Object Detection model#

This tutorial reveals end-to-end solution from installation to model export and optimization for object detection task on a specific example.

To learn more about Object Detection task, refer to Object Detection.

On this page, we show how to train, validate, export and optimize ATSS model on WGISD public dataset.

To have a specific example in this tutorial, all commands will be run on the ATSS model. It’s a medium model, that achieves relatively high accuracy while keeping the inference fast.

The process has been tested on the following configuration.

Ubuntu 20.04
NVIDIA GeForce RTX 3090
Intel(R) Core(TM) i9-11900
CUDA Toolkit 11.8

Setup virtual environment#

1. You can follow the installation process from a quick start guide to create a universal virtual environment for OpenVINO™ Training Extensions.

2. Activate your virtual environment:

.otx/bin/activate
# or by this line, if you created an environment, using tox
. venv/otx/bin/activate

Dataset preparation#

Note

Currently, we support the following object detection dataset formats:

1. Clone a repository with WGISD dataset.

mkdir data ; cd data
git clone https://github.com/thsant/wgisd.git
cd wgisd
git checkout 6910edc5ae3aae8c20062941b1641821f0c30127

This dataset contains images of grapevines with the annotation for different varieties of grapes.

CDY - Chardonnay
CFR - Cabernet Franc
CSV - Cabernet Sauvignon
SVB - Sauvignon Blanc
SYH - Syrah

It’s a great example to start with. The model achieves high accuracy right from the beginning of the training due to relatively large and focused objects. Also, these objects are distinguished by a person, so we can check inference results just by looking at images.

this image uploaded from this `source <https://github.com/thsant/wgisd/blob/master/data/CDY_2015.jpg>`_

2. To run the training using auto-configuration feature, we need to reformat the dataset according to this structure:

wgisd
├── annotations/
    ├── instances_train.json
    ├── instances_val.json
    └── instances_test.json
├──images/
    ├── train
    ├── val
    └── test

We can do that by running these commands:

# format images folder
mv data images

# format annotations folder
mv coco_annotations annotations

# rename annotations to meet *_train.json pattern
mv annotations/train_bbox_instances.json annotations/instances_train.json
mv annotations/test_bbox_instances.json annotations/instances_val.json
cp annotations/instances_val.json annotations/instances_test.json

cd ../..

Training#

1. First of all, you need to choose which object detection model you want to train. The list of supported recipes for object detection is available with the command line below.

Note

The characteristics and detailed comparison of the models could be found in Explanation section.

CLI

(otx) ...$ otx find --task DETECTION --pattern atss
┏━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┓
┃ Task      ┃ Model Name            ┃ Recipe Path                                                    ┃
┡━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┩
│ DETECTION │ atss_mobilenetv2_tile │ src/otx/recipe/detection/atss_mobilenetv2_tile.yaml            │
│ DETECTION │ atss_resnext101       │ src/otx/recipe/detection/atss_resnext101.yaml                  │
│ DETECTION │ atss_mobilenetv2      │ src/otx/recipe/detection/atss_mobilenetv2.yaml                 │
└───────────┴───────────────────────┴────────────────────────────────────────────────────────────────┘

API

from otx.engine.utils.api import list_models

model_lists = list_models(task="DETECTION", pattern="atss")
print(model_lists)
'''
[
    'atss_mobilenetv2',
    'atss_mobilenetv2_tile',
    'atss_resnext101',
]
'''

2. On this step we will configure configuration with:

all necessary configs for atss_mobilenetv2
train/validation sets, based on provided annotation.

It may be counterintuitive, but for --data_root we need to pass the path to the dataset folder root (in our case it’s data/wgisd) instead of the folder with validation images. This is because the function automatically detects annotations and images according to the expected folder structure we achieved above.

Let’s check the object detection configuration running the following command:

# or its config path
(otx) ...$ otx train --config  src/otx/recipe/detection/atss_mobilenetv2.yaml \
                     --data_root data/wgisd \
                     --work_dir otx-workspace \
                     --print_config

...
data_root: data/wgisd
work_dir: otx-workspace
callback_monitor: val/map_50
disable_infer_num_classes: false
engine:
  task: DETECTION
  device: auto
data:
...

Note

If you want to get configuration as yaml file, please use --print_config parameter and > configs.yaml.

(otx) ...$ otx train --config  src/otx/recipe/detection/atss_mobilenetv2.yaml --data_root data/wgisd --print_config > configs.yaml
# Update configs.yaml & Train configs.yaml
(otx) ...$ otx train --config configs.yaml

3. otx train trains a model (a particular model recipe) on a dataset and results:

Here are the main outputs can expect with CLI: - {work_dir}/{timestamp}/checkpoints/epoch_*.ckpt - a model checkpoint file. - {work_dir}/{timestamp}/configs.yaml - The configuration file used in the training can be reused to reproduce the training. - {work_dir}/.latest - The results of each of the most recently executed subcommands are soft-linked. This allows you to skip checkpoints and config file entry as a workspace.

CLI (auto-config)

(otx) ...$ otx train --data_root data/wgisd

CLI (with config)

(otx) ...$ otx train --config src/otx/recipe/detection/atss_mobilenetv2.yaml --data_root data/wgisd

API (from_config)

from otx.engine import Engine

data_root = "data/wgisd"
recipe = "src/otx/recipe/detection/atss_mobilenetv2.yaml"

engine = Engine.from_config(
          config_path=recipe,
          data_root=data_root,
          work_dir="otx-workspace",
        )

engine.train(...)

API

from otx.engine import Engine

data_root = "data/wgisd"

engine = Engine(
          model="atss_mobilenetv2",
          data_root=data_root,
          work_dir="otx-workspace",
        )

engine.train(...)

4. (Optional) Additionally, we can tune training parameters such as batch size, learning rate, patience epochs or warm-up iterations. Learn more about specific parameters using otx train --help -v or otx train --help -vv.

For example, to decrease the batch size to 4, fix the number of epochs to 100, extend the command line above with the following line.

CLI

(otx) ...$ otx train ... --data.train_subset.batch_size 4 \
                         --max_epochs 100

API

from otx.core.config.data import SubsetConfig
from otx.core.data.module import OTXDataModule
from otx.engine import Engine

datamodule = OTXDataModule(..., train_subset=SubsetConfig(..., batch_size=4))

engine = Engine(..., datamodule=datamodule)

engine.train(max_epochs=100)

5. The training result checkpoints/*.ckpt file is located in {work_dir} folder, while training logs can be found in the {work_dir}/{timestamp} dir.

Note

We also can visualize the training using Tensorboard as these logs are located in {work_dir}/{timestamp}/tensorboard.

otx-workspace
├── 20240403_134256/
    ├── csv/
    ├── checkpoints/
    |   └── epoch_*.pth
    ├── tensorboard/
    └── configs.yaml
└── .latest
    └── train/
...

The training time highly relies on the hardware characteristics, for example on 1 NVIDIA GeForce RTX 3090 the training took about 3 minutes.

After that, we have the PyTorch object detection model trained with OpenVINO™ Training Extensions, which we can use for evaluation, export, optimization and deployment.

Evaluation#

1. otx test runs evaluation of a trained model on a particular dataset.

Test function receives test annotation information and model snapshot, trained in previous step.

The default metric is mAP_50 measure.

2. That’s how we can evaluate the snapshot in otx-workspace folder on WGISD dataset and save results to otx-workspace:

CLI (with work_dir)

(otx) ...$ otx test --work_dir otx-workspace
┏━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━━┓
┃        Test metric        ┃       DataLoader 0        ┃
┡━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━━┩
│      test/data_time       │   0.025369757786393166    │
│       test/map_50         │    0.8693901896476746     │
│      test/iter_time       │    0.08180806040763855    │
└───────────────────────────┴───────────────────────────┘

CLI (with config)

(otx) ...$ otx test --config  src/otx/recipe/detection/atss_mobilenetv2.yaml \
                    --data_root data/wgisd \
                    --checkpoint otx-workspace/20240312_051135/checkpoints/epoch_033.ckpt
┏━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━━┓
┃        Test metric        ┃       DataLoader 0        ┃
┡━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━━┩
│      test/data_time       │   0.025369757786393166    │
│       test/map_50         │    0.8693901896476746     │
│      test/iter_time       │    0.08180806040763855    │
└───────────────────────────┴───────────────────────────┘

API

engine.test()

3. The output of {work_dir}/{timestamp}/csv/version_0/metrics.csv consists of a dict with target metric name and its value.

Optimization#

1. We can further optimize the model with otx optimize. It uses PTQ depending on the model and transforms it to INT8 format.

PTQ optimization is used for models exported in the OpenVINO™ IR format. It decreases the floating-point precision to integer precision of the exported model by performing the post-training optimization.

To learn more about optimization, refer to NNCF repository.

2. Command example for optimizing OpenVINO™ model (.xml) with OpenVINO™ PTQ.

CLI

(otx) ...$ otx optimize  --work_dir otx-workspace \
                         --checkpoint otx-workspace/20240312_052847/exported_model.xml

...
Statistics collection ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 100% 30/30 • 0:00:14 • 0:00:00
Applying Fast Bias correction ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 100% 58/58 • 0:00:02 • 0:00:00
Elapsed time: 0:00:24.958733

API

ckpt_path = "otx-workspace/20240312_052847/exported_model.xml"
engine.optimize(checkpoint=ckpt_path)

The optimization time highly relies on the hardware characteristics, for example on Intel(R) Core(TM) i9-11900 it took about 25 seconds. Please note, that PTQ will take some time without logging to optimize the model.

Note

You can also pass export_demo_package=True parameter to obtain exportable_code.zip archive with packed optimized model and demo package. Please refer to export tutorial.

3. Finally, we can also evaluate the optimized model by passing it to the otx test function.

CLI

(otx) ...$ otx test --work_dir otx-workspace \
                    --checkpoint otx-workspace/20240312_055042/optimized_model.xml \
                    --engine.device cpu

...
┏━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━━┓
┃        Test metric        ┃       DataLoader 0        ┃
┡━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━━┩
│       test/map_50         │    0.8693901896476746     │
└───────────────────────────┴───────────────────────────┘
Elapsed time: 0:00:10.260521

API

ckpt_path = "otx-workspace/20240312_055042/optimized_model.xml"
engine.test(checkpoint=ckpt_path)

Now we have fully trained, optimized and exported an efficient model representation ready-to-use object detection model.

Object Detection model#

Setup virtual environment#

Dataset preparation#

Training#

Evaluation#

Export#

Optimization#