main.py

import datetime
import os

from torch.nn import Linear
from torchvision.transforms import v2

import data.dataset
from torch.optim.lr_scheduler import CosineAnnealingLR
import pandas
from torchmetrics.classification import MulticlassAccuracy, MulticlassAveragePrecision, MulticlassF1Score
from torchvision.models import mobilenet_v3_small, MobileNet_V3_Small_Weights
from torch import nn, optim

import torch
from tqdm import tqdm

from torch.utils.data import random_split

import mlflow

if __name__ == '__main__':
    mlflow.set_tracking_uri('http://localhost:5000')
    curr_date = datetime.datetime.now()
    os.mkdir(f"outputs/{curr_date}")
    # Input data files are available in the read-only "../input/" directory
    # For example, running this (by clicking run or pressing Shift+Enter) will list all files under the input directory
    with mlflow.start_run():
        device = torch.device("cuda")
        augmentation_transforms = v2.Compose([
            v2.RandomHorizontalFlip(),
            v2.RandomVerticalFlip(),
            v2.RandomGrayscale(),
            v2.RandomAutocontrast(),
            v2.RandomRotation(45),
        ]).to("cuda")
        dataset = data.dataset.OrangeDataset("/home/jarric/orange_dataset/processed/FIELD IMAGES/")

        train_size = int(0.75 * len(dataset))
        val_size = len(dataset) - train_size
        train_dataset, val_dataset = random_split(dataset, [train_size, val_size])

        # Create data loaders
        batch_size = 32

        train_loader = torch.utils.data.DataLoader(
            train_dataset,
            batch_size=batch_size,
            shuffle=False,
            num_workers=8,

        )

        val_loader = torch.utils.data.DataLoader(
            val_dataset,
            batch_size=batch_size,
            shuffle=False,
            num_workers=8,

        )
        epochs = 100

        mobilenet_v3_model = mobilenet_v3_small(weights=MobileNet_V3_Small_Weights.IMAGENET1K_V1.DEFAULT).to(device)

        for param in mobilenet_v3_model.parameters(): # freeze layers
            param.requires_grad = False

        mobilenet_v3_model.classifier[3] = Linear(in_features=1024, out_features=3, bias=True)
        mobilenet_v3_model.classifier[3].requires_grad = True

        mobilenet_v3_model.cuda()
        loss_fn = nn.CrossEntropyLoss().to(device)

        optimizer = optim.AdamW(
            mobilenet_v3_model.parameters(),
            lr=1e-4,
            weight_decay=1e-4)

        reporting_interval_train = 50
        reporting_interval_val = 10

        loss_fn = loss_fn.to(device)

        # metrics
        acc_metric = MulticlassAccuracy(num_classes=3).to(device)
        ap_metric = MulticlassAveragePrecision(num_classes=3, average="macro").to(device)
        f1_metric = MulticlassF1Score(num_classes=3).to(device)


        train_step = 0
        val_step = 0
        for epoch in tqdm(range(0, epochs)):
            train_loss = 0
            avg_accuracy = 0
            cur_iter = 0
            average_precision = 0
            f1_score_avg = 0
            mobilenet_v3_model.train()
            for images, labels in tqdm(train_loader, leave=False):
                images, labels = images.to(device), labels.to(device)

                images = augmentation_transforms(images)
                outputs = mobilenet_v3_model(images)
                loss = loss_fn(outputs, labels)
                train_loss += loss.item()

                optimizer.zero_grad()
                loss.backward()
                optimizer.step()
                _, predicted = torch.max(outputs.data, 1)
                avg_accuracy += acc_metric(predicted, labels)
                average_precision += ap_metric(outputs, labels)
                f1_score_avg += f1_metric(predicted, labels)

            train_loss /= len(train_loader)
            avg_accuracy /= len(train_loader)
            average_precision /= len(train_loader)
            f1_score_avg /= len(train_loader)
            mlflow.log_metric("train_loss", train_loss, step=epoch)
            mlflow.log_metric("train_avg_accuracy", avg_accuracy, step=epoch)
            mlflow.log_metric("train_average_precision", average_precision, step=epoch)
            mlflow.log_metric("f1_score_avg", f1_score_avg, step=epoch)

            val_loss = 0
            val_accuracy = 0
            cur_iter = 0
            average_precision_val = 0
            f1_score_avg_val = 0
            mobilenet_v3_model.eval()
            with torch.no_grad():
                for images, labels in tqdm(val_loader, leave=False):
                    images, labels = images.to(device), labels.to(device)

                    # Forward pass
                    outputs = mobilenet_v3_model(images)
                    loss = loss_fn(outputs, labels)
                    val_loss += loss.item()
                    _, predicted = torch.max(outputs.data, 1)
                    average_precision_val += ap_metric(outputs, labels)
                    val_accuracy += acc_metric(predicted, labels)
                    f1_score_avg_val += f1_metric(predicted, labels)

            val_loss /= len(val_loader)
            val_accuracy /= len(val_loader)
            average_precision_val /= len(val_loader)
            f1_score_avg_val /= len(val_loader)
            mlflow.log_metric("val_loss", val_loss, step=epoch)
            mlflow.log_metric("val_avg_accuracy", val_accuracy, step=epoch)
            mlflow.log_metric("val_average_precision", average_precision_val, step=epoch)
            mlflow.log_metric("val_f1_score_avg", f1_score_avg_val, step=epoch)


            torch.save(mobilenet_v3_model, f"outputs/{curr_date}/model_{epoch}_finetuned.pt")
            mlflow.log_artifact(f"outputs/{curr_date}/model_{epoch}_finetuned.pt")