter/quantizer/config/mixed_bit_weight_quant.cfg
The mixed bit weight quantification configuration file is as follows:
[common_quant_param]
quant_type=WEIGHT_QUANT
# Weight quantization support the number of bits [0,16], Set to 0 is mixed bit quantization, otherwise it is fixed bit quantization
bit_num=0
# Layers with size of weights exceeds threshold `min_quant_weight_size` will be quantized.
min_quant_weight_size=5000
# Layers with channel size of weights exceeds threshold `min_quant_weight_channel` will be quantized.
min_quant_weight_channel=5
[mixed_bit_weight_quant_param]
# Initialization scale in (0,1).
# A larger value can get a larger compression ratio, but it may also cause a larger error.
init_scale=0.02
Users can adjust the weighted parameters according to the model and their own needs.
The init_scale default value is 0.02, and the compression rate is equivalent to the compression effect of 6-7 fixed bits quantization.
Mixed bits need to search for the best bit, and the waiting time may be longer. If you need to view the log, you can set export GLOG_v=1 before the execution to print the relevant Info level log.
Fixed Bit Weight Quantization
Fixed-bit weighting supports fixed-bit quantization between 1 and 16, and users can adjust the weighting parameters according to the requirement.
The general form of the fixed bit weight quantization conversion command is:
./converter_lite --fmk=ModelType --modelFile=ModelFilePath --outputFile=ConvertedModelPath --configFile=/mindspore/lite/tools/converter/quantizer/config/fixed_bit_weight_quant.cfg
The fixed bit weight quantization configuration file is as follows:
[common_quant_param]
quant_type=WEIGHT_QUANT
# Weight quantization support the number of bits [0,16], Set to 0 is mixed bit quantization, otherwise it is fixed bit quantization
bit_num=8
# Layers with size of weights exceeds threshold `min_quant_weight_size` will be quantized.
min_quant_weight_size=0
# Layers with channel size of weights exceeds threshold `min_quant_weight_channel` will be quantized.
min_quant_weight_channel=16
Ascend ON_THE_FLY Quantization
Ascend ON_THE_FLY quantization means runtime weight dequantization. At this stage, only the MINDIR model is supported.
We must add configuration about [ascend_context]
for Ascend ON_THE_FLY quantification as follow:
[common_quant_param]
quant_type=WEIGHT_QUANT
# Weight quantization support the number of bits (0,16]
bit_num=8
# Layers with size of weights exceeds threshold `min_quant_weight_size` will be quantized.
min_quant_weight_size=5000
# Layers with channel size of weights exceeds threshold `min_quant_weight_channel` will be quantized.
min_quant_weight_channel=5
[weight_quant_param]
dequant_strategy=ON_THE_FLY
# If set to true, it will enable PerChannel quantization, or set to false to enable PerLayer quantization.
per_channel=True
# Whether to correct the quantization error. Recommended to set to true.
bias_correction=False
[ascend_context]
# The converted model is suitable for Ascend GE processes
provider=ge
Partial Model Accuracy Result
Model |
Test Dataset |
FP32 Model Accuracy |
Weight Quantization Accuracy (8 bits) |
---|---|---|---|
77.60% |
77.53% |
||
70.96% |
70.56% |
||
71.56% |
71.53% |
All the preceding results are obtained in the x86 environment.
Full Quantization
In CV scenarios where the model running speed needs to be improved and the model running power consumption needs to be reduced, the full quantization after training can be used. The following describes how to use full quantization and its effects.
To calculate a quantization parameter of an activation value, you need to provide a calibration dataset. It is recommended that the calibration dataset be obtained from the actual inference scenario and can represent the actual input of a model. The number of data records is about 100 - 500, and the calibration dataset needs to be processed into the Format of NHWC
.
For image data, currently supports channel pack, normalization, resize, center crop processing. The user can set the corresponding parameter according to the preprocessing operation requirements.
Full quantization config’s info must include [common_quant_param]
, [data_preprocess_param]
, [full_quant_param]
.
Note:
The model calibration data must be co-distributed with the training data, and the Format of the calibration data and the inputs of the exported floating-point model need to be consistent.
The general form of the full quantization conversion command is:
./converter_lite --fmk=ModelType --modelFile=ModelFilePath --outputFile=ConvertedModelPath --configFile=/mindspore/lite/tools/converter/quantizer/config/full_quant.cfg
CPU
The full CPU quantization complete configuration file is shown below:
[common_quant_param]
quant_type=FULL_QUANT
# Full quantization support 8bit
bit_num=8
[data_preprocess_param]
# Calibration dataset path, the format is input_name_1:input_1_dir,input_name_2:input_2_dir
# Full quantification must provide correction dataset
calibrate_path=input_name_1:/mnt/image/input_1_dir,input_name_2:input_2_dir
# Calibration data size
calibrate_size=100
# Input type supports IMAGE or BIN
# When set to IMAGE, the image data will be read
# When set to BIN, the `.bin` binary file will be read
input_type=IMAGE
# The output format of the preprocessed image
# Supports RGB or GRAY or BGR
image_to_format=RGB
# Image normalization
# dst = (src - mean) / std
normalize_mean=[127.5, 127.5, 127.5]
normalize_std=[127.5, 127.5, 127.5]
# Image resize
resize_width=224
resize_height=224
# Resize method supports LINEAR or NEAREST or CUBIC
resize_method=LINEAR
# Image center crop
center_crop_width=224
center_crop_height=224
[full_quant_param]
# Activation quantized method supports MAX_MIN or KL or REMOVAL_OUTLIER
activation_quant_method=MAX_MIN
# Whether to correct the quantization error. Recommended to set to true.
bias_correction=true
Full quantization requires the execution of inference, and the waiting time may be long. If you need to view the log, you can set export GLOG_v=1 before execution for printing the related Info level log.
[full_quant_param]
# Activation quantized method supports MAX_MIN or KL or REMOVAL_OUTLIER
activation_quant_method=MAX_MIN
# Whether to correct the quantization error. Recommended to set to true.
bias_correction=true
# If set to true, it will enable PerChannel quantization, or set to false to enable PerLayer quantization.
per_channel=true
The full quantization parameter (PerLayer quantization type) [full_quant_param]
configuration is as follows:
[full_quant_param]
# Activation quantized method supports MAX_MIN or KL or REMOVAL_OUTLIER
activation_quant_method=MAX_MIN
# Whether to correct the quantization error. Recommended to set to true.
bias_correction=true
# If set to true, it will enable PerChannel quantization, or set to false to enable PerLayer quantization.
per_channel=false
Partial Model Accuracy Result
Model |
Test Dataset |
quant_method |
FP32 Model Accuracy |
Full Quantization Accuracy (8 bits) |
Description |
---|---|---|---|---|---|
KL |
77.60% |
77.40% |
Randomly select 100 images from the ImageNet Validation dataset as a calibration dataset. |
||
KL |
70.96% |
70.31% |
Randomly select 100 images from the ImageNet Validation dataset as a calibration dataset. |
||
MAX_MIN |
71.56% |
71.16% |
Randomly select 100 images from the ImageNet Validation dataset as a calibration dataset. |
All the preceding results are obtained in the x86 environment.
NVDIA
NVIDIA GPU full quantization parameter configuration. Just add a new configuration target_device=NVGPU
to [full_quant_param]
:
[full_quant_param]
# Activation quantized method supports MAX_MIN or KL or REMOVAL_OUTLIER
activation_quant_method=MAX_MIN
# Supports specific hardware backends
target_device=NVGPU
DSP
DSP full quantization parameter configuration only need add target_device=DSP
to [full_quant_param]
, and the command is as follows:
[full_quant_param]
# Activation quantized method supports MAX_MIN or KL or REMOVAL_OUTLIER
activation_quant_method=MAX_MIN
# Whether to correct the quantization error.
bias_correction=false
# Supports specific hardware backends
target_device=DSP
Ascend
Ascend quantization need to set optimize
to ascend_oriented
for converter tools and we also need to set environment for Ascend.
Ascend quantization static shape parameter configuration
In the static shape scenario, the general form of the conversion command for Ascend quantization as follow:
./converter_lite --fmk=ModelType --modelFile=ModelFilePath --outputFile=ConvertedModelPath --configFile=/mindspore/lite/tools/converter/quantizer/config/full_quant.cfg --optimize=ascend_oriented
Ascend static shape quantizion only need add
target_device=ASCEND
to[full_quant_param]
likes as follows:[full_quant_param] # Activation quantized method supports MAX_MIN or KL or REMOVAL_OUTLIER activation_quant_method=MAX_MIN # Whether to correct the quantization error. bias_correction=true # Supports specific hardware backends target_device=ASCEND
Ascend quantization also support dynamic shape. It is worth noting that the conversion command must set the same inputShape as the calibration dataset. For details, please refer to Conversion Tool Parameter Description.
In the dynamic shape scenario, the general form of the conversion command for Ascend quantization as follow:
./converter_lite --fmk=ModelType --modelFile=ModelFilePath --outputFile=ConvertedModelPath --configFile=/mindspore/lite/tools/converter/quantizer/config/full_quant.cfg --optimize=ascend_oriented --inputShape="inTensorName_1:1,32,32,4"
We must add configuration about
[ascend_context]
for dynamic shape as follow:[full_quant_param] # Activation quantized method supports MAX_MIN or KL or REMOVAL_OUTLIER activation_quant_method=MAX_MIN # Whether to correct the quantization error. bias_correction=true # Supports specific hardware backends target_device=ASCEND [ascend_context] input_shape=input_1:[-1,32,32,4] dynamic_dims=[1~4],[8],[16] # "-1" in input_shape indicates that the batch size is dynamic.
Dynamic Quantization
In NLP scenarios where the model running speed needs to be improved and the model running power consumption needs to be reduced, the dynamic quantization after training can be used. The following describes how to use dynamic quantization and its effects.
In dynamic quantization, the weights are quantized at the convert, and the activation are quantized at the runtime. Compared to static quantization, no calibration dataset is required.
The general form of the dynamic quantization conversion command is:
./converter_lite --fmk=ModelType --modelFile=ModelFilePath --outputFile=ConvertedModelPath --configFile=/mindspore/lite/tools/converter/quantizer/config/dynamic_quant.cfg
The dynamic quantization profile is as follows:
[common_quant_param]
quant_type=DYNAMIC_QUANT
bit_num=8
[dynamic_quant_param]
# If set to ALWC, it will enable activation perlayer and weight perchannel quantization. If set to ACWL, it will enable activation perchannel and weight perlayer quantization. Default value is ALWC.
quant_strategy=ACWL
In order to ensure the quantization accuracy, the dynamic quantization does not support setting the FP16 mode .
The dynamic quantization will have a further acceleration effect on the ARM architecture that supports SDOT instructions.
Partial Model Performance Results
tinybert_encoder
Model Type |
Runtime Mode |
Model Size(M) |
RAM(K) |
Latency(ms) |
Cos-Similarity |
Compression Ratio |
Memory compared to FP32 |
Latency compared to FP32 |
---|---|---|---|---|---|---|---|---|
FP32 |
FP32 |
20 |
29,029 |
9.916 |
1 |
|||
FP32 |
FP16 |
20 |
18,208 |
5.75 |
0.99999 |
1 |
-37.28% |
-42.01% |
FP16 |
FP16 |
12 |
18,105 |
5.924 |
0.99999 |
1.66667 |
-37.63% |
-40.26% |
Weight Quant(8 Bit) |
FP16 |
5.3 |
19,324 |
5.764 |
0.99994 |
3.77358 |
-33.43% |
-41.87% |
Dynamic Quant |
INT8+FP32 |
5.2 |
15,709 |
4.517 |
0.99668 |
3.84615 |
-45.89% |
-54.45% |
tinybert_decoder
Model Type |
Runtime Mode |
Model Size(M) |
RAM(K) |
Latency(ms) |
Cos-Similarity |
Compression Ratio |
Memory compared to FP32 |
Latency compared to FP32 |
---|---|---|---|---|---|---|---|---|
FP32 |
FP32 |
43 |
51,355 |
4.161 |
1 |
|||
FP32 |
FP16 |
43 |
29,462 |
2.184 |
0.99999 |
1 |
-42.63% |
-47.51% |
FP16 |
FP16 |
22 |
29,440 |
2.264 |
0.99999 |
1.95455 |
-42.67% |
-45.59% |
Weight Quant(8 Bit) |
FP16 |
12 |
32,285 |
2.307 |
0.99998 |
3.58333 |
-37.13% |
-44.56% |
Dynamic Quant |
INT8+FP32 |
12 |
22,181 |
2.074 |
0.9993 |
3.58333 |
-56.81% |
-50.16% |
Quantization Debug
Turn on the quantization Debug function, you can get the data distribution statistics report, which is used to evaluate the quantization error and assist the decision-making model (operator) whether it is suitable for quantization. For full quantification, N data distribution statistics reports will be generated according to the number of correction datasets provided, that is, one report will be generated for each round; for weighting, only one data distribution statistics report will be generated.
When setting the debug_info_save_path
parameter, the relevant debug report will be generated in the /home/workspace/mindspore/debug_info_save_path
folder:
[common_quant_param]
debug_info_save_path=/home/workspace/mindspore/debug_info_save_path
The quantized output summary report output_summary.csv
contains the accuracy information of the output layer Tensor of the entire network. The related fields are as follows:
Type |
Name |
---|---|
Round |
The calibration training round |
TensorName |
The tensor name |
CosineSimilarity |
Cosine similarity compared with the original data |
The data distribution statistics report report_*.csv
will count the original data distribution of each Tensor and the data distribution after dequantization of the quantized Tensor. The relevant fields of the data distribution statistics report are as follows:
Type |
Name |
---|---|
NodeName |
The node name |
NodeType |
The node type |
TensorName |
The tensor name |
InOutFlag |
The input or output tensor |
DataTypeFlag |
The data type, use Origin for original model, use Dequant for quantization model |
TensorTypeFlag |
The data types such as input and output, use Activation, and constants, etc., use Weight. |
Min |
The minimum value |
Q1 |
The 25% quantile |
Median |
The median |
Q3 |
The 75% quantile |
MAX |
The maximum |
Mean |
The mean |
Var |
The var |
Sparsity |
The sparsity |
Clip |
The Clip |
CosineSimilarity |
Cosine similarity compared with the original data |
The quantization parameter report quant_param.csv
contains the quantization parameter information of all quantized Tensors. The related fields of the quantization parameter are as follows:
Type |
Name |
---|---|
NodeName |
The node name |
NodeType |
The node type |
TensorName |
The tensor name |
ElementsNum |
The Tensor elements num |
Dims |
The tensor dims |
Scale |
The quantization parameter scale |
ZeroPoint |
The quantization parameter zeropoint |
Bits |
The number of quantization bits |
CorrectionVar |
Bias correction coefficient-variance |
CorrectionMean |
Bias correction coefficient-mean |
Mixed bit quantization is non-standard quantization, the quantization parameter file may not exist.
Skipping Quantization Node
Quantization is to convert the float32 operator to the int8 operator. The current quantization strategy is to quantify all the nodes contained in a certain type of operator that can be supported, but there are some nodes that are more sensitive and will cause larger errors after quantization. At the same time, the inference speed of some layers after quantization is much lower than that of float16. It supports non-quantization of the specified layer, which can effectively improve the accuracy and inference speed.
Below is an example of conv2d_1
add_8
concat_1
without quantifying the three nodes:
[common_quant_param]
# Supports WEIGHT_QUANT or FULL_QUANT
quant_type=FULL_QUANT
# Weight quantization support the number of bits [0,16], Set to 0 is mixed bit quantization, otherwise it is fixed bit quantization
# Full quantization support 8bit
bit_num=8
# Set the name of the operator that skips the quantization, and use `,` to split between multiple operators.
skip_quant_node=conv2d_1,add_8,concat_1
Recommendations
By filtering
InOutFlag == Output && DataTypeFlag == Dequant
, the output layer of all quantization operators can be filtered out, and the accuracy loss of the operator can be judged by looking at the quantized outputCosineSimilarity
, the closer to 1 the smaller the loss.For merging operators such as Add and Concat, if the distribution of
min
andmax
between different input Tensors is quite different, which is likely to cause large errors, you can setskip_quant_node
to not quantize them.For operators with a higher cutoff rate
Clip
, you can setskip_quant_node
to not quantize it.