Introduce

This is our team's source code of the TaobaoLive Commodity Identify Competition of Tianchi Big Data Platform. Our team PAI got the sixth place in this competition.

Document

The Technical Document (CN-zh)

Object Detection

Configuration File

from config import get_args_efficientdet
opt = get_args_efficientdet()
print(opt)

Data Preparation

from process_data import processTrain, processValidation
label = {}
label['label2index'] = {}
label['index2label'] = {}
processTrain(label)
processValidation(label)

data from [images, videos] is in path data/[train, validation]_[images, videos] and data/[train, validation]_[images, videos]_annotation.json

[train, validation]_[images, videos]_annotation.json

{
    "annotations": [
        {
            "img_name": "{item id}_{frame}.jpg",
            "annotations": [
                {
                    "label": 22,
                    "viewpoint": 0,
                    "display": 1,
                    "instance_id": 0,
                    "box": [221, 507, 614, 1039]
                },
                ...
            ]
        },
        ...
    ]
}

Dataset

from dataset import EfficientdetDataset
dataset = EfficientdetDataset(root_dir, mode, imgORvdo, transform, maxLen, PAD)

EfficientDet

from config import get_args_efficientdet
from efficientdet.efficientdet import EfficientDet
opt = get_args_efficientdet()
model = EfficientDet(opt)

Train EfficientDet

python train_efficientdet.py

Validation EfficientDet

python validation_efficientdet.py

Match

Configuration File

from config import get_args_arcface
opt = get_args_arcface()
print(opt)

Data Preparation

from process_data import processTrain, processValidation
saveNumpyInstance('data', 'train', (256, 256))
saveNumpyInstance('data', 'validation', (256, 256))
createInstance2Label('data')
createInstanceID()

the shape of matrix in data/[train, validation]_instance is $256\times256\times3$

the input of arcface is $3\times224\times224$

Dataset

Arcface

from dataset import ArcfaceDataset
dataset = ArcfaceDataset(root_dir, mode, size, flip_x, maxLen, PAD, imgORvdo)

Triplet Loss

from dataset import TripletDataset
dataset = TripletDataset(root_dir, mode, size, flip_x)

Hard Triplet Loss

from dataset import HardTripletDataset
dataset = HardTripletDataset(root_dir, mode, size, flip_x, n_samples)

Arcface

Backbone

from config import get_args_arcface
from arcface.resnest import ResNeSt
opt = get_args_arcface()
backbone = ResNeSt(opt)

Head

from config import get_args_arcface
from arcface.head import Arcface
opt = get_args_arcface()
head = Arcface(opt)

Train Match Model

Arcface

python train_arcface.py

Triplet Loss

python train_triplet.py

Hard Triplet Loss

python train_hardtriplet.py

Validate Match Model

python validation_arcface.py

Validate with both object detection and match model

python validation_all.py
python cal_score.py

Docker

cd docker

Build

docker build -t image_name .

Push

docker push image_name

Test

nvidia-docker run -v /data/validation_dataset_part1:/tcdata/test_dataset_3w image_name sh run.sh

knectt / the-taobaolive-commodity-identify-competition Goto Github PK

the-taobaolive-commodity-identify-competition's Introduction

Introduce

Document

Object Detection

Configuration File

Data Preparation

Dataset

EfficientDet

Train EfficientDet

Validation EfficientDet

Match

Configuration File

Data Preparation

Dataset

Arcface

Triplet Loss

Hard Triplet Loss

Arcface

Backbone

Head

Train Match Model

Arcface

Triplet Loss

Hard Triplet Loss

Validate Match Model

Validate with both object detection and match model

Docker

Build

Push

Test

the-taobaolive-commodity-identify-competition's People

Contributors

Recommend Projects

Recommend Topics

Recommend Org