Okay

How can i send

I think probably you are model is overfit i guess and it can occurs due to many causes. Could me share your code i will see throughly what is causing it dropping it to low i mean overfitting it .

why do not try using yolov11m Version I have got 82 score for this when i fine tuning it . Here is link to my Notebook on Kaggle https://www.kaggle.com/code/muhammadqasimshabbir/zindi1-2ghana-crop-disease/

https://www.kaggle.com/code/muhammadqasimshabbir/zindilacuna-malaria-detection-challenge/ this is notebook for this competeion

and should we drop class NEG for training?

This is my code where i do not drop negative class but still got 82 score yolov11m fine tuning

!pip install ultralytics !pip install tqdm import pandas as pd import os from pathlib import Path import shutil from sklearn.model_selection import train_test_split from tqdm.notebook import tqdm import cv2 import yaml import matplotlib.pyplot as plt from ultralytics import YOLO import multiprocessing # Set the data directory DATA_DIR = Path('/kaggle/input/lacuna-malaria-detection-dataset') # Load train and test files train = pd.read_csv(DATA_DIR / 'Train.csv') test = pd.read_csv(DATA_DIR / 'Test.csv') ss = pd.read_csv(DATA_DIR / 'SampleSubmission.csv') # Add an image_path column train['image_path'] = [Path('/kaggle/input/lacuna-malaria-detection-dataset/images' + x) for x in train.Image_ID] test['image_path'] = [Path('/kaggle/input/lacuna-malaria-detection-dataset/images' + x) for x in test.Image_ID] # Map string classes to integers (label encoding targets) train['class_id'] = train['class'].map({'Trophozoite': 0, 'NEG': 1, 'WBC': 2}) # Split data into training and validation train_unique_imgs_df = train.drop_duplicates(subset=['Image_ID'], ignore_index=True) X_train, X_val = train_test_split(train_unique_imgs_df, test_size=0.25, stratify=train_unique_imgs_df['class'], random_state=42) X_train = train[train.Image_ID.isin(X_train.Image_ID)] X_val = train[train.Image_ID.isin(X_val.Image_ID)] # Check the shapes of training and validation data print(X_train.shape, X_val.shape) # Define directories for images and labels TRAIN_IMAGES_DIR = Path('/kaggle/working/train/images') VAL_IMAGES_DIR = Path('/kaggle/working/val/images') TEST_IMAGES_DIR = Path('/kaggle/working/test/images') TRAIN_LABELS_DIR = Path('/kaggle/working/train/labels') VAL_LABELS_DIR = Path('/kaggle/working/val/labels') TEST_LABELS_DIR = Path('/kaggle/working/test/labels') # Create necessary directories for DIR in [TRAIN_IMAGES_DIR, VAL_IMAGES_DIR, TEST_IMAGES_DIR, TRAIN_LABELS_DIR, VAL_LABELS_DIR, TEST_LABELS_DIR]: if DIR.exists(): shutil.rmtree(DIR) DIR.mkdir(parents=True, exist_ok=True) # Copy train, val, and test images to their respective dirs for img in tqdm(X_train.image_path.unique()): shutil.copy(img, TRAIN_IMAGES_DIR / img.parts[-1]) for img in tqdm(X_val.image_path.unique()): shutil.copy(img, VAL_IMAGES_DIR / img.parts[-1]) for img in tqdm(test.image_path.unique()): shutil.copy(img, TEST_IMAGES_DIR / img.parts[-1]) # Function to convert the bounding boxes to YOLO format and save them def save_yolo_annotation(row): image_path, class_id, output_dir = row['image_path'], row['class_id'], row['output_dir'] img = cv2.imread(str(image_path)) if img is None: raise ValueError(f"Could not read image from path: {image_path}") height, width, _ = img.shape label_file = Path(output_dir) / f"{Path(image_path).stem}.txt" ymin, xmin, ymax, xmax = row['ymin'], row['xmin'], row['ymax'], row['xmax'] # Normalize the coordinates x_center = (xmin + xmax) / 2 / width y_center = (ymin + ymax) / 2 / height bbox_width = (xmax - xmin) / width bbox_height = (ymax - ymin) / height with open(label_file, 'a') as f: f.write(f"{class_id} {x_center:.6f} {y_center:.6f} {bbox_width:.6f} {bbox_height:.6f}\n") # Parallelize the annotation saving process def process_dataset(dataframe, output_dir): dataframe['output_dir'] = output_dir with multiprocessing.Pool() as pool: list(tqdm(pool.imap(save_yolo_annotation, dataframe.to_dict('records')), total=len(dataframe))) # Save train and validation labels to their respective dirs process_dataset(X_train, TRAIN_LABELS_DIR) process_dataset(X_val, VAL_LABELS_DIR) # Create a data.yaml file required by YOLO class_names = train['class'].unique().tolist() num_classes = len(class_names) data_yaml = { 'train': str(TRAIN_IMAGES_DIR), 'val': str(VAL_IMAGES_DIR), 'nc': num_classes, 'names': class_names } # Save the data.yaml file yaml_path = Path('data.yaml') with open(yaml_path, 'w') as file: yaml.dump(data_yaml, file, default_flow_style=False) # Load a YOLO pretrained model model = YOLO('yolov8m.pt').to(device)

# Fine-tune model to our data model.train( data='data.yaml', # Path to the dataset configuration epochs=10, # Number of epochs imgsz=1024, # Image size batch=16, # Batch size # device='0', device=device patience=3 ) # Validate the model on the validation set model.val() # Make predictions on test images and save them to a CSV file image_files = os.listdir(TEST_IMAGES_DIR) all_data = [] for image_file in tqdm(image_files): img_path = os.path.join(TEST_IMAGES_DIR, image_file) results = model(img_path) boxes = results[0].boxes.xyxy.tolist() classes = results[0].boxes.cls.tolist() confidences = results[0].boxes.conf.tolist() names = results[0].names incorrect_prediction = 0 if not boxes: incorrect_prediction = incorrect_prediction + 1 all_data.append({ 'Image_ID': image_file, 'class': 'NEG', 'confidence': 1.0, 'ymin': 0, 'xmin': 0, 'ymax': 0, 'xmax': 0 }) else: for box, cls, conf in zip(boxes, classes, confidences): x1, y1, x2, y2 = box detected_class = names[int(cls)] all_data.append({ 'Image_ID': image_file, 'class': detected_class, 'confidence': conf, 'ymin': y1, 'xmin': x1, 'ymax': y2, 'xmax': x2 }) print(f'========== Total Incorrect predction or dummy prediction {incorrect_prediction}================') # Convert the results to a DataFrame and save it sub = pd.DataFrame(all_data) sub.to_csv('/kaggle/working/benchmark_submission.csv', index=False)

map({'Trophozoite': 0, 'NEG': 1, 'WBC': 2}) #

Here

But here it seems like you used yolov8m not

Acutully When i published this notebook i was using old version of yolo which is yolov8 But latest version is Yolov11 which which i have set Here you can use this notebook only you have to make a small change here model = YOLO('yolov8m.pt')--->>> model = YOLO('yolov11m.pt') Yolov11 has four varaination (n(nano) , s(small) ,m(medium),l(large),x(extralarge))