Convert and Optimize YOLOv8 with OpenVINO™

This Jupyter notebook can be launched after a local installation only.

The YOLOv8 algorithm developed by Ultralytics is a cutting-edge, state-of-the-art (SOTA) model that is designed to be fast, accurate, and easy to use, making it an excellent choice for a wide range of object detection, image segmentation, and image classification tasks. More details about its realization can be found in the original model repository.

This tutorial demonstrates step-by-step instructions on how to run apply quantization with accuracy control to PyTorch YOLOv8. The advanced quantization flow allows to apply 8-bit quantization to the model with control of accuracy metric. This is achieved by keeping the most impactful operations within the model in the original precision. The flow is based on the Basic 8-bit quantization and has the following differences:

• Besides the calibration dataset, a validation dataset is required to compute the accuracy metric. Both datasets can refer to the same data in the simplest case.

• Validation function, used to compute accuracy metric is required. It can be a function that is already available in the source framework or a custom function.

• Since accuracy validation is run several times during the quantization process, quantization with accuracy control can take more time than the Basic 8-bit quantization flow.

• The resulted model can provide smaller performance improvement than the Basic 8-bit quantization flow because some of the operations are kept in the original precision.

NOTE: Currently, 8-bit quantization with accuracy control in NNCF is available only for models in OpenVINO representation.

The steps for the quantization with accuracy control are described below.

Table of contents:

• Prerequisites

• Get Pytorch model and OpenVINO IR model

  • Define validator and data loader

  • Prepare calibration and validation datasets

  • Prepare validation function

• Run quantization with accuracy control

• Compare Accuracy and Performance of the Original and Quantized Models

Prerequisites

Install necessary packages.

%pip install -q "openvino>=2023.1.0"
%pip install -q "nncf>=2.6.0"
%pip install -q "ultralytics==8.0.43" --extra-index-url https://download.pytorch.org/whl/cpu

Get Pytorch model and OpenVINO IR model

Generally, PyTorch models represent an instance of the torch.nn.Module class, initialized by a state dictionary with model weights. We will use the YOLOv8 nano model (also known as yolov8n) pre-trained on a COCO dataset, which is available in this repo. Similar steps are also applicable to other YOLOv8 models. Typical steps to obtain a pre-trained model:

1. Create an instance of a model class.

2. Load a checkpoint state dict, which contains the pre-trained model weights.

In this case, the creators of the model provide an API that enables converting the YOLOv8 model to ONNX and then to OpenVINO IR. Therefore, we do not need to do these steps manually.

import os
from pathlib import Path

from ultralytics import YOLO
from ultralytics.yolo.cfg import get_cfg
from ultralytics.yolo.data.utils import check_det_dataset
from ultralytics.yolo.engine.validator import BaseValidator as Validator
from ultralytics.yolo.utils import DATASETS_DIR
from ultralytics.yolo.utils import DEFAULT_CFG
from ultralytics.yolo.utils import ops
from ultralytics.yolo.utils.metrics import ConfusionMatrix

ROOT = os.path.abspath('')

MODEL_NAME = "yolov8n-seg"

model = YOLO(f"{ROOT}/{MODEL_NAME}.pt")
args = get_cfg(cfg=DEFAULT_CFG)
args.data = "coco128-seg.yaml"

Load model.

import openvino as ov


model_path = Path(f"{ROOT}/{MODEL_NAME}_openvino_model/{MODEL_NAME}.xml")
if not model_path.exists():
    model.export(format="openvino", dynamic=True, half=False)

ov_model = ov.Core().read_model(model_path)

Define validator and data loader

The original model repository uses a Validator wrapper, which represents the accuracy validation pipeline. It creates dataloader and evaluation metrics and updates metrics on each data batch produced by the dataloader. Besides that, it is responsible for data preprocessing and results postprocessing. For class initialization, the configuration should be provided. We will use the default setup, but it can be replaced with some parameters overriding to test on custom data. The model has connected the ValidatorClass method, which creates a validator class instance.

validator = model.ValidatorClass(args)
validator.data = check_det_dataset(args.data)
data_loader = validator.get_dataloader(f"{DATASETS_DIR}/coco128-seg", 1)

validator.is_coco = True
validator.class_map = ops.coco80_to_coco91_class()
validator.names = model.model.names
validator.metrics.names = validator.names
validator.nc = model.model.model[-1].nc
validator.nm = 32
validator.process = ops.process_mask
validator.plot_masks = []

val: Scanning /home/ea/work/openvino_notebooks/notebooks/230-yolov8-optimization/datasets/coco128-seg/labels/train2017.cache... 126 images, 2 backgrounds, 0 corrupt: 100%|██████████| 128/128 [00:00<?, ?it/s

Prepare calibration and validation datasets

We can use one dataset as calibration and validation datasets. Name it quantization_dataset.

from typing import Dict

import nncf


def transform_fn(data_item: Dict):
    input_tensor = validator.preprocess(data_item)["img"].numpy()
    return input_tensor


quantization_dataset = nncf.Dataset(data_loader, transform_fn)

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

Prepare validation function

from functools import partial

import torch
from nncf.quantization.advanced_parameters import AdvancedAccuracyRestorerParameters


def validation_ac(
    compiled_model: ov.CompiledModel,
    validation_loader: torch.utils.data.DataLoader,
    validator: Validator,
    num_samples: int = None,
    log=True
) -> float:
    validator.seen = 0
    validator.jdict = []
    validator.stats = []
    validator.batch_i = 1
    validator.confusion_matrix = ConfusionMatrix(nc=validator.nc)
    num_outputs = len(compiled_model.outputs)

    counter = 0
    for batch_i, batch in enumerate(validation_loader):
        if num_samples is not None and batch_i == num_samples:
            break
        batch = validator.preprocess(batch)
        results = compiled_model(batch["img"])
        if num_outputs == 1:
            preds = torch.from_numpy(results[compiled_model.output(0)])
        else:
            preds = [
                torch.from_numpy(results[compiled_model.output(0)]),
                torch.from_numpy(results[compiled_model.output(1)]),
            ]
        preds = validator.postprocess(preds)
        validator.update_metrics(preds, batch)
        counter += 1
    stats = validator.get_stats()
    if num_outputs == 1:
        stats_metrics = stats["metrics/mAP50-95(B)"]
    else:
        stats_metrics = stats["metrics/mAP50-95(M)"]
    if log:
        print(f"Validate: dataset length = {counter}, metric value = {stats_metrics:.3f}")

    return stats_metrics


validation_fn = partial(validation_ac, validator=validator, log=False)

Run quantization with accuracy control

You should provide the calibration dataset and the validation dataset. It can be the same dataset. - parameter max_drop defines the accuracy drop threshold. The quantization process stops when the degradation of accuracy metric on the validation dataset is less than the max_drop. The default value is 0.01. NNCF will stop the quantization and report an error if the max_drop value can’t be reached. - drop_type defines how the accuracy drop will be calculated: ABSOLUTE (used by default) or RELATIVE. - ranking_subset_size - size of a subset that is used to rank layers by their contribution to the accuracy drop. Default value is 300, and the more samples it has the better ranking, potentially. Here we use the value 25 to speed up the execution.

NOTE: Execution can take tens of minutes and requires up to 15 GB of free memory

quantized_model = nncf.quantize_with_accuracy_control(
    ov_model,
    quantization_dataset,
    quantization_dataset,
    validation_fn=validation_fn,
    max_drop=0.01,
    preset=nncf.QuantizationPreset.MIXED,
    subset_size=128,
    advanced_accuracy_restorer_parameters=AdvancedAccuracyRestorerParameters(
        ranking_subset_size=25
    ),
)

2023-10-10 09:55:44.477778: I tensorflow/core/util/port.cc:110] oneDNN custom operations are on. You may see slightly different numerical results due to floating-point round-off errors from different computation orders. To turn them off, set the environment variable TF_ENABLE_ONEDNN_OPTS=0.
2023-10-10 09:55:44.516624: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
To enable the following instructions: AVX2 AVX512F AVX512_VNNI FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
2023-10-10 09:55:45.324364: W tensorflow/compiler/tf2tensorrt/utils/py_utils.cc:38] TF-TRT Warning: Could not find TensorRT
Statistics collection: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 128/128 [00:16<00:00,  7.79it/s]
Applying Fast Bias correction: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 75/75 [00:03<00:00, 18.84it/s]

INFO:nncf:Validation of initial model was started

INFO:nncf:Elapsed Time: 00:00:00
INFO:nncf:Elapsed Time: 00:00:05
INFO:nncf:Metric of initial model: 0.366118260036709
INFO:nncf:Collecting values for each data item using the initial model
INFO:nncf:Elapsed Time: 00:00:06
INFO:nncf:Validation of quantized model was started
INFO:nncf:Elapsed Time: 00:00:01
INFO:nncf:Elapsed Time: 00:00:04
INFO:nncf:Metric of quantized model: 0.3418411101103462
INFO:nncf:Collecting values for each data item using the quantized model
INFO:nncf:Elapsed Time: 00:00:05
INFO:nncf:Accuracy drop: 0.024277149926362818 (DropType.ABSOLUTE)
INFO:nncf:Accuracy drop: 0.024277149926362818 (DropType.ABSOLUTE)
INFO:nncf:Total number of quantized operations in the model: 91
INFO:nncf:Number of parallel processes to rank quantized operations: 6
INFO:nncf:ORIGINAL metric is used to rank quantizers
INFO:nncf:Calculating ranking score for groups of quantizers
INFO:nncf:Elapsed Time: 00:02:16
INFO:nncf:Changing the scope of quantizer nodes was started
INFO:nncf:Reverted 1 operations to the floating-point precision:
    /model.22/Mul_5
INFO:nncf:Accuracy drop with the new quantization scope is 0.013359187935064742 (DropType.ABSOLUTE)
INFO:nncf:Reverted 1 operations to the floating-point precision:
    /model.1/conv/Conv/WithoutBiases
INFO:nncf:Accuracy drop with the new quantization scope is 0.01287864227202773 (DropType.ABSOLUTE)
INFO:nncf:Reverted 1 operations to the floating-point precision:
    /model.2/cv1/conv/Conv/WithoutBiases
INFO:nncf:Algorithm completed: achieved required accuracy drop 0.007027355074555763 (DropType.ABSOLUTE)
INFO:nncf:3 out of 91 were reverted back to the floating-point precision:
    /model.22/Mul_5
    /model.1/conv/Conv/WithoutBiases
    /model.2/cv1/conv/Conv/WithoutBiases

Compare Accuracy and Performance of the Original and Quantized Models

Now we can compare metrics of the Original non-quantized OpenVINO IR model and Quantized OpenVINO IR model to make sure that the max_drop is not exceeded.

core = ov.Core()
quantized_compiled_model = core.compile_model(model=quantized_model, device_name='CPU')
compiled_ov_model = core.compile_model(model=ov_model, device_name='CPU')

pt_result = validation_ac(compiled_ov_model, data_loader, validator)
quantized_result = validation_ac(quantized_compiled_model, data_loader, validator)


print(f'[Original OpenVino]: {pt_result:.4f}')
print(f'[Quantized OpenVino]: {quantized_result:.4f}')

Validate: dataset length = 128, metric value = 0.368
Validate: dataset length = 128, metric value = 0.360
[Original OpenVino]: 0.3677
[Quantized OpenVino]: 0.3602

And compare performance.

from pathlib import Path
# Set model directory
MODEL_DIR = Path("model")
MODEL_DIR.mkdir(exist_ok=True)

ir_model_path = MODEL_DIR / 'ir_model.xml'
quantized_model_path = MODEL_DIR / 'quantized_model.xml'

# Save models to use them in the commandline banchmark app
ov.save_model(ov_model, ir_model_path, compress_to_fp16=False)
ov.save_model(quantized_model, quantized_model_path, compress_to_fp16=False)

# Inference Original model (OpenVINO IR)
! benchmark_app -m $ir_model_path -shape "[1,3,640,640]" -d CPU -api async

[Step 1/11] Parsing and validating input arguments
[ INFO ] Parsing input parameters
[Step 2/11] Loading OpenVINO Runtime
[ INFO ] OpenVINO:
[ INFO ] Build ................................. 2023.2.0-12713-47c2a91b6b6
[ INFO ]
[ INFO ] Device info:
[ INFO ] CPU
[ INFO ] Build ................................. 2023.2.0-12713-47c2a91b6b6
[ 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 PerformanceMode.THROUGHPUT.
[Step 4/11] Reading model files
[ INFO ] Loading model files
[ INFO ] Read model took 27.11 ms
[ INFO ] Original model I/O parameters:
[ INFO ] Model inputs:
[ INFO ]     images (node: images) : f32 / [...] / [?,3,?,?]
[ INFO ] Model outputs:
[ INFO ]     output0 (node: output0) : f32 / [...] / [?,116,?]
[ INFO ]     output1 (node: output1) : f32 / [...] / [?,32,8..,8..]
[Step 5/11] Resizing model to match image sizes and given batch
[ INFO ] Model batch size: 1
[ INFO ] Reshaping model: 'images': [1,3,640,640]
[ INFO ] Reshape model took 13.41 ms
[Step 6/11] Configuring input of the model
[ INFO ] Model inputs:
[ INFO ]     images (node: images) : u8 / [N,C,H,W] / [1,3,640,640]
[ INFO ] Model outputs:
[ INFO ]     output0 (node: output0) : f32 / [...] / [1,116,8400]
[ INFO ]     output1 (node: output1) : f32 / [...] / [1,32,160,160]
[Step 7/11] Loading the model to the device
[ INFO ] Compile model took 274.70 ms
[Step 8/11] Querying optimal runtime parameters
[ INFO ] Model:
[ INFO ]   NETWORK_NAME: torch_jit
[ INFO ]   OPTIMAL_NUMBER_OF_INFER_REQUESTS: 12
[ INFO ]   NUM_STREAMS: 12
[ INFO ]   AFFINITY: Affinity.CORE
[ INFO ]   INFERENCE_NUM_THREADS: 36
[ INFO ]   PERF_COUNT: False
[ INFO ]   INFERENCE_PRECISION_HINT: <Type: 'float32'>
[ INFO ]   PERFORMANCE_HINT: PerformanceMode.THROUGHPUT
[ INFO ]   EXECUTION_MODE_HINT: ExecutionMode.PERFORMANCE
[ INFO ]   PERFORMANCE_HINT_NUM_REQUESTS: 0
[ INFO ]   ENABLE_CPU_PINNING: True
[ INFO ]   SCHEDULING_CORE_TYPE: SchedulingCoreType.ANY_CORE
[ INFO ]   ENABLE_HYPER_THREADING: True
[ INFO ]   EXECUTION_DEVICES: ['CPU']
[ INFO ]   CPU_DENORMALS_OPTIMIZATION: False
[ INFO ]   CPU_SPARSE_WEIGHTS_DECOMPRESSION_RATE: 1.0
[Step 9/11] Creating infer requests and preparing input tensors
[ WARNING ] No input files were given for input 'images'!. This input will be filled with random values!
[ INFO ] Fill input 'images' with random values
[Step 10/11] Measuring performance (Start inference asynchronously, 12 inference requests, limits: 60000 ms duration)
[ INFO ] Benchmarking in inference only mode (inputs filling are not included in measurement loop).
[ INFO ] First inference took 42.88 ms
[Step 11/11] Dumping statistics report
[ INFO ] Execution Devices:['CPU']
[ INFO ] Count:            7716 iterations
[ INFO ] Duration:         60104.07 ms
[ INFO ] Latency:
[ INFO ]    Median:        88.61 ms
[ INFO ]    Average:       93.27 ms
[ INFO ]    Min:           52.72 ms
[ INFO ]    Max:           181.74 ms
[ INFO ] Throughput:   128.38 FPS

# Inference Quantized model (OpenVINO IR)
! benchmark_app -m $quantized_model_path -shape "[1,3,640,640]" -d CPU -api async

[Step 1/11] Parsing and validating input arguments
[ INFO ] Parsing input parameters
[Step 2/11] Loading OpenVINO Runtime
[ INFO ] OpenVINO:
[ INFO ] Build ................................. 2023.2.0-12713-47c2a91b6b6
[ INFO ]
[ INFO ] Device info:
[ INFO ] CPU
[ INFO ] Build ................................. 2023.2.0-12713-47c2a91b6b6
[ 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 PerformanceMode.THROUGHPUT.
[Step 4/11] Reading model files
[ INFO ] Loading model files
[ INFO ] Read model took 32.74 ms
[ INFO ] Original model I/O parameters:
[ INFO ] Model inputs:
[ INFO ]     images (node: images) : f32 / [...] / [?,3,?,?]
[ INFO ] Model outputs:
[ INFO ]     output0 (node: output0) : f32 / [...] / [?,116,?]
[ INFO ]     output1 (node: output1) : f32 / [...] / [?,32,8..,8..]
[Step 5/11] Resizing model to match image sizes and given batch
[ INFO ] Model batch size: 1
[ INFO ] Reshaping model: 'images': [1,3,640,640]
[ INFO ] Reshape model took 18.09 ms
[Step 6/11] Configuring input of the model
[ INFO ] Model inputs:
[ INFO ]     images (node: images) : u8 / [N,C,H,W] / [1,3,640,640]
[ INFO ] Model outputs:
[ INFO ]     output0 (node: output0) : f32 / [...] / [1,116,8400]
[ INFO ]     output1 (node: output1) : f32 / [...] / [1,32,160,160]
[Step 7/11] Loading the model to the device
[ INFO ] Compile model took 574.58 ms
[Step 8/11] Querying optimal runtime parameters
[ INFO ] Model:
[ INFO ]   NETWORK_NAME: torch_jit
[ INFO ]   OPTIMAL_NUMBER_OF_INFER_REQUESTS: 12
[ INFO ]   NUM_STREAMS: 12
[ INFO ]   AFFINITY: Affinity.CORE
[ INFO ]   INFERENCE_NUM_THREADS: 36
[ INFO ]   PERF_COUNT: False
[ INFO ]   INFERENCE_PRECISION_HINT: <Type: 'float32'>
[ INFO ]   PERFORMANCE_HINT: PerformanceMode.THROUGHPUT
[ INFO ]   EXECUTION_MODE_HINT: ExecutionMode.PERFORMANCE
[ INFO ]   PERFORMANCE_HINT_NUM_REQUESTS: 0
[ INFO ]   ENABLE_CPU_PINNING: True
[ INFO ]   SCHEDULING_CORE_TYPE: SchedulingCoreType.ANY_CORE
[ INFO ]   ENABLE_HYPER_THREADING: True
[ INFO ]   EXECUTION_DEVICES: ['CPU']
[ INFO ]   CPU_DENORMALS_OPTIMIZATION: False
[ INFO ]   CPU_SPARSE_WEIGHTS_DECOMPRESSION_RATE: 1.0
[Step 9/11] Creating infer requests and preparing input tensors
[ WARNING ] No input files were given for input 'images'!. This input will be filled with random values!
[ INFO ] Fill input 'images' with random values
[Step 10/11] Measuring performance (Start inference asynchronously, 12 inference requests, limits: 60000 ms duration)
[ INFO ] Benchmarking in inference only mode (inputs filling are not included in measurement loop).
[ INFO ] First inference took 31.29 ms
[Step 11/11] Dumping statistics report
[ INFO ] Execution Devices:['CPU']
[ INFO ] Count:            15900 iterations
[ INFO ] Duration:         60077.25 ms
[ INFO ] Latency:
[ INFO ]    Median:        42.02 ms
[ INFO ]    Average:       45.18 ms
[ INFO ]    Min:           25.42 ms
[ INFO ]    Max:           117.81 ms
[ INFO ] Throughput:   264.66 FPS