script.py

import pandas as pd
import numpy as np
import os
from tqdm import tqdm
from PIL import Image
import torch
import torch.nn as nn
import torchvision.transforms as T
from torchvision.models import resnet50

def is_gpu_available():
    return torch.cuda.is_available()

class ResNetClassifier(nn.Module):
    def __init__(self, num_classes, metadata_size):
        super(ResNetClassifier, self).__init__()
        self.resnet = resnet50(pretrained=True)
        self.resnet.fc = nn.Identity()  # Remove the fully connected layer
        self.metadata_fc = nn.Linear(metadata_size, 128)
        self.classifier = nn.Linear(2048 + 128, num_classes)  # 2048 is the output size of ResNet50

    def forward(self, x, metadata_features):
        resnet_features = self.resnet(x)
        metadata_features = self.metadata_fc(metadata_features)
        combined_features = torch.cat((resnet_features, metadata_features), dim=1)
        logits = self.classifier(combined_features)
        return logits

class PytorchWorker:
    def __init__(self, model_path: str, num_classes: int, metadata_size: int):
        self.device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
        print(f"Using device: {self.device}")
        self.model = self._load_model(model_path, num_classes, metadata_size)
        self.transforms = T.Compose([T.Resize((224, 224)),
                                     T.ToTensor(),
                                     T.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])])

    def _load_model(self, model_path, num_classes, metadata_size):
        model = ResNetClassifier(num_classes, metadata_size)
        model.load_state_dict(torch.load(model_path, map_location=self.device))
        return model.to(self.device).eval()

    def predict_image(self, image: Image.Image, metadata_features: np.ndarray) -> list:
        input_tensor = self.transforms(image).unsqueeze(0).to(self.device)
        metadata_tensor = torch.tensor(metadata_features).unsqueeze(0).to(self.device)
        with torch.no_grad():
            logits = self.model(input_tensor, metadata_tensor)
        return logits.tolist()

def make_submission(test_metadata, model_path, num_classes, metadata_size, output_csv_path="./submission.csv", images_root_path="/tmp/data/private_testset"):
    model = PytorchWorker(model_path, num_classes, metadata_size)
    predictions = []
    for _, row in tqdm(test_metadata.iterrows(), total=len(test_metadata)):
        image_path = os.path.join(images_root_path, row['image_path'])
        test_image = Image.open(image_path).convert("RGB")
        metadata_features = row.drop(['image_path', 'class_id']).values.astype(np.float32)
        logits = model.predict_image(test_image, metadata_features)
        predictions.append(np.argmax(logits))
    test_metadata["class_id"] = predictions
    user_pred_df = test_metadata.drop_duplicates("observation_id", keep="first")
    user_pred_df[["observation_id", "class_id"]].to_csv(output_csv_path, index=None)

if __name__ == "__main__":
    import zipfile
    with zipfile.ZipFile("/tmp/data/private_testset.zip", 'r') as zip_ref:
        zip_ref.extractall("/tmp/data")
    
    MODEL_PATH = "pytorch_model.pth"
    metadata_file_path = "./SnakeCLEF2024-TestMetadata.csv"
    test_metadata = pd.read_csv(metadata_file_path)
    num_classes = 1784
    metadata_size = len(test_metadata.columns) - 2  # Excluding 'image_path' and 'class_id'

    make_submission(
        test_metadata=test_metadata,
        model_path=MODEL_PATH,
        num_classes=num_classes,
        metadata_size=metadata_size
    )