MNIST - BlockAI

'''
---BlockAI/MNIST Auto Generate Code---
Author		: BlockAI
Project Name: MNIST
Project Link: https://blockai.kr/BlockAI/MNIST (BlockAI)
Create Date	: 2023-11-29

---Requirements---
# 사용자의 환경(OS, CUDA 등)에 따라 라이브러리 버전을 맞춰주세요
pip install torch==1.12 torchvision==0.13.1 torchtext==0.13.1 torchaudio==0.12.1
pip install pytorch-lightning==2.0.4
pip install tqdm
pip install pandas
pip install scikit-learn
pip install transformers
pip install timm

---Folder Structure---
--📂 data
   |--📂 train
      |--📁 LABEL_NAME_1
         |--🖼️ IMAGE_FILE
         |--🖼️ IMAGE_FILE
         |--🖼️ ...
      |--📁 LABEL_NAME_2
         |--🖼️ IMAGE_FILE
         |--🖼️ IMAGE_FILE
         |--🖼️ ...
      |--📁 ...
   |--📂 val
      |--📁 LABEL_NAME_1
         |--🖼️ IMAGE_FILE
         |--🖼️ IMAGE_FILE
         |--🖼️ ...
      |--📁 LABEL_NAME_2
         |--🖼️ IMAGE_FILE
         |--🖼️ IMAGE_FILE
         |--🖼️ ...
      |--📁 ...
   |--📂 test
      |--📁 LABEL_NAME_1
         |--🖼️ IMAGE_FILE
         |--🖼️ IMAGE_FILE
         |--🖼️ ...
      |--📁 LABEL_NAME_2
         |--🖼️ IMAGE_FILE
         |--🖼️ IMAGE_FILE
         |--🖼️ ...
      |--📁 ...
--📄 MNIST.py
--📄 MNIST.ipynb
--📄 requirements.txt
'''

import os
import argparse
import copy
import random
from glob import glob

from tqdm import tqdm
import numpy as np
import pandas as pd
from sklearn import preprocessing
from sklearn.model_selection import train_test_split
from PIL import Image

import torch
import pytorch_lightning as pl
import torchvision

path_sep = os.sep


# https://pytorch.org/tutorials/beginner/basics/data_tutorial.html#creating-a-custom-dataset-for-your-files
class Dataset(torch.utils.data.Dataset):
    def __init__(self, inputs, targets=[], input_transform=None):
        self.inputs = inputs
        self.targets = targets
        self.input_transform = input_transform

    # 학습 및 추론 과정에서 데이터를 1개씩 꺼내오는 곳
    def __getitem__(self, idx):
        # 정답이 있다면 if문을, 없다면 else문을 수행합니다
        if len(self.targets) == 0:
            return self.input_transform(Image.open(self.inputs[idx]))
        else:
            return self.input_transform(Image.open(self.inputs[idx])), torch.tensor(self.targets[idx])

    # 입력하는 개수만큼 데이터를 사용합니다
    # 'return 100'이면 1에폭에 100개의 데이터만 사용합니다
    def __len__(self):
        return len(self.inputs)


# https://pytorch-lightning.readthedocs.io/en/stable/extensions/datamodules.html
class Dataloader(pl.LightningDataModule):
# 데이터의 종류에 따라 코드 수정이 필요할 수 있습니다
    def __init__(self, data_folder, batch_size, train_ratio, shuffle):
        super().__init__()
        self.data_folder = data_folder
        self.batch_size = batch_size
        self.train_ratio = train_ratio
        self.shuffle = shuffle

        self.train_dataset = None
        self.val_dataset = None
        self.test_dataset = None
        self.predict_dataset = None

    def get_inputs_targets(self, data):
        inputs = []
        targets = []

        for data_path in data:
            # 이미지 경로를 저장합니다
            inputs.append(data_path)
            # 이미지 경로에서 타겟 값을 분리하여 저장합니다
            targets.append(data_path.split(path_sep)[-2])

        return inputs, targets

    def setup(self, stage='fit'):
        # train 폴더에 있는 모든 이미지 경로를 받아옵니다
        # 학습 데이터와 검증 데이터셋을 비율에 맞춰 분리하고, 데이터셋을 만듭니다
        total_data = sorted(glob(os.path.join(self.data_folder, 'train', '*', '*')))
        total_inputs, total_targets = self.get_inputs_targets(total_data)
        train_inputs, val_inputs, train_targets, val_targets = train_test_split(total_inputs, total_targets, train_size=self.train_ratio)

        # test 폴더에 있는 모든 이미지 경로를 받아와서 데이터셋을 만듭니다
        test_data = sorted(glob(os.path.join(self.data_folder, 'test', '*', '*')))
        test_inputs, test_targets = self.get_inputs_targets(test_data)

        # target 값을 encoding 해줍니다
        self.target_encoder = preprocessing.LabelEncoder()
        train_targets = self.target_encoder.fit_transform(train_targets)
        val_targets = self.target_encoder.transform(val_targets)
        test_targets = self.target_encoder.transform(test_targets)

        train_transform = torchvision.transforms.Compose([
            torchvision.transforms.v2.ToTensor(),
            torchvision.transforms.v2.Normalize(mean=0.1307, std=0.3081, inplace=False),
        ])
        test_transform = torchvision.transforms.Compose([
            torchvision.transforms.v2.ToTensor(),
            torchvision.transforms.v2.Normalize(mean=0.1307, std=0.3081, inplace=False),
        ])

        self.train_dataset = Dataset(train_inputs, train_targets, train_transform)
        self.val_dataset = Dataset(val_inputs, val_targets, train_transform)
        self.test_dataset = Dataset(test_inputs, test_targets, test_transform)
        self.predict_dataset = Dataset(test_inputs, [], test_transform)

    def train_dataloader(self):
        return torch.utils.data.DataLoader(self.train_dataset, batch_size=self.batch_size, shuffle=args.shuffle)

    def val_dataloader(self):
        return torch.utils.data.DataLoader(self.val_dataset, batch_size=self.batch_size)

    def test_dataloader(self):
        return torch.utils.data.DataLoader(self.test_dataset, batch_size=self.batch_size)

    def predict_dataloader(self):
        return torch.utils.data.DataLoader(self.predict_dataset, batch_size=self.batch_size)


# https://pytorch-lightning.readthedocs.io/en/stable/common/lightning_module.html
class Model(pl.LightningModule):
    def __init__(self):
        super().__init__()
        self.save_hyperparameters()

        self.linear_1 = torch.nn.Linear(in_features=784, out_features=4096)
        self.linear_2 = torch.nn.Linear(in_features=4096, out_features=512)
        self.linear_3 = torch.nn.Linear(in_features=512, out_features=10)

        self.dropout_1 = torch.nn.Dropout(p=0.2)
        self.relu_1 = torch.nn.ReLU()
        self.crossentropyloss_1 = torch.nn.CrossEntropyLoss()

    def forward(self, x_mnistimage):
        x_0 = torch.flatten(x_mnistimage)
        x_0 = self.linear_1(x_0)
        x_0 = self.dropout_1(x_0)
        x_0 = self.relu_1(x_0)
        x_0 = self.linear_2(x_0)
        x_0 = self.dropout_1(x_0)
        x_0 = self.relu_1(x_0)
        x_0 = self.linear_3(x_0)

        return x_0

    def training_step(self, batch, batch_idx):
        x_mnistimage, y = batch
        x_0 = self(x_mnistimage)
        loss = self.crossentropyloss_1(x_0, y.long().squeeze())
        self.log("train_loss", loss)

        return loss

    def validation_step(self, batch, batch_idx):
        x_mnistimage, y = batch
        x_0 = self(x_mnistimage)
        loss = self.crossentropyloss_1(x_0, y.long().squeeze())
        self.log("val_loss", loss)

        return loss

    def test_step(self, batch, batch_idx):
        x_mnistimage, y = batch
        x_0 = self(x_mnistimage)
        loss = self.crossentropyloss_1(x_0, y.long().squeeze())
        self.log("test_loss", loss)

        return loss

    def predict_step(self, batch, batch_idx):
        x_mnistimage = batch
        x_0 = self(x_mnistimage)

        return x_0

    def configure_optimizers(self):
        # 이곳에서 lr 값을 변경할 수 있습니다
        optimizer = torch.optim.AdamW(self.parameters(), lr=0.0005)
        return optimizer


if __name__ == '__main__':
    # https://docs.python.org/ko/3/library/argparse.html
    # 하이퍼 파라미터 등 각종 설정값을 입력받습니다
    # 터미널 실행 예시 : python3 run.py --batch_size=64 ...
    # 실행 시 '--batch_size=64' 같은 인자를 입력하지 않으면 default 값이 기본으로 실행됩니다
    parser = argparse.ArgumentParser()
    parser.add_argument('--data_folder', default='data')
    parser.add_argument('--batch_size', default=32)
    parser.add_argument('--max_epoch', default=1)
    parser.add_argument('--shuffle', default=True)
    parser.add_argument('--train_ratio', default=0.9)
    args = parser.parse_args()

    dataloader = Dataloader(args.data_folder, args.batch_size, args.train_ratio, args.shuffle)
    model = Model()
    # https://pytorch-lightning.readthedocs.io/en/stable/common/trainer.html
    # 학습 및 추론을 위한 Trainer 설정
    trainer = pl.Trainer(accelerator='gpu', devices=1, max_epochs=args.max_epoch)
    trainer.fit(model=model, datamodule=dataloader)
    # trainer.test(model=model, datamodule=dataloader)
    # predictions = trainer.predict(model=model, datamodule=dataloader)