YOLO-NAS Training with ClassBalancedSampler & YoloDarknetFormatDetectionDataset triggers RuntimeError

[dkrief]

🐛 Describe the bug

I'm experiencing a bug while attempting to train a YOLO-NAS model using the YoloDarknetFormatDetectionDataset. The error occurs when invoking the train method, specifically in the _bbox_loss function within the PPYoloELoss module of the SUPERGRADIENTS library. This issue is observed only when using the YoloDarknetFormatDetectionDataset (CODE_2), whereas using coco_detection_yolo_format_train (CODE_1) does not result in this error.

Details & Error Logs:

In CODE_2, while implementing a class-balanced sampling strategy for overcoming class imbalance in a dataset, the following runtime error surfaces during model training:

RuntimeError: numel: integer multiplication overflow

Steps to Reproduce:

1. Use the YoloDarknetFormatDetectionDataset to load datasets for training.
2. Employ the ClassBalancedSampler and initialize the DataLoader.
3. Execute the training process.
4. Encounter the overflow error within the loss computation method _bbox_loss.

Additional Information:

The error seems to come from the line involving torch.masked_select(pred_bboxes, bbox_mask).reshape([-1, 4]) in super_gradients/training/losses/ppyolo_loss.py

Code

import os
import logging
import yaml
import itertools
from super_gradients.training import models, Trainer
from super_gradients.training.losses import PPYoloELoss
from super_gradients.training.datasets.detection_datasets import YoloDarknetFormatDetectionDataset
from torch.utils.data import DataLoader
from tqdm import tqdm
import numpy as np
import glob
from super_gradients.training.utils.collate_fn import DetectionCollateFN
from super_gradients.training.datasets.samplers.class_balanced_sampler import ClassBalancedSampler
import torch

DATASET_RELATIVE_PATH = 'ComputerUse_original_resized'  ## yolov5 format

# Configure logging (if needed)
def get_class_frequency(dataloader):
    class_frequency = {}
    for batch_x, batch_y in tqdm(dataloader, total=len(dataloader), desc="Calculating class frequencies"):
        for cls in batch_y[:, -1]:
            cls = int(cls.item())
            class_frequency[cls] = class_frequency.get(cls, 0) + 1

    return class_frequency

# Configure the logging system
logging.basicConfig(level=logging.DEBUG)

def main():

    HOME = os.getcwd()
    logging.info("Working directory: %s", HOME)
    DEVICE = 'cuda' if torch.cuda.is_available() else 'cpu'
    logging.info("Device: %s", DEVICE)

    EXPERIMENT_NAME = 'test'
    max_epochs = 100
    initial_lr = 1e-4

    if not DATASET_RELATIVE_PATH:
        raise ValueError(
            "For local datasets, 'DATASET_RELATIVE_PATH' must be specified.")

    if not os.path.isdir(DATASET_RELATIVE_PATH):
        raise FileNotFoundError(
            f"Local dataset directory not found at: {DATASET_RELATIVE_PATH}")

    # Path to data.yaml within the local dataset directory
    data_yaml_path = os.path.join(DATASET_RELATIVE_PATH, 'data.yaml')
    if not os.path.isfile(data_yaml_path):
        raise FileNotFoundError(
            f"data.yaml not found in the local dataset directory: {data_yaml_path}")

    # Load and parse data.yaml
    with open(data_yaml_path, 'r') as f:
        try:
            data_yaml = yaml.safe_load(f)
        except yaml.YAMLError as e:
            raise ValueError(f"Error parsing data.yaml: {e}")

    # Extract class names
    CLASSES = data_yaml.get('names')
    if not CLASSES:
        raise ValueError(
            "No class names found in data.yaml under 'names' key.")

    logging.info(f"Loaded classes from data.yaml: {CLASSES}")

    # Extract data paths
    train_images_dir = data_yaml.get('train')
    val_images_dir = data_yaml.get('val')
    test_images_dir = data_yaml.get('test')  # Optional

    if not all([train_images_dir, val_images_dir]):
        raise ValueError("data.yaml must contain 'train' and 'val' paths.")

    # Define label directories based on data.yaml paths
    train_labels_dir = os.path.join(
        os.path.dirname(train_images_dir), 'labels')
    val_labels_dir = os.path.join(
        os.path.dirname(val_images_dir), 'labels')
    test_labels_dir = os.path.join(os.path.dirname(
        test_images_dir), 'labels') if test_images_dir else None

    # Preprocess to remove label files not associated with any image
    def clean_label_files(images_dir, labels_dir):
        image_files = set(os.path.splitext(os.path.basename(f))[0] for f in glob.glob(
            os.path.join(images_dir, "*.*")) if os.path.isfile(f))
        label_files = glob.glob(os.path.join(labels_dir, "*.txt"))
        orphan_labels = []
        for label_file in label_files:
            label_name = os.path.splitext(os.path.basename(label_file))[0]
            if label_name not in image_files:
                orphan_labels.append(label_file)
        if orphan_labels:
            logging.warning(
                f"Removing {len(orphan_labels)} orphan label files from {labels_dir}")
            for label_file in orphan_labels:
                os.remove(label_file)

    clean_label_files(train_images_dir, train_labels_dir)
    clean_label_files(val_images_dir, val_labels_dir)
    if test_images_dir and test_labels_dir:
        clean_label_files(test_images_dir, test_labels_dir)

    CHECKPOINT_DIR = os.path.join(HOME, 'checkpoints')


    # Initialize training dataset
    train_dataset = YoloDarknetFormatDetectionDataset(
                    data_dir=DATASET_RELATIVE_PATH, images_dir="train/images", labels_dir="train/labels", classes=CLASSES
                )

    # Initialize validation dataset
    val_dataset = YoloDarknetFormatDetectionDataset(
                    data_dir=DATASET_RELATIVE_PATH, images_dir="valid/images", labels_dir="valid/labels", classes=CLASSES
                )

    # Get initial class balance for training
    initial_balance = train_dataset.get_dataset_classes_information()
    initial_class_balance = initial_balance.sum(axis=0)
    initial_discrepancy = initial_class_balance.max() / initial_class_balance.min()

    print(
        f" \n\nBEFORE BALANCE:"
        f" Most frequent class (#{np.argmax(initial_class_balance)}) appears {initial_class_balance.max()} times, which is {initial_discrepancy:.2f}x"
        f" more frequent than the least frequent class (#{np.argmin(initial_class_balance)}) that appears only {initial_class_balance.min()} times!"
    )
    
    # Create DataLoader for training without sampling
    vanilla_dataloader = DataLoader(
        train_dataset,
        batch_size=8,
        drop_last=False,
        collate_fn=DetectionCollateFN()
    )
    vanilla_sampled_class_balance = np.zeros_like(initial_class_balance)

    # Calculate class frequencies for vanilla DataLoader
    for k, v in get_class_frequency(vanilla_dataloader).items():
        vanilla_sampled_class_balance[k] = v


    # Verify that vanilla sampling matches initial class balance
    np.testing.assert_equal(vanilla_sampled_class_balance, initial_class_balance)  # no special sampling

    # Define parameter grid for oversampling
    #oversample_thresholds = [None, 0.5, 1.0, 1.5, 2.0, 3.0, 3.5, 4.0, 4.5, 5.0]
    #oversample_aggressivenesses = [0.5, 1.0, 1.5, 2.0, 3.0, 3.5, 4.0, 4.5, 5.0]
    oversample_thresholds = [2.5]
    oversample_aggressivenesses = [3.0]
    parameter_grid = list(itertools.product(oversample_thresholds, oversample_aggressivenesses))

    best_discrepancy = float('inf')
    best_setting = None
    
    # Iterate over all settings to find the best one
    for oversample_threshold, oversample_aggressiveness in parameter_grid:
        sampler = ClassBalancedSampler(
            train_dataset,
            oversample_threshold=oversample_threshold,
            oversample_aggressiveness=oversample_aggressiveness
        )
        train_dataloader_balanced = DataLoader(
            train_dataset,
            batch_size=8,
            drop_last=False,
            collate_fn=DetectionCollateFN(),
            sampler=sampler
        )
        balanced_sampled_class_balance = np.zeros_like(initial_class_balance)

        # Calculate class frequencies for balanced DataLoader
        for k, v in get_class_frequency(train_dataloader_balanced).items():
            balanced_sampled_class_balance[k] = v

        # Handle division by zero if any class has zero samples
        min_count = balanced_sampled_class_balance.min()
        if min_count == 0:
            print(f"Skipping setting (oversample_threshold={oversample_threshold}, oversample_aggressiveness={oversample_aggressiveness}) due to zero samples in a class.")
            continue  # Skip this setting as it's invalid

        balanced_discrepancy = balanced_sampled_class_balance.max() / min_count

        print(
            f"AFTER BALANCE (oversample_threshold={oversample_threshold}, oversample_aggressiveness={oversample_aggressiveness}):"
            f" Most frequent class (#{np.argmax(balanced_sampled_class_balance)}) appears {balanced_sampled_class_balance.max()} times,"
            f" which is {balanced_discrepancy:.2f}x more frequent than the least frequent class (#{np.argmin(balanced_sampled_class_balance)})"
            f" that appears only {balanced_sampled_class_balance.min()} times!"
        )

        # Update best setting if current discrepancy is lower
        if balanced_discrepancy < best_discrepancy:
            best_discrepancy = balanced_discrepancy
            best_setting = (oversample_threshold, oversample_aggressiveness)

    if best_setting is None:
        raise ValueError("No valid sampling settings found. Please check your parameter grid and dataset.")

    print(
        f"\nOptimal Sampling Settings:"
        f" oversample_threshold={best_setting[0]}, oversample_aggressiveness={best_setting[1]}"
        f" with balanced_discrepancy={best_discrepancy:.2f}"
    )

    optimal_sampler = ClassBalancedSampler(
        train_dataset,
        oversample_threshold=best_setting[0],
        oversample_aggressiveness=best_setting[1]
    )
    train_dataloader_optimal = DataLoader(
        train_dataset,
        batch_size=8,
        drop_last=False,
        collate_fn=DetectionCollateFN(),
        sampler=optimal_sampler
    )

    val_dataloader = DataLoader(
        val_dataset,
        batch_size=8,
        shuffle=False,  # No need to shuffle validation data
        drop_last=False,
        collate_fn=DetectionCollateFN()
    )

    model = models.get(
        'yolo_nas_l',
        num_classes=len(CLASSES),
        pretrained_weights="coco"
    )

    trainer = Trainer(experiment_name=EXPERIMENT_NAME,
                        ckpt_root_dir=CHECKPOINT_DIR)

    loss_fn = PPYoloELoss(
        use_static_assigner=False,
        use_varifocal_loss=False,
        num_classes=len(CLASSES),
    )

    train_params = {
        "max_epochs": max_epochs,
        "loss": loss_fn,
        "criterion_params": None,
        "warmup_mode": "LinearEpochLRWarmup",
        "lr_warmup_epochs": 5,
        "lr_warmup_steps": 0,
        "warmup_initial_lr": 1e-6,
        "initial_lr": initial_lr,
        "metric_to_watch": '[email protected]',
        "batch_accumulate": 4,
    }

    train_params["loss_logging_items_names"] = ["Total Loss", "PPYoloE Loss"]

    trainer.train(
        model=model,
        training_params=train_params,
        train_loader=train_dataloader_optimal,
        valid_loader=val_dataloader
    )

if __name__ == "__main__":
    main()

Many thanks for your assistance !

Versions

PyTorch version: 1.13.1+cu117
Is debug build: False
CUDA used to build PyTorch: 11.7
ROCM used to build PyTorch: N/A

OS: Ubuntu 24.04.1 LTS (x86_64)
GCC version: (Ubuntu 13.3.0-6ubuntu2~24.04) 13.3.0
Clang version: Could not collect
CMake version: Could not collect
Libc version: glibc-2.39

Python version: 3.9.2rc1 (default, Jan 4 2025, 03:03:13) [GCC 13.3.0] (64-bit runtime)
Python platform: Linux-5.15.167.4-microsoft-standard-WSL2-x86_64-with-glibc2.39
Is CUDA available: True
CUDA runtime version: 12.0.140
CUDA_MODULE_LOADING set to: LAZY
GPU models and configuration: GPU 0: NVIDIA GeForce RTX 3070
Nvidia driver version: 566.14
cuDNN version: Could not collect
HIP runtime version: N/A
MIOpen runtime version: N/A
Is XNNPACK available: True

CPU:
Architecture: x86_64
CPU op-mode(s): 32-bit, 64-bit
Address sizes: 46 bits physical, 48 bits virtual
Byte Order: Little Endian
CPU(s): 24
On-line CPU(s) list: 0-23
Vendor ID: GenuineIntel
Model name: 12th Gen Intel(R) Core(TM) i9-12900
CPU family: 6
Model: 151
Thread(s) per core: 2
Core(s) per socket: 12
Socket(s): 1
Stepping: 2
BogoMIPS: 4838.40
Flags: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ss ht syscall nx pdpe1gb rdtscp lm constant_tsc rep_good nopl xtopology tsc_reliable nonstop_tsc cpuid pni pclmulqdq vmx ssse3 fma cx16 pcid sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes xsave avx f16c rdrand hypervisor lahf_lm abm 3dnowprefetch invpcid_single ssbd ibrs ibpb stibp ibrs_enhanced tpr_shadow vnmi ept vpid ept_ad fsgsbase tsc_adjust bmi1 avx2 smep bmi2 erms invpcid rdseed adx smap clflushopt clwb sha_ni xsaveopt xsavec xgetbv1 xsaves avx_vnni umip waitpkg gfni vaes vpclmulqdq rdpid movdiri movdir64b fsrm md_clear serialize flush_l1d arch_capabilities
Virtualization: VT-x
Hypervisor vendor: Microsoft
Virtualization type: full
L1d cache: 576 KiB (12 instances)
L1i cache: 384 KiB (12 instances)
L2 cache: 15 MiB (12 instances)
L3 cache: 30 MiB (1 instance)
Vulnerability Gather data sampling: Not affected
Vulnerability Itlb multihit: Not affected
Vulnerability L1tf: Not affected
Vulnerability Mds: Not affected
Vulnerability Meltdown: Not affected
Vulnerability Mmio stale data: Not affected
Vulnerability Reg file data sampling: Vulnerable: No microcode
Vulnerability Retbleed: Mitigation; Enhanced IBRS
Vulnerability Spec rstack overflow: Not affected
Vulnerability Spec store bypass: Mitigation; Speculative Store Bypass disabled via prctl and seccomp
Vulnerability Spectre v1: Mitigation; usercopy/swapgs barriers and __user pointer sanitization
Vulnerability Spectre v2: Mitigation; Enhanced / Automatic IBRS; IBPB conditional; RSB filling; PBRSB-eIBRS SW sequence; BHI SW loop, KVM SW loop
Vulnerability Srbds: Not affected
Vulnerability Tsx async abort: Not affected

Versions of relevant libraries:
[pip3] numpy==1.23.0
[pip3] nvidia-cublas-cu12==12.4.5.8
[pip3] nvidia-cuda-cupti-cu12==12.4.127
[pip3] nvidia-cuda-nvrtc-cu12==12.4.127
[pip3] nvidia-cuda-runtime-cu12==12.4.127
[pip3] nvidia-cudnn-cu12==9.1.0.70
[pip3] nvidia-cufft-cu12==11.2.1.3
[pip3] nvidia-curand-cu12==10.3.5.147
[pip3] nvidia-cusolver-cu12==11.6.1.9
[pip3] nvidia-cusparse-cu12==12.3.1.170
[pip3] nvidia-nccl-cu12==2.21.5
[pip3] nvidia-nvjitlink-cu12==12.4.127
[pip3] nvidia-nvtx-cu12==12.4.127
[pip3] onnx==1.15.0
[pip3] onnxruntime==1.15.0
[pip3] onnxsim==0.4.36
[pip3] torch==1.13.1+cu117
[pip3] torchaudio==0.13.1+cu117
[pip3] torchmetrics==0.8.0
[pip3] torchvision==0.14.1+cu117
[pip3] triton==3.1.0