Skip to content

Latest commit

 

History

History
127 lines (100 loc) · 4.37 KB

README.md

File metadata and controls

127 lines (100 loc) · 4.37 KB

Tiny3D

Tiny3D website

Documentation, Tutorials and examples

Introduction

Tiny3D is a next generation of 3D object detection service production system.

Features

Tiny3D solution embodies four transformative features:

  1. A Performance Optimization Engine for 3d object detection online/offline inference services product performance optimization. Through this engine users can easily get a high accuracy and high speed 3d object detection service/competetion result in a Data-Centeric AI way. Our Performance Optimization Engine can easily be a Plug-in to any machine learning system.
  2. One line of code to complete dataset editing, model training, model testing, model compression, model deployment.
  3. One line of code to Fine-grained data editing on different size datasets or single data.
  4. A user-friendly web interface for a developer team to product a 3d object detection service pictorially, in a low-code fashion. [currently not supported]

Example1: Fast access to a high-precision 3d object detection service

step-1: Edit the data using different data operation method to get high quality dataset

from tiny3d.data import dataset_edit

dataset_edit(dataset_input_path, dataset_output_path, denoise_method=None, 
             simulation_method='Snow', filter_method=None, 
             augmentation_method=None, qualification_method=None)

step-2: Train a model on edited dataset

from tiny3d.deephub import Pointpillars
from tiny3d.engine import build_dataset, engine, fit

model = Pointpillars()
model = engine(model)

dataset_train = build_dataset(train_dataset_path)
dataset_val = build_dataset(val_dataset_path)

fit(dataset_train=dataset_train, dataset_val=dataset_val, torch_model=model)

step-3: Compress a trained model

from tiny3d.model.model_compressor import prune, quant 

prune(model)
quant(model)

step-4: Deploy a model

from tiny3d.model.model_deployor import deploy 

backend_file = deploy(model, backend='tensorrt', output_file=output_model_path)

step-5: Provide a model serving

from tiny3d.model.model_server import payload

PyTorch_REST_API_URL = 'http://127.0.0.1:1234/'
url = PyTorch_REST_API_URL + 'transfer'

# Submit the request
requests.post(url, files=payload).json()

Example2: Through data editing to further improve the accuracy of 3d object detection service.

step-1: Real world bad case visulization or potential bad case visulization

# obtain the potential bad case, you can also specify a real world bad case data path.
from tiny3d.engine import inference
from tiny3d.data.data_qualificator import lidar_qualificate

bad_case_data_path = './'
prediction = inference(model, dataset)
bad_case = lidar_qualificate(dataset, prediction, Topk=100, save_path=bad_case_data_path)

# Visualization
from tiny3d.data.data_visulizator import visualize

visualize(bad_case_data_path)

step-2: Edit the dataset to get more bad case data to imporove the 3d object detection service accuracy on bad case data in real world case.

from tiny3d.data import dataset_edit

dataset_edit(bad_case_data_path, dataset_output_path, denoise_method=None, 
             simulation_method='Snow, Rain', filter_method=None, 
             augmentation_method=None, qualification_method=None)

step-3: Re-train the model on a edited data and deploy it.

Operations Tiny3D currently supported

Lidar data operations currently supported

  • lidar data loading
  • lidar data sampling
  • lidar data preprocessing/cleaning
  • lidar data denoising
  • lidar data outlier detection
  • lidar data augmentation
  • lidar data simulation

Lidar based 3d object detection model operations currently supported

  • model compression
  • model deploy and serve
  • model ensemble

Data-Model co-operations currently supported

  • training
  • testing
  • bad case visulization

TODO

1. Reorganize the code and docs

2. Add more data ops

  • lidar data selection
  • lidar data robustion
  • lidar data privacy
  • lidar data domain adptation
  • lidar data auto-labeling
  • data drift emergency

3. Add visual interaction interface.

Acknowlegement