Food Recognition/ Ingredient Recognition

• 사진 안에 있는 음식이 어떤 음식인지 인식할 수 있는 인공지능 모델 개발
• 사진 안에 있는 식재료가 어떤 식재료인지 인식할 수 있는 인공지능 모델 개발

Getting Started

You can start on any computer that can learn deep learning. If you want to learn fast, use GPU-workstation.

Training Prerequisites

python version == 3.6.9

torch~=1.9.1
torchvision~=0.10.1
Pillow~=7.0.0
natsort~=7.0.1
sklearn~=0.0
scikit-learn~=0.24.2
tqdm~=4.42.0
numpy~=1.18.1
tensorflow~=2.2.0
tensorflow-gpu~=2.2.0
tensorboard~=2.7.0
matplotlib~=3.1.2

pip install -r requirements.txt

• If you only try inference,

torch~=1.9.1
torchvision~=0.10.1
Pillow~=7.0.0
natsort~=7.0.1
numpy~=1.18.1

How can I request Checkpoint?

If you respond to GoogleForms, we will share the download link within a few days. Currently, the shared checkpoint is ResNet152.

• Resquest JIT Traced Checkpoint
• Request Pytorch Checkpoint

Running the Test

• Use TorchScript

test_model = torch.jit.load('./jit_traced_torch_model_name.pt', map_location='cpu')
sample_data = torch.randn(1, 3, 512, 512) # (1, channel, width, height)
out_data = test_model(sample_data)

• pytorch

python inference.py --config configure_name.json --image image_name.jpg --label labels.txt

Quick Start Training Guide

1. You need to create a configuration first.
2. Then execute the following command:

python train.py --configuration configuration_name.json

Dataset

• KOREA AI HUB DATASET - 한국 음식 이미지 데이터셋
• Fruit and Vegetable Image Dataset - 과일과 채소 데이터셋
• Vegetable Image Dataset - 채소 데이터셋
• Private Dataset (INGD_V1, INGD_V2)

Support TorchScript

• Introduction to TorchScript - TorchScript 소개 및 튜토리얼
• Deep Java Library Pytorch - Pytorch용 Deep Java Library Engine Provider 소개 및 튜토리얼

Baseline Results - Food Recognition

Pretrained Model	Accuracy	Loss	epoch	note
VGG16	0.077	5.001	-	early stop, the performance is terrible
RESNET50	81.94	0.78	60	early stop,
RESNET152	73.77	0.973	20	comming soon!
WIDERESNET50_2	72.52	0.998	20	comming soon!
MOBILENET V2	81.96	0.72	240	cool, stop training
DENSENET121	45.94	4.3338e+7	40	early stop,

Baseline Results - Ingredient Recognition

Pretrained Model	Accuracy	Loss	epoch	note	dataset	num of class
VGG16	-	-	-	poor accuracy	INGD_V1 (private)	58
RESNET50	-	-	-	poor accuracy	INGD_V1 (private)	58
RESNET152	95.44	0.68	250	fruits and vegs only	Food and Vegetable Image Dataset	58
RESNET152	92.19	0.41	376	nice accuracy	INGD_V1 (private)	58
WIDERESNET50_2	-	-	-	poor accuracy	INGD_V1 (private)	58
MOBILENET V2	82.55	0.70	282	not bad	INGD_V1 (private)	58
DENSENET121	-	-	-	poor accuracy	INGD_V1 (private)	58

Stage2 Result - Ingredient Recognition

Pretrained Model	Accuracy	Loss	epoch	note	dataset	num of class
RESNET152	83.03	0.71	40	now available!	INGD_V2 (private)	238
MOBILENET V2				comming soon!	INGD_V2 (private)	238

Built With

• waverDeep - model architecture, setup train/test pipeline

Computing resources

GPU RESOURCE	RAM	COUNT	NOTE
NVIDIA TITAN RTX	24G	2	training
NVIDIA GeForce GTX 1080TI	12G	1	develop, test or etc

License

This project is licensed under the MIT License - see the LICENSE.md file for details