Quantize NLP models with Post-Training Quantization in NNCF¶

This tutorial is also available as a Jupyter notebook that can be cloned directly from GitHub. See the installation guide for instructions to run this tutorial locally on Windows, Linux or macOS.

This tutorial demonstrates how to apply INT8 quantization to the Natural Language Processing model known as BERT, using the Post-Training Quantization API (NNCF library). A fine-tuned HuggingFace BERT PyTorch model, trained on the Microsoft Research Paraphrase Corpus (MRPC), will be used. 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.

!pip install -q nncf datasets evaluate

Imports¶

import os
import sys
import time
from pathlib import Path
from zipfile import ZipFile
from typing import Iterable
from typing import Any

import numpy as np
import torch
from openvino import runtime as ov
from openvino.tools import mo
from openvino.runtime import serialize, Model
import nncf
from nncf.parameters import ModelType
from transformers import BertForSequenceClassification, BertTokenizer
import datasets
import evaluate

sys.path.append("../utils")
from notebook_utils import download_file

/opt/home/k8sworker/cibuilds/ov-notebook/OVNotebookOps-408/.workspace/scm/ov-notebook/.venv/lib/python3.8/site-packages/openvino/offline_transformations/__init__.py:10: FutureWarning: The module is private and following namespace offline_transformations will be removed in the future, use openvino.runtime.passes instead!
  warnings.warn(

INFO:nncf:NNCF initialized successfully. Supported frameworks detected: torch, tensorflow, onnx, openvino

Settings¶

# Set the data and model directories, source URL and the filename of the model.
DATA_DIR = "data"
MODEL_DIR = "model"
MODEL_LINK = "https://download.pytorch.org/tutorial/MRPC.zip"
FILE_NAME = MODEL_LINK.split("/")[-1]
PRETRAINED_MODEL_DIR = os.path.join(MODEL_DIR, "MRPC")

os.makedirs(DATA_DIR, exist_ok=True)
os.makedirs(MODEL_DIR, exist_ok=True)

Prepare the Model¶

Perform the following: - Download and unpack pre-trained BERT model for MRPC by PyTorch. - Convert the model to the ONNX. - Run Model Optimizer to convert the model from the ONNX representation to the OpenVINO Intermediate Representation (OpenVINO 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]

Convert the original PyTorch model 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,
        }
        torch.onnx.export(
            model,
            (inputs["input_ids"], inputs["attention_mask"], inputs["token_type_ids"]),
            path,
            opset_version=11,
            input_names=["input_ids", "attention_mask", "token_type_ids"],
            output_names=["output"]
        )
        print("ONNX model saved to {}".format(path))


torch_model = BertForSequenceClassification.from_pretrained(PRETRAINED_MODEL_DIR)
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¶

Use Model Optimizer Python API to convert the model to OpenVINO IR with FP32 precision. For more information about Model Optimizer Python API, see the Model Optimizer Developer Guide.

ir_model_xml = onnx_model_path.with_suffix(".xml")

# Convert the ONNX model to OpenVINO IR FP32.
if not ir_model_xml.exists():
    model = mo.convert_model(onnx_model_path)
    serialize(model, str(ir_model_xml))

Prepare the Dataset¶

We download the General Language Understanding Evaluation (GLUE) dataset for the MRPC task from HuggingFace datasets. Then, we tokenize the data with a pre-trained BERT tokenizer from HuggingFace.

def create_data_source():
    raw_dataset = datasets.load_dataset('glue', 'mrpc', split='validation')
    tokenizer = BertTokenizer.from_pretrained(PRETRAINED_MODEL_DIR)

    def _preprocess_fn(examples):
        texts = (examples['sentence1'], examples['sentence2'])
        result = tokenizer(*texts, padding='max_length', max_length=MAX_SEQ_LENGTH, truncation=True)
        result['labels'] = examples['label']
        return result
    processed_dataset = raw_dataset.map(_preprocess_fn, batched=True, batch_size=1)

    return processed_dataset


data_source = create_data_source()

[ WARNING ] Found cached dataset glue (/opt/home/k8sworker/.cache/huggingface/datasets/glue/mrpc/1.0.0/dacbe3125aa31d7f70367a07a8a9e72a5a0bfeb5fc42e75c9db75b96da6053ad)

[ WARNING ]  Found cached dataset glue (/opt/home/k8sworker/.cache/huggingface/datasets/glue/mrpc/1.0.0/dacbe3125aa31d7f70367a07a8a9e72a5a0bfeb5fc42e75c9db75b96da6053ad)

[ WARNING ] Loading cached processed dataset at /opt/home/k8sworker/.cache/huggingface/datasets/glue/mrpc/1.0.0/dacbe3125aa31d7f70367a07a8a9e72a5a0bfeb5fc42e75c9db75b96da6053ad/cache-b5f4c739eb2a4a9f.arrow

[ WARNING ]  Loading cached processed dataset at /opt/home/k8sworker/.cache/huggingface/datasets/glue/mrpc/1.0.0/dacbe3125aa31d7f70367a07a8a9e72a5a0bfeb5fc42e75c9db75b96da6053ad/cache-b5f4c739eb2a4a9f.arrow

Optimize model using NNCF Post-training Quantization API¶

NNCF provides a suite of advanced algorithms for Neural Networks inference optimization in OpenVINO with minimal accuracy drop. We will use 8-bit quantization in post-training mode (without the fine-tuning pipeline) to optimize BERT.

Note: NNCF Post-training Quantization is available as a preview feature in OpenVINO 2022.3 release. Fully functional support will be provided in the next releases.

The optimization process contains the following steps:

Create a Dataset for quantization
Run nncf.quantize for getting an optimized model
Serialize OpenVINO IR model using openvino.runtime.serialize function

# Load the network in OpenVINO Runtime.
core = ov.Core()
model = core.read_model(ir_model_xml)
INPUT_NAMES = [x.any_name for x in model.inputs]


def transform_fn(data_item):
    """
    Extract the model's input from the data item.
    The data item here is the data item that is returned from the data source per iteration.
    This function should be passed when the data item cannot be used as model's input.
    """
    inputs = {
        name: np.asarray(data_item[name], dtype=np.int64) for name in INPUT_NAMES
    }
    return inputs


calibration_dataset = nncf.Dataset(data_source, transform_fn)
# Quantize the model. By specifying model_type, we specify additional transformer patterns in the model.
quantized_model = nncf.quantize(model, calibration_dataset,
                                model_type=ModelType.TRANSFORMER)

INFO:openvino.tools.pot.pipeline.pipeline:Inference Engine version:                2022.3.0-9052-9752fafe8eb-releases/2022/3
INFO:openvino.tools.pot.pipeline.pipeline:Model Optimizer version:                 2022.3.0-9052-9752fafe8eb-releases/2022/3
INFO:openvino.tools.pot.pipeline.pipeline:Post-Training Optimization Tool version: 2022.3.0-9052-9752fafe8eb-releases/2022/3
INFO:openvino.tools.pot.statistics.collector:Start computing statistics for algorithms : DefaultQuantization
INFO:openvino.tools.pot.statistics.collector:Computing statistics finished
INFO:openvino.tools.pot.pipeline.pipeline:Start algorithm: DefaultQuantization
INFO:openvino.tools.pot.algorithms.quantization.default.algorithm:Start computing statistics for algorithm : ActivationChannelAlignment
INFO:openvino.tools.pot.algorithms.quantization.default.algorithm:Computing statistics finished
INFO:openvino.tools.pot.algorithms.quantization.default.algorithm:Start computing statistics for algorithms : MinMaxQuantization,FastBiasCorrection
INFO:openvino.tools.pot.algorithms.quantization.default.algorithm:Computing statistics finished
INFO:openvino.tools.pot.pipeline.pipeline:Finished: DefaultQuantization
 ===========================================================================

compressed_model_xml = 'quantized_bert_mrpc.xml'
ov.serialize(quantized_model, compressed_model_xml)

Load and Test OpenVINO Model¶

To load and test converted model, perform the following: * Load the model and compile it for CPU. * Prepare the input. * Run the inference. * Get the answer from the 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 the model for a specific device.
compiled_model_int8 = core.compile_model(model=model, device_name="CPU")

output_layer = compiled_model_int8.outputs[0]

The Data Source 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_source[sample_idx]
inputs = {k: torch.unsqueeze(torch.tensor(sample[k]), 0) for k in ['input_ids', 'token_type_ids', 'attention_mask']}

result = compiled_model_int8(inputs)[output_layer]
result = np.argmax(result)

print(f"Text 1: {sample['sentence1']}")
print(f"Text 2: {sample['sentence2']}")
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 F1-score of FP32 and INT8 models¶

def validate(model: Model, dataset: Iterable[Any]) -> float:
    """
    Evaluate the model on GLUE dataset.
    Returns F1 score metric.
    """
    compiled_model = core.compile_model(model, device_name='CPU')
    output_layer = compiled_model.output(0)

    metric = evaluate.load('glue', 'mrpc')
    INPUT_NAMES = [x.any_name for x in compiled_model.inputs]
    for batch in dataset:
        inputs = [
            np.expand_dims(np.asarray(batch[key], dtype=np.int64), 0) for key in INPUT_NAMES
        ]
        outputs = compiled_model(inputs)[output_layer]
        predictions = outputs[0].argmax(axis=-1)
        metric.add_batch(predictions=[predictions], references=[batch['labels']])
    metrics = metric.compute()
    f1_score = metrics['f1']

    return f1_score


print('Checking the accuracy of the original model:')
metric = validate(model, data_source)
print(f'F1 score: {metric:.4f}')

print('Checking the accuracy of the quantized model:')
metric = validate(quantized_model, data_source)
print(f'F1 score: {metric:.4f}')

Checking the accuracy of the original model:
F1 score: 0.8927
Checking the accuracy of the quantized model:
F1 score: 0.9014

Compare Performance of the Original, Converted and Quantized Models¶

Compare the original PyTorch model with OpenVINO converted and quantized models (FP32, INT8) to see the difference in performance. It is expressed in 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.
dynamic_shapes = {}
for input_layer in model.inputs:
    input_shape = input_layer.partial_shape
    input_shape[1] = -1

    dynamic_shapes[input_layer] = input_shape

model.reshape(dynamic_shapes)

# Compile the model for a specific device.
compiled_model_fp32 = core.compile_model(model=model, device_name="CPU")

num_samples = 50
sample = data_source[0]
inputs = {k: torch.unsqueeze(torch.tensor(sample[k]), 0) for k in ['input_ids', 'token_type_ids', 'attention_mask']}

with torch.no_grad():
    start = time.perf_counter()
    for _ in range(num_samples):
        torch_model(torch.vstack(list(inputs.values())))
    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 OpenVINO Runtime/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"OpenVINO IR INT8 model in OpenVINO Runtime/CPU: {time_ir / num_samples:.3f} "
    f"seconds per sentence, SPS: {num_samples / time_ir:.2f}"
)

PyTorch model on CPU: 0.071 seconds per sentence, SPS: 14.03
IR FP32 model in OpenVINO Runtime/CPU: 0.020 seconds per sentence, SPS: 50.00
OpenVINO IR INT8 model in OpenVINO Runtime/CPU: 0.009 seconds per sentence, SPS: 113.12

Finally, measure the inference performance of OpenVINO FP32 and INT8 models. For this purpose, use Benchmark Tool in OpenVINO.

Note: The benchmark_app tool is able to measure the performance of the OpenVINO Intermediate Representation (OpenVINO IR) models only. For more accurate performance, run benchmark_app in a terminal/command prompt after closing other applications. Run benchmark_app -m model.xml -d CPU to benchmark async inference on CPU for one minute. Change CPU to GPU to benchmark on GPU. Run benchmark_app --help to see an overview of all command-line options.

# Inference FP32 model (OpenVINO IR)
! benchmark_app -m $ir_model_xml -d CPU -api sync

[Step 1/11] Parsing and validating input arguments
[ INFO ] Parsing input parameters
[Step 2/11] Loading OpenVINO Runtime
[ INFO ] OpenVINO:
[ INFO ] Build ................................. 2022.3.0-9052-9752fafe8eb-releases/2022/3
[ INFO ]
[ INFO ] Device info:
[ INFO ] CPU
[ INFO ] Build ................................. 2022.3.0-9052-9752fafe8eb-releases/2022/3
[ INFO ]
[ INFO ]
[Step 3/11] Setting device configuration
[ WARNING ] Performance hint was not explicitly specified in command line. Device(CPU) performance hint will be set to LATENCY.
[Step 4/11] Reading model files
[ INFO ] Loading model files
[ INFO ] Read model took 203.52 ms
[ INFO ] Original model I/O parameters:
[ INFO ] Model inputs:
[ INFO ]     input_ids (node: input_ids) : i64 / [...] / [1,128]
[ INFO ]     attention_mask (node: attention_mask) : i64 / [...] / [1,128]
[ INFO ]     token_type_ids (node: token_type_ids) : i64 / [...] / [1,128]
[ INFO ] Model outputs:
[ INFO ]     output (node: output) : f32 / [...] / [1,2]
[Step 5/11] Resizing model to match image sizes and given batch
[ INFO ] Model batch size: 1
[Step 6/11] Configuring input of the model
[ INFO ] Model inputs:
[ INFO ]     input_ids (node: input_ids) : i64 / [...] / [1,128]
[ INFO ]     attention_mask (node: attention_mask) : i64 / [...] / [1,128]
[ INFO ]     token_type_ids (node: token_type_ids) : i64 / [...] / [1,128]
[ INFO ] Model outputs:
[ INFO ]     output (node: output) : f32 / [...] / [1,2]
[Step 7/11] Loading the model to the device
[ INFO ] Compile model took 218.71 ms
[Step 8/11] Querying optimal runtime parameters
[ INFO ] Model:
[ INFO ]   NETWORK_NAME: torch_jit
[ INFO ]   OPTIMAL_NUMBER_OF_INFER_REQUESTS: 1
[ INFO ]   NUM_STREAMS: 1
[ INFO ]   AFFINITY: Affinity.CORE
[ INFO ]   INFERENCE_NUM_THREADS: 12
[ INFO ]   PERF_COUNT: False
[ INFO ]   INFERENCE_PRECISION_HINT: <Type: 'float32'>
[ INFO ]   PERFORMANCE_HINT: PerformanceMode.LATENCY
[ INFO ]   PERFORMANCE_HINT_NUM_REQUESTS: 0
[Step 9/11] Creating infer requests and preparing input tensors
[ 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 'attention_mask'!. This input will be filled with random values!
[ WARNING ] No input files were given for input 'token_type_ids'!. This input will be filled with random values!
[ INFO ] Fill input 'input_ids' with random values
[ INFO ] Fill input 'attention_mask' with random values
[ INFO ] Fill input 'token_type_ids' with random values
[Step 10/11] Measuring performance (Start inference synchronously, limits: 60000 ms duration)
[ INFO ] Benchmarking in inference only mode (inputs filling are not included in measurement loop).
[ INFO ] First inference took 30.87 ms
[Step 11/11] Dumping statistics report
[ INFO ] Count:            3097 iterations
[ INFO ] Duration:         60007.44 ms
[ INFO ] Latency:
[ INFO ]    Median:        19.31 ms
[ INFO ]    Average:       19.29 ms
[ INFO ]    Min:           18.67 ms
[ INFO ]    Max:           22.77 ms
[ INFO ] Throughput:   51.77 FPS

# Inference INT8 model (OpenVINO IR)
! benchmark_app -m $compressed_model_xml -d CPU -api sync

[Step 1/11] Parsing and validating input arguments
[ INFO ] Parsing input parameters
[Step 2/11] Loading OpenVINO Runtime
[ INFO ] OpenVINO:
[ INFO ] Build ................................. 2022.3.0-9052-9752fafe8eb-releases/2022/3
[ INFO ]
[ INFO ] Device info:
[ INFO ] CPU
[ INFO ] Build ................................. 2022.3.0-9052-9752fafe8eb-releases/2022/3
[ INFO ]
[ INFO ]
[Step 3/11] Setting device configuration
[ WARNING ] Performance hint was not explicitly specified in command line. Device(CPU) performance hint will be set to LATENCY.
[Step 4/11] Reading model files
[ INFO ] Loading model files
[ INFO ] Read model took 101.68 ms
[ INFO ] Original model I/O parameters:
[ INFO ] Model inputs:
[ INFO ]     input_ids:0 , input_ids (node: input_ids) : i64 / [...] / [1,128]
[ INFO ]     attention_mask (node: attention_mask) : i64 / [...] / [1,128]
[ INFO ]     token_type_ids:0 , token_type_ids (node: token_type_ids) : i64 / [...] / [1,128]
[ INFO ] Model outputs:
[ INFO ]     output (node: output) : f32 / [...] / [1,2]
[Step 5/11] Resizing model to match image sizes and given batch
[ INFO ] Model batch size: 1
[Step 6/11] Configuring input of the model
[ INFO ] Model inputs:
[ INFO ]     input_ids:0 , input_ids (node: input_ids) : i64 / [...] / [1,128]
[ INFO ]     attention_mask (node: attention_mask) : i64 / [...] / [1,128]
[ INFO ]     token_type_ids:0 , token_type_ids (node: token_type_ids) : i64 / [...] / [1,128]
[ INFO ] Model outputs:
[ INFO ]     output (node: output) : f32 / [...] / [1,2]
[Step 7/11] Loading the model to the device
[ INFO ] Compile model took 387.60 ms
[Step 8/11] Querying optimal runtime parameters
[ INFO ] Model:
[ INFO ]   NETWORK_NAME: torch_jit
[ INFO ]   OPTIMAL_NUMBER_OF_INFER_REQUESTS: 1
[ INFO ]   NUM_STREAMS: 1
[ INFO ]   AFFINITY: Affinity.CORE
[ INFO ]   INFERENCE_NUM_THREADS: 12
[ INFO ]   PERF_COUNT: False
[ INFO ]   INFERENCE_PRECISION_HINT: <Type: 'float32'>
[ INFO ]   PERFORMANCE_HINT: PerformanceMode.LATENCY
[ INFO ]   PERFORMANCE_HINT_NUM_REQUESTS: 0
[Step 9/11] Creating infer requests and preparing input tensors
[ 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 'attention_mask'!. This input will be filled with random values!
[ WARNING ] No input files were given for input 'token_type_ids'!. This input will be filled with random values!
[ INFO ] Fill input 'input_ids' with random values
[ INFO ] Fill input 'attention_mask' with random values
[ INFO ] Fill input 'token_type_ids' with random values
[Step 10/11] Measuring performance (Start inference synchronously, limits: 60000 ms duration)
[ INFO ] Benchmarking in inference only mode (inputs filling are not included in measurement loop).
[ INFO ] First inference took 16.64 ms
[Step 11/11] Dumping statistics report
[ INFO ] Count:            6650 iterations
[ INFO ] Duration:         60001.71 ms
[ INFO ] Latency:
[ INFO ]    Median:        8.84 ms
[ INFO ]    Average:       8.93 ms
[ INFO ]    Min:           7.61 ms
[ INFO ]    Max:           10.80 ms
[ INFO ] Throughput:   113.16 FPS

Quantize NLP models with Post-Training Quantization ​in NNCF¶