Install the ultralytics package for YOLOv8 (if not already installed) !pip install…
- User
Install the ultralytics package for YOLOv8 (if not already installed)
!pip install ultralytics
Import necessary libraries
import pandas as pd import numpy as np import torch from torch.utils.data import Dataset, DataLoader from tqdm.notebook import tqdm from torchvision import transforms from PIL import Image import torch.nn as nn
Suppress warnings from YOLO and transformers
import warnings warnings.filterwarnings("ignore")
# Mount Google Drive if needed
from google.colab import drive
drive.mount('/content/drive')
Load the CSV file
csv_file = '/kaggle/input/scenario32/scenario32_dev.csv' # Replace with your actual path df = pd.read_csv(csv_file)
Import YOLOv8 model
from ultralytics import YOLO
Suppress YOLO logging output
import logging logging.getLogger('ultralytics').setLevel(logging.ERROR)
Load the pre-trained YOLOv8 segmentation model
yolo_model = YOLO('yolov8l-seg.pt') # Using 'n'ano model for speed and low resource consumption
Create a custom dataset class
class BeamDataset(Dataset): def init(self, dataframe, transform=None, label_mapping=None): self.data = dataframe.reset_index(drop=True) self.image_paths = self.data['scenario32/'+'unit1_rgb'] self.labels = self.data['unit1_beam'] self.transform = transform self.label_mapping = label_mapping
# Map labels to indices self.labels = self.labels.map(self.label_mapping).astype(int) def __len__(self): return len(self.data) def __getitem__(self, idx): img_path = self.image_paths.iloc[idx] # Load image image = Image.open(img_path).convert('RGB') # Perform YOLOv8 segmentation without verbose output results = yolo_model(img_path, verbose=False) # Get the segmentation mask masks = results[0].masks.data.cpu().numpy() # Shape: (num_objects, H, W) # If no objects detected, create an empty mask if len(masks) == 0: seg_mask = np.zeros((image.height, image.width), dtype=np.uint8) else: # Combine all masks into a single mask seg_mask = np.sum(masks, axis=0) seg_mask = np.clip(seg_mask, 0, 1).astype(np.uint8) # Resize to match image size if necessary seg_mask = Image.fromarray(seg_mask * 255).resize(image.size) # Convert segmentation mask to tensor if self.transform: image = self.transform(image) seg_mask = self.transform(seg_mask.convert('RGB')) else: transform = transforms.ToTensor() image = transform(image) seg_mask = transform(seg_mask.convert('RGB')) label = self.labels.iloc[idx] return {'image': image, 'seg_mask': seg_mask, 'label': label}Define the transforms
from transformers import ViTImageProcessor
Use a smaller model for practical GPU memory usage
model_name = 'google/vit-large-patch16-224-in21k' image_processor = ViTImageProcessor.from_pretrained(model_name)
common_transforms = transforms.Compose([ transforms.Resize((224, 224)), transforms.ToTensor(), transforms.Normalize(mean=image_processor.image_mean, std=image_processor.image_std) ])
Create label mapping
all_labels = df['unit1_beam'].unique() label_mapping = {label: idx for idx, label in enumerate(sorted(all_labels))} num_labels = len(label_mapping)
Split the dataset into training and testing sets
from sklearn.model_selection import train_test_split
train_df, test_df = train_test_split(df, test_size=0.2, random_state=42)
train_dataset = BeamDataset(train_df, transform=common_transforms, label_mapping=label_mapping) test_dataset = BeamDataset(test_df, transform=common_transforms, label_mapping=label_mapping)
train_loader = DataLoader(train_dataset, batch_size=8, shuffle=True, num_workers=0) test_loader = DataLoader(test_dataset, batch_size=8, shuffle=False, num_workers=0)
Load the ViT model
from transformers import ViTModel
Load pre-trained ViT models for images and segmentation masks
vit_image = ViTModel.from_pretrained(model_name) vit_segmentation = ViTModel.from_pretrained(model_name)
Freeze ViT models if desired
for param in vit_image.parameters():
param.requires_grad = False
for param in vit_segmentation.parameters():
param.requires_grad = False
Define the combined model
class BeamPredictionModel(nn.Module): def init(self, num_labels): super(BeamPredictionModel, self).init() self.vit_image = vit_image self.vit_segmentation = vit_segmentation self.dropout = nn.Dropout(0.1) self.classifier = nn.Linear(vit_image.config.hidden_size * 2, num_labels)
def forward(self, image, seg_mask): # Pass through ViT models image_outputs = self.vit_image(image) seg_outputs = self.vit_segmentation(seg_mask) # Concatenate embeddings combined_representation = torch.cat( (image_outputs.pooler_output, seg_outputs.pooler_output), dim=1 ) # Classification x = self.dropout(combined_representation) logits = self.classifier(x) return logitsInitialize the model
num_labels = len(label_mapping)
model = BeamPredictionModel(num_labels) device = torch.device('cuda' if torch.cuda.is_available() else 'cpu') model.to(device)
criterion = nn.CrossEntropyLoss() optimizer = torch.optim.AdamW(model.parameters(), lr=1e-5)
Use mixed precision
scaler = torch.cuda.amp.GradScaler()
Import matplotlib for plotting
import matplotlib.pyplot as plt
Initialize lists to store metrics
train_losses = [] val_losses = [] val_top1_accuracies = [] val_top2_accuracies = [] val_top3_accuracies = [] val_dba_scores = []
Training loop
num_epochs = 5 # Adjust the number of epochs as needed
for epoch in range(num_epochs): model.train() total_loss = 0 progress_bar = tqdm(train_loader, desc=f"Epoch {epoch+1}/{num_epochs}")
for batch in progress_bar: images = batch['image'].to(device) seg_masks = batch['seg_mask'].to(device) labels = batch['label'].to(device) optimizer.zero_grad() with torch.cuda.amp.autocast(): logits = model(images, seg_masks) loss = criterion(logits, labels) scaler.scale(loss).backward() scaler.step(optimizer) scaler.update() total_loss += loss.item() progress_bar.set_postfix(loss=loss.item()) avg_loss = total_loss / len(train_loader) train_losses.append(avg_loss) print(f"Epoch [{epoch+1}/{num_epochs}], Training Loss: {avg_loss:.4f}") # Evaluate on the test set model.eval() total_correct_top1 = 0 total_correct_top2 = 0 total_correct_top3 = 0 total_samples = 0 total_dba = 0 total_val_loss = 0 with torch.no_grad(): for batch in test_loader: images = batch['image'].to(device) seg_masks = batch['seg_mask'].to(device) labels = batch['label'].to(device) logits = model(images, seg_masks) val_loss = criterion(logits, labels) total_val_loss += val_loss.item() # Get top-k predictions _, preds_topk = logits.topk(3, dim=1) correct_topk = preds_topk.eq(labels.view(-1, 1)) # Shape: (batch_size, 3) # Compute Top-1 accuracy total_correct_top1 += correct_topk[:, 0].sum().item() # Compute Top-2 accuracy total_correct_top2 += correct_topk[:, :2].any(dim=1).sum().item() # Compute Top-3 accuracy total_correct_top3 += correct_topk.any(dim=1).sum().item() # Compute DBA score probs = torch.softmax(logits, dim=1) batch_size = labels.size(0) for i in range(batch_size): prob = probs[i] label = labels[i] sorted_probs, indices = prob.sort(descending=True) rank = (indices == label).nonzero(as_tuple=True)[0].item() + 1 # Rank starts from 1 total_dba += 1.0 / rank total_samples += labels.size(0) avg_val_loss = total_val_loss / len(test_loader) val_losses.append(avg_val_loss) top1_accuracy = total_correct_top1 / total_samples top2_accuracy = total_correct_top2 / total_samples top3_accuracy = total_correct_top3 / total_samples avg_dba = total_dba / total_samples val_top1_accuracies.append(top1_accuracy) val_top2_accuracies.append(top2_accuracy) val_top3_accuracies.append(top3_accuracy) val_dba_scores.append(avg_dba) print(f"Validation Loss: {avg_val_loss:.4f}") print(f"Validation Top-1 Accuracy: {top1_accuracy:.4f}") print(f"Validation Top-2 Accuracy: {top2_accuracy:.4f}") print(f"Validation Top-3 Accuracy: {top3_accuracy:.4f}") print(f"Validation DBA Score: {avg_dba:.4f}")Plotting Training Loss over Epochs
plt.figure(figsize=(8, 6)) plt.plot(range(1, num_epochs+1), train_losses, label='Training Loss') plt.plot(range(1, num_epochs+1), val_losses, label='Validation Loss') plt.xlabel('Epoch') plt.ylabel('Loss') plt.title('Training and Validation Loss over Epochs') plt.legend() plt.grid(True) plt.show()
Plotting Top-1, Top-2, Top-3 Accuracy over Epochs
plt.figure(figsize=(8, 6)) plt.plot(range(1, num_epochs+1), val_top1_accuracies, label='Top-1 Accuracy') plt.plot(range(1, num_epochs+1), val_top2_accuracies, label='Top-2 Accuracy') plt.plot(range(1, num_epochs+1), val_top3_accuracies, label='Top-3 Accuracy') plt.xlabel('Epoch') plt.ylabel('Accuracy') plt.title('Top-1, Top-2, Top-3 Accuracy over Epochs') plt.legend() plt.grid(True) plt.show()
Plotting DBA Score over Epochs
plt.figure(figsize=(8, 6)) plt.plot(range(1, num_epochs+1), val_dba_scores, label='DBA Score', marker='o') plt.xlabel('Epoch') plt.ylabel('DBA Score') plt.title('DBA Score over Epochs') plt.legend() plt.grid(True) plt.show()
what is error in this code
- Assistant
The error in your code is due to incorrect handling of the image paths in your custom
BeamDatasetclass. Specifically, the line:self.image_paths = self.data['scenario32/'+'unit1_rgb']is causing a
KeyErrorbecause you're attempting to access a column named'scenario32/unit1_rgb'in your dataframe, which likely doesn't exist. Instead, you probably have a column named'unit1_rgb'that contains relative paths or filenames of your images.Why This Causes an Error:
- Incorrect Column Name: By concatenating
'scenario32/'with'unit1_rgb'inside the brackets, you're looking for a column in the dataframe named'scenario32/unit1_rgb', which is unlikely to exist. - Misuse of String Concatenation: You're trying to concatenate a directory path to the data inside the dataframe, but you're doing it in a way that alters the column name instead of the data within the column.
How to Fix It:
You need to construct the full image paths by concatenating the base directory path with the filenames provided in the
'unit1_rgb'column. Here's how you can modify your__init__method in theBeamDatasetclass:class BeamDataset(Dataset): def __init__(self, dataframe, transform=None, label_mapping=None): self.data = dataframe.reset_index(drop=True) # Correctly construct the image paths self.image_paths = '/kaggle/input/scenario32/' + self.data['unit1_rgb'] self.labels = self.data['unit1_beam'] self.transform = transform self.label_mapping = label_mapping # Map labels to indices self.labels = self.labels.map(self.label_mapping).astype(int)Explanation:
- Constructing Full Paths: By adding
'/kaggle/input/scenario32/'to the front ofself.data['unit1_rgb'], you create the full paths to your image files. - Accessing Correct Column: Ensure you're accessing the correct column
'unit1_rgb'from your dataframe.
Your Corrected Dataset Class:
class BeamDataset(Dataset): def __init__(self, dataframe, transform=None, label_mapping=None): self.data = dataframe.reset_index(drop=True) self.image_paths = '/kaggle/input/scenario32/' + self.data['unit1_rgb'] self.labels = self.data['unit1_beam'] self.transform = transform self.label_mapping = label_mapping # Map labels to indices self.labels = self.labels.map(self.label_mapping).astype(int) def __len__(self): return len(self.data) def __getitem__(self, idx): img_path = self.image_paths.iloc[idx] # Load image image = Image.open(img_path).convert('RGB') # Perform YOLOv8 segmentation without verbose output results = yolo_model(img_path, verbose=False) # Get the segmentation mask masks = results[0].masks.data.cpu().numpy() # Shape: (num_objects, H, W) # If no objects detected, create an empty mask if len(masks) == 0: seg_mask = np.zeros((image.height, image.width), dtype=np.uint8) else: # Combine all masks into a single mask seg_mask = np.sum(masks, axis=0) seg_mask = np.clip(seg_mask, 0, 1).astype(np.uint8) # Resize to match image size if necessary seg_mask = Image.fromarray(seg_mask * 255).resize(image.size) # Convert segmentation mask to tensor if self.transform: image = self.transform(image) seg_mask = self.transform(seg_mask.convert('RGB')) else: transform = transforms.ToTensor() image = transform(image) seg_mask = transform(seg_mask.convert('RGB')) label = self.labels.iloc[idx] return {'image': image, 'seg_mask': seg_mask, 'label': label}Additional Tips:
- Check the DataFrame Columns: It's always good practice to check the columns of your dataframe using
print(df.columns)to ensure you're accessing them correctly. - Verify File Existence: You may also want to verify that the constructed file paths actually point to existing files to avoid
FileNotFoundError. - Handle Relative Paths: If your image paths in
'unit1_rgb'are already full paths, you may not need to prepend'/kaggle/input/scenario32/'.
Example Verification:
print(df.columns) print(df['unit1_rgb'].head()) print(df['unit1_beam'].head())By correcting the way you construct your image paths, your code should run without the
KeyErrorand proceed with training your model. - Incorrect Column Name: By concatenating