Japanese Food Classification

This repository contains training, evaluation and testing of classification models implementated from stratch using PyTorch.

Classes

Ramen 🍜 , Sashimi 🐟 , Sushi 🍣 , Takoyaki 🍡

Available Models

Model	Variants
VGG	11, 13, 16, 19
Resnet	18, 34, 50, 101, 152
Mobilenet	v2, v3-small, v3-large
Shufflenet	v2
ConvNeXt	tiny, small, base, large

1. Dependencies

Please install essential dependencies (see requirements.txt)

1.1. Create conda virtual environment

conda create -n jpfood python==3.8

1.2. Activate environment

conda activate jpfood

1.3. Install dependencies

pip install -r requirements.txt

2. Dataset

The dataset used in this repository is a subset of Food-101 dataset.
The original dataset contains 1000 images for each class.

In this repository, train - 200, validation - 100 and test - 100 images are used.
Japanese food dataset can be downloaded from link

3. Prepare pretrained models

Model implementations are available in models/{model}.py folder.
To initialize PyTorch's pretrained model weights to our model architecture, run:

python convert.py --model resnet --variant 50 

options:
    --model : Model name {vgg, resnet, mobilenet}
    --variant : Model variant 
    --width_mulit : channel width ratio for mobilenet
    --num_class : 1000 default for ImageNet classes

Pretrained models will be saved as pretrained/{model}.pt

4. Prepare Config

Change model and variant of desired model before training and testing in config.yaml.

model:
    name: 'resnet'
    variant: '50'

Hyperparameters recommendation:

Model	Optimizer	Learning_rate
VGG	sgd	0.001
Resnet	sgd	0.001
Mobilenet	sgd	0.001
Shufflenet	sgd	0.005
ConvNeXt	adamw	0.00005

5. Training and evaluation

For training, run:

python train.py

options:
    --eval : To evaluate the model without training again
    --eval_model : model path for evaluation

The result for each model training is saved in experiments/{model_name}{datetime}/ folder.

6. Inference

For inference, run:

python inference.py --image {image_path} --viz

options:
    --image : Folder path to inference images
    --viz   : Whether to visualize inferece results

for C++ inference, change required inputs and run:

python trace_model.py 

to convert pytorch model to script model. After tracing the model, use it in libtorch-inference repo.

Inference Result

\

Reference

• Torch Vision
• Zero to Mastery Learn PyTorch for Deep Learning