Quantize NLP models with OpenVINO Post-Training Optimization Tool ¶
This tutorial demonstrates how to apply INT8 quantization to the Natural Language Processing model known as BERT, using the Post-Training Optimization Tool API (part of the OpenVINO Toolkit). We will use a fine-tuned HuggingFace BERT PyTorch model trained on the Microsoft Research Paraphrase Corpus (MRPC). The tutorial is designed to be extendable to custom models and datasets. It consists of the following steps:
Download and prepare the BERT model and MRPC dataset
Define data loading and accuracy validation functionality
Prepare the model for quantization
Run optimization pipeline
Load and test quantized model
Compare the performance of the original, converted and quantized models
Imports¶
import os
import sys
import time
import warnings
from pathlib import Path
from zipfile import ZipFile
import numpy as np
import torch
from addict import Dict
from compression.api import DataLoader as POTDataLoader
from compression.api import Metric
from compression.engines.ie_engine import IEEngine
from compression.graph import load_model, save_model
from compression.graph.model_utils import compress_model_weights
from compression.pipeline.initializer import create_pipeline
from openvino import runtime as ov
from torch.utils.data import TensorDataset
from transformers import BertForSequenceClassification, BertTokenizer
from transformers import (
glue_convert_examples_to_features as convert_examples_to_features,
)
from transformers import glue_output_modes as output_modes
from transformers import glue_processors as processors
sys.path.append("../utils")
from notebook_utils import download_file
Settings¶
# Set the data and model directories, model source URL and model filename
DATA_DIR = "data"
MODEL_DIR = "model"
MODEL_LINK = "https://download.pytorch.org/tutorial/MRPC.zip"
FILE_NAME = MODEL_LINK.split("/")[-1]
os.makedirs(DATA_DIR, exist_ok=True)
os.makedirs(MODEL_DIR, exist_ok=True)
Prepare the Model¶
Next steps include: - Download and unpack pre-trained BERT model for MRPC by PyTorch - Convert model to ONNX - Run OpenVINO Model Optimizer tool to convert the model from the ONNX representation to the OpenVINO Intermediate Representation (IR)
download_file(MODEL_LINK, directory=MODEL_DIR, show_progress=True)
with ZipFile(f"{MODEL_DIR}/{FILE_NAME}", "r") as zip_ref:
zip_ref.extractall(MODEL_DIR)
model/MRPC.zip: 0%| | 0.00/387M [00:00<?, ?B/s]
Import all dependencies to load the original PyTorch model and convert it to the ONNX representation.
BATCH_SIZE = 1
MAX_SEQ_LENGTH = 128
def export_model_to_onnx(model, path):
with torch.no_grad():
default_input = torch.ones(1, MAX_SEQ_LENGTH, dtype=torch.int64)
inputs = {
"input_ids": default_input,
"attention_mask": default_input,
"token_type_ids": default_input,
}
symbolic_names = {0: "batch_size", 1: "max_seq_len"}
torch.onnx.export(
model,
(inputs["input_ids"], inputs["attention_mask"], inputs["token_type_ids"]),
path,
opset_version=11,
do_constant_folding=True,
input_names=["input_ids", "input_mask", "segment_ids"],
output_names=["output"],
dynamic_axes={
"input_ids": symbolic_names,
"input_mask": symbolic_names,
"segment_ids": symbolic_names,
},
)
print("ONNX model saved to {}".format(path))
torch_model = BertForSequenceClassification.from_pretrained(os.path.join(MODEL_DIR, "MRPC"))
onnx_model_path = Path(MODEL_DIR) / "bert_mrpc.onnx"
if not onnx_model_path.exists():
export_model_to_onnx(torch_model, onnx_model_path)
ONNX model saved to model/bert_mrpc.onnx
Convert the ONNX Model to OpenVINO IR¶
ir_model_xml = onnx_model_path.with_suffix(".xml")
ir_model_bin = onnx_model_path.with_suffix(".bin")
# convert ONNX model to IR FP32
if not ir_model_xml.exists():
!mo --input_model $onnx_model_path --output_dir $MODEL_DIR --model_name $ir_model_xml.stem --input input_ids,input_mask,segment_ids --input_shape [1,128],[1,128],[1,128] --output output --data_type FP32
Model Optimizer arguments:
Common parameters:
- Path to the Input Model: /home/runner/work/openvino_notebooks/openvino_notebooks/notebooks/105-language-quantize-bert/model/bert_mrpc.onnx
- Path for generated IR: /home/runner/work/openvino_notebooks/openvino_notebooks/notebooks/105-language-quantize-bert/model
- IR output name: bert_mrpc
- Log level: ERROR
- Batch: Not specified, inherited from the model
- Input layers: input_ids,input_mask,segment_ids
- Output layers: output
- Input shapes: [1,128],[1,128],[1,128]
- Source layout: Not specified
- Target layout: Not specified
- Layout: Not specified
- Mean values: Not specified
- Scale values: Not specified
- Scale factor: Not specified
- Precision of IR: FP32
- Enable fusing: True
- User transformations: Not specified
- Reverse input channels: False
- Enable IR generation for fixed input shape: False
- Use the transformations config file: None
Advanced parameters:
- Force the usage of legacy Frontend of Model Optimizer for model conversion into IR: False
- Force the usage of new Frontend of Model Optimizer for model conversion into IR: False
OpenVINO runtime found in: /opt/hostedtoolcache/Python/3.8.12/x64/lib/python3.8/site-packages/openvino
OpenVINO runtime version: 2022.1.0-7019-cdb9bec7210-releases/2022/1
Model Optimizer version: 2022.1.0-7019-cdb9bec7210-releases/2022/1
[ SUCCESS ] Generated IR version 11 model.
[ SUCCESS ] XML file: /home/runner/work/openvino_notebooks/openvino_notebooks/notebooks/105-language-quantize-bert/model/bert_mrpc.xml
[ SUCCESS ] BIN file: /home/runner/work/openvino_notebooks/openvino_notebooks/notebooks/105-language-quantize-bert/model/bert_mrpc.bin
[ SUCCESS ] Total execution time: 3.25 seconds.
[ SUCCESS ] Memory consumed: 1024 MB.
It's been a while, check for a new version of Intel(R) Distribution of OpenVINO(TM) toolkit here https://software.intel.com/content/www/us/en/develop/tools/openvino-toolkit/download.html?cid=other&source=prod&campid=ww_2022_bu_IOTG_OpenVINO-2022-1&content=upg_all&medium=organic or on the GitHub*
[ INFO ] The model was converted to IR v11, the latest model format that corresponds to the source DL framework input/output format. While IR v11 is backwards compatible with OpenVINO Inference Engine API v1.0, please use API v2.0 (as of 2022.1) to take advantage of the latest improvements in IR v11.
Find more information about API v2.0 and IR v11 at https://docs.openvino.ai
Prepare MRPC Task Dataset¶
To run this tutorial, you will need to download the General Language Understanding Evaluation (GLUE) data for the MRPC task from HuggingFace. The code below will download a script that fetches the MRPC dataset.
download_file(
"https://raw.githubusercontent.com/huggingface/transformers/f98ef14d161d7bcdc9808b5ec399981481411cc1/utils/download_glue_data.py",
show_progress=False,
)
PosixPath('/home/runner/work/openvino_notebooks/openvino_notebooks/notebooks/105-language-quantize-bert/download_glue_data.py')
from download_glue_data import format_mrpc
format_mrpc(DATA_DIR, "")
Processing MRPC...
Local MRPC data not specified, downloading data from https://dl.fbaipublicfiles.com/senteval/senteval_data/msr_paraphrase_train.txt
Completed!
Define DataLoader for POT¶
In this step, we need to define DataLoader based on POT API. It will
be used to collect statistics for quantization and run model evaluation.
We use helper functions from the HuggingFace Transformers to do the data
preprocessing. It takes raw text data and encodes sentences and words
producing three model inputs. For more details about the data
preprocessing and tokenization please refer to this
description.
class MRPCDataLoader(POTDataLoader):
# Required methods
def __init__(self, config):
"""Constructor
:param config: data loader specific config
"""
if not isinstance(config, Dict):
config = Dict(config)
super().__init__(config)
self._task = config["task"].lower()
self._model_dir = config["model_dir"]
self._data_dir = config["data_source"]
self._batch_size = config["batch_size"]
self._max_length = config["max_length"]
self.examples = []
self._prepare_dataset()
def __len__(self):
"""Returns size of the dataset"""
return len(self.dataset)
def __getitem__(self, index):
"""
Returns annotation, data and metadata at the specified index.
Possible formats:
(index, annotation), data
(index, annotation), data, metadata
"""
if index >= len(self):
raise IndexError
batch = self.dataset[index]
batch = tuple(t.detach().cpu().numpy() for t in batch)
inputs = {"input_ids": batch[0], "input_mask": batch[1], "segment_ids": batch[2]}
labels = batch[3]
return (index, labels), inputs
# Methods specific to the current implementation
def _prepare_dataset(self):
"""Prepare dataset"""
tokenizer = BertTokenizer.from_pretrained(self._model_dir, do_lower_case=True)
processor = processors[self._task]()
output_mode = output_modes[self._task]
label_list = processor.get_labels()
examples = processor.get_dev_examples(self._data_dir)
features = convert_examples_to_features(
examples,
tokenizer,
label_list=label_list,
max_length=self._max_length,
output_mode=output_mode,
)
all_input_ids = torch.unsqueeze(torch.tensor([f.input_ids for f in features], dtype=torch.long), 1)
all_attention_mask = torch.unsqueeze(torch.tensor([f.attention_mask for f in features], dtype=torch.long), 1)
all_token_type_ids = torch.unsqueeze(torch.tensor([f.token_type_ids for f in features], dtype=torch.long), 1)
all_labels = torch.unsqueeze(torch.tensor([f.label for f in features], dtype=torch.long), 1)
self.dataset = TensorDataset(
all_input_ids, all_attention_mask, all_token_type_ids, all_labels
)
self.examples = examples
Define Accuracy Metric Calculation¶
At this step the Metric interface for MRPC task metrics is
implemented. It is used for validating the accuracy of the models.
class Accuracy(Metric):
# Required methods
def __init__(self):
super().__init__()
self._name = "Accuracy"
self._matches = []
@property
def value(self):
"""Returns accuracy metric value for the last model output."""
return {self._name: self._matches[-1]}
@property
def avg_value(self):
"""Returns accuracy metric value for all model outputs."""
return {self._name: np.ravel(self._matches).mean()}
def update(self, output, target):
"""
Updates prediction matches.
:param output: model output
:param target: annotations
"""
if len(output) > 1:
raise Exception(
"The accuracy metric cannot be calculated " "for a model with multiple outputs"
)
output = np.argmax(output)
match = output == target[0]
self._matches.append(match)
def reset(self):
"""
Resets collected matches
"""
self._matches = []
def get_attributes(self):
"""
Returns a dictionary of metric attributes {metric_name: {attribute_name: value}}.
Required attributes: 'direction': 'higher-better' or 'higher-worse'
'type': metric type
"""
return {self._name: {"direction": "higher-better", "type": "accuracy"}}
Run Quantization Pipeline¶
Here we define a configuration for our quantization pipeline and run it.
Please note that we use built-in IEEngine implementation of
Engine interface from the POT API for model inference.
warnings.filterwarnings("ignore") # Suppress accuracychecker warnings
model_config = Dict({"model_name": "bert_mrpc", "model": ir_model_xml, "weights": ir_model_bin})
engine_config = Dict({"device": "CPU"})
dataset_config = {
"task": "mrpc",
"data_source": os.path.join(DATA_DIR, "MRPC"),
"model_dir": os.path.join(MODEL_DIR, "MRPC"),
"batch_size": BATCH_SIZE,
"max_length": MAX_SEQ_LENGTH,
}
algorithms = [
{
"name": "DefaultQuantization",
"params": {
"target_device": "ANY",
"model_type": "transformer",
"preset": "performance",
"stat_subset_size": 250,
},
}
]
# Step 1: Load the model
model = load_model(model_config=model_config)
# Step 2: Initialize the data loader
data_loader = MRPCDataLoader(config=dataset_config)
# Step 3 (Optional. Required for AccuracyAwareQuantization): Initialize the metric
metric = Accuracy()
# Step 4: Initialize the engine for metric calculation and statistics collection
engine = IEEngine(config=engine_config, data_loader=data_loader, metric=metric)
# Step 5: Create a pipeline of compression algorithms
pipeline = create_pipeline(algo_config=algorithms, engine=engine)
# Step 6 (Optional): Evaluate the original model. Print the results
fp_results = pipeline.evaluate(model=model)
if fp_results:
print("FP32 model results:")
for name, value in fp_results.items():
print(f"{name}: {value:.5f}")
FP32 model results:
Accuracy: 0.86029
# Step 7: Execute the pipeline
warnings.filterwarnings("ignore") # Suppress accuracychecker warnings
print(
f"Quantizing model with {algorithms[0]['params']['preset']} preset and {algorithms[0]['name']}"
)
start_time = time.perf_counter()
compressed_model = pipeline.run(model=model)
end_time = time.perf_counter()
print(f"Quantization finished in {end_time - start_time:.2f} seconds")
# Step 8 (Optional): Compress model weights to quantized precision
# in order to reduce the size of final .bin file
compress_model_weights(model=compressed_model)
# Step 9: Save the compressed model to the desired path
compressed_model_paths = save_model(model=compressed_model, save_path=MODEL_DIR, model_name="quantized_bert_mrpc"
)
compressed_model_xml = compressed_model_paths[0]["model"]
Quantizing model with performance preset and DefaultQuantization
Quantization finished in 117.36 seconds
# Step 10 (Optional): Evaluate the compressed model and print the results
int_results = pipeline.evaluate(model=compressed_model)
if int_results:
print("INT8 model results:")
for name, value in int_results.items():
print(f"{name}: {value:.5f}")
INT8 model results:
Accuracy: 0.85784
Load and Test OpenVINO Model¶
We will load and test converted model. The next steps include: * load model and compile for CPU * prepare the input * run the inference * get the answer from model output
core = ov.Core()
# read the model from files
model = core.read_model(model=compressed_model_xml)
# assign dynamic shapes to every input layer
for input_layer in model.inputs:
input_shape = input_layer.partial_shape
input_shape[1] = -1
model.reshape({input_layer: input_shape})
# compile model for specific device
compiled_model_int8 = core.compile_model(model=model, device_name="CPU")
output_layer = compiled_model_int8.outputs[0]
The Data Loader returns a pair of sentences (indicated by
sample_idx) and the inference compares these sentences and outputs
whether their meaning is the same. You can test other sentences by
changing sample_idx to another value (from 0 to 407).
sample_idx = 5
sample = data_loader.examples[sample_idx]
inputs = data_loader[sample_idx][1]
result = compiled_model_int8(inputs)[output_layer]
result = np.argmax(result)
print(f"Text 1: {sample.text_a}")
print(f"Text 2: {sample.text_b}")
print(f"The same meaning: {'yes' if result == 1 else 'no'}")
Text 1: Wal-Mart said it would check all of its million-plus domestic workers to ensure they were legally employed .
Text 2: It has also said it would review all of its domestic employees more than 1 million to ensure they have legal status .
The same meaning: yes
Compare Performance of the Original, Converted and Quantized Models¶
To see the performance difference we will compare the original PyTorch model with OpenVINO converted and quantized models (FP32, INT8). We use sentences per second (SPS) measure, which is the same as frames per second (FPS) for images.
model = core.read_model(model=ir_model_xml)
# assign dynamic shapes to every input layer
for input_layer in model.inputs:
input_shape = input_layer.partial_shape
input_shape[1] = -1
model.reshape({input_layer: input_shape})
# compile model for specific device
compiled_model_fp32 = core.compile_model(model=model, device_name="CPU")
num_samples = 50
inputs = data_loader[0][1]
with torch.no_grad():
start = time.perf_counter()
for _ in range(num_samples):
torch_model(torch.as_tensor(list(inputs.values())).squeeze())
end = time.perf_counter()
time_torch = end - start
print(
f"PyTorch model on CPU: {time_torch / num_samples:.3f} seconds per sentence, "
f"SPS: {num_samples / time_torch:.2f}"
)
start = time.perf_counter()
for _ in range(num_samples):
compiled_model_fp32(inputs)
end = time.perf_counter()
time_ir = end - start
print(
f"IR FP32 model in Inference Engine/CPU: {time_ir / num_samples:.3f} "
f"seconds per sentence, SPS: {num_samples / time_ir:.2f}"
)
start = time.perf_counter()
for _ in range(num_samples):
compiled_model_int8(inputs)
end = time.perf_counter()
time_ir = end - start
print(
f"IR INT8 model in Inference Engine/CPU: {time_ir / num_samples:.3f} "
f"seconds per sentence, SPS: {num_samples / time_ir:.2f}"
)
PyTorch model on CPU: 0.284 seconds per sentence, SPS: 3.52
IR FP32 model in Inference Engine/CPU: 0.102 seconds per sentence, SPS: 9.83
IR INT8 model in Inference Engine/CPU: 0.074 seconds per sentence, SPS: 13.57
Finally, we will measure the inference performance of OpenVINO FP32 and INT8 models. To do this, we use Benchmark Tool - OpenVINO’s inference performance measurement tool.
NOTE:
benchmark_appis able to measure the performance of the Intermediate Representation (IR) models only. For more accurate performance, we recommended runningbenchmark_appin a terminal/command prompt after closing other applications. Runbenchmark_app -m model.xml -d CPUto benchmark async inference on CPU for one minute. ChangeCPUtoGPUto benchmark on GPU. Runbenchmark_app --helpto see an overview of all command line options.
# Inference FP32 model (IR)
! benchmark_app -m $ir_model_xml -d CPU -api sync
[Step 1/11] Parsing and validating input arguments
[ WARNING ] -nstreams default value is determined automatically for a device. Although the automatic selection usually provides a reasonable performance, but it still may be non-optimal for some cases, for more information look at README.
[Step 2/11] Loading OpenVINO
[ WARNING ] PerformanceMode was not explicitly specified in command line. Device CPU performance hint will be set to LATENCY.
[ INFO ] OpenVINO:
API version............. 2022.1.0-7019-cdb9bec7210-releases/2022/1
[ INFO ] Device info
CPU
openvino_intel_cpu_plugin version 2022.1
Build................... 2022.1.0-7019-cdb9bec7210-releases/2022/1
[Step 3/11] Setting device configuration
[Step 4/11] Reading network files
[ INFO ] Read model took 232.25 ms
[Step 5/11] Resizing network to match image sizes and given batch
[ INFO ] Network batch size: 1
[Step 6/11] Configuring input of the model
[ INFO ] Model input 'input_ids' precision i64, dimensions ([...]): 1 128
[ INFO ] Model input 'input_mask' precision i64, dimensions ([...]): 1 128
[ INFO ] Model input 'segment_ids' precision i64, dimensions ([...]): 1 128
[ INFO ] Model output 'output' precision f32, dimensions ([...]): 1 2
[Step 7/11] Loading the model to the device
[ INFO ] Compile model took 422.10 ms
[Step 8/11] Querying optimal runtime parameters
[ INFO ] DEVICE: CPU
[ INFO ] AVAILABLE_DEVICES , ['']
[ INFO ] RANGE_FOR_ASYNC_INFER_REQUESTS , (1, 1, 1)
[ INFO ] RANGE_FOR_STREAMS , (1, 2)
[ INFO ] FULL_DEVICE_NAME , Intel(R) Xeon(R) Platinum 8272CL CPU @ 2.60GHz
[ INFO ] OPTIMIZATION_CAPABILITIES , ['WINOGRAD', 'FP32', 'FP16', 'INT8', 'BIN', 'EXPORT_IMPORT']
[ INFO ] CACHE_DIR ,
[ INFO ] NUM_STREAMS , 1
[ INFO ] INFERENCE_NUM_THREADS , 0
[ INFO ] PERF_COUNT , False
[ INFO ] PERFORMANCE_HINT_NUM_REQUESTS , 0
[Step 9/11] Creating infer requests and preparing input data
[ INFO ] Create 1 infer requests took 0.14 ms
[ WARNING ] No input files were given for input 'input_ids'!. This input will be filled with random values!
[ WARNING ] No input files were given for input 'input_mask'!. This input will be filled with random values!
[ WARNING ] No input files were given for input 'segment_ids'!. This input will be filled with random values!
[ INFO ] Fill input 'input_ids' with random values
[ INFO ] Fill input 'input_mask' with random values
[ INFO ] Fill input 'segment_ids' with random values
[Step 10/11] Measuring performance (Start inference synchronously, inference only: True, limits: 60000 ms duration)
[ INFO ] Benchmarking in inference only mode (inputs filling are not included in measurement loop).
[ INFO ] First inference took 100.79 ms
[Step 11/11] Dumping statistics report
Count: 616 iterations
Duration: 60080.51 ms
Latency:
Median: 96.31 ms
AVG: 97.43 ms
MIN: 93.80 ms
MAX: 221.52 ms
Throughput: 10.38 FPS
# Inference INT8 model (IR)
! benchmark_app -m $compressed_model_xml -d CPU -api sync
[Step 1/11] Parsing and validating input arguments
[ WARNING ] -nstreams default value is determined automatically for a device. Although the automatic selection usually provides a reasonable performance, but it still may be non-optimal for some cases, for more information look at README.
[Step 2/11] Loading OpenVINO
[ WARNING ] PerformanceMode was not explicitly specified in command line. Device CPU performance hint will be set to LATENCY.
[ INFO ] OpenVINO:
API version............. 2022.1.0-7019-cdb9bec7210-releases/2022/1
[ INFO ] Device info
CPU
openvino_intel_cpu_plugin version 2022.1
Build................... 2022.1.0-7019-cdb9bec7210-releases/2022/1
[Step 3/11] Setting device configuration
[Step 4/11] Reading network files
[ INFO ] Read model took 116.27 ms
[Step 5/11] Resizing network to match image sizes and given batch
[ INFO ] Network batch size: 1
[Step 6/11] Configuring input of the model
[ INFO ] Model input 'input_ids' precision i64, dimensions ([...]): 1 128
[ INFO ] Model input 'input_mask' precision i64, dimensions ([...]): 1 128
[ INFO ] Model input 'segment_ids' precision i64, dimensions ([...]): 1 128
[ INFO ] Model output 'output' precision f32, dimensions ([...]): 1 2
[Step 7/11] Loading the model to the device
[ INFO ] Compile model took 270.87 ms
[Step 8/11] Querying optimal runtime parameters
[ INFO ] DEVICE: CPU
[ INFO ] AVAILABLE_DEVICES , ['']
[ INFO ] RANGE_FOR_ASYNC_INFER_REQUESTS , (1, 1, 1)
[ INFO ] RANGE_FOR_STREAMS , (1, 2)
[ INFO ] FULL_DEVICE_NAME , Intel(R) Xeon(R) Platinum 8272CL CPU @ 2.60GHz
[ INFO ] OPTIMIZATION_CAPABILITIES , ['WINOGRAD', 'FP32', 'FP16', 'INT8', 'BIN', 'EXPORT_IMPORT']
[ INFO ] CACHE_DIR ,
[ INFO ] NUM_STREAMS , 1
[ INFO ] INFERENCE_NUM_THREADS , 0
[ INFO ] PERF_COUNT , False
[ INFO ] PERFORMANCE_HINT_NUM_REQUESTS , 0
[Step 9/11] Creating infer requests and preparing input data
[ INFO ] Create 1 infer requests took 0.18 ms
[ WARNING ] No input files were given for input 'input_ids'!. This input will be filled with random values!
[ WARNING ] No input files were given for input 'input_mask'!. This input will be filled with random values!
[ WARNING ] No input files were given for input 'segment_ids'!. This input will be filled with random values!
[ INFO ] Fill input 'input_ids' with random values
[ INFO ] Fill input 'input_mask' with random values
[ INFO ] Fill input 'segment_ids' with random values
[Step 10/11] Measuring performance (Start inference synchronously, inference only: True, limits: 60000 ms duration)
[ INFO ] Benchmarking in inference only mode (inputs filling are not included in measurement loop).
[ INFO ] First inference took 84.54 ms
[Step 11/11] Dumping statistics report
Count: 853 iterations
Duration: 60010.94 ms
Latency:
Median: 69.95 ms
AVG: 70.26 ms
MIN: 67.90 ms
MAX: 105.54 ms
Throughput: 14.30 FPS