Unable to Achieve 90% Accuracy with PerBatch Method in CORDS_SL_CIFAR10_Custom_Train.ipynb

[football-prince]

Dear CORDS team,

I am trying to replicate the results from your paper "GRAD-MATCH: Gradient Matching based Data Subset Selection for Efficient Deep Model Training" using the provided code in cords/examples/SL/image_classification/python_notebooks/CORDS_SL_CIFAR10_Custom_Train.ipynb. However, I am encountering a discrepancy in performance.

I have followed the instructions in the notebook and implemented a custom dataloader for the PerBatch subset selection method, as the original code did not provide it. After training for 300 epochs, I am only able to achieve 82% accuracy, whereas the paper reports 90% accuracy. I would appreciate any help in identifying potential issues or misconfigurations that could be causing this gap.

Here are the relevant parts of the code and parameter settings I used:

from cords.utils.data.dataloader.SL.adaptive import GradMatchDataLoader, CRAIGDataLoader, RandomDataLoader
from dotmap import DotMap

selection_strategy = 'GradMatch'

dss_args = dict(
    model=model,
    loss=criterion_nored,
    eta=0.01,                  # Learning rate
    num_classes=10,            # Number of classes in the dataset, e.g., CIFAR-10
    num_epochs=300,            # Total number of epochs
    device='cuda',             # Device set to GPU
    fraction=0.1,              # Fraction of the subset size
    select_every=20,           # Frequency of subset selection every few epochs
    kappa=0,                   # Kappa value
    linear_layer=True,         # Whether to use a linear layer
    selection_type='PerBatch', # GradMatch selection type like 'PerClass', 'PerBatch'
    greedy='Stochastic',       # Type of greedy selection
    valid=False,               # Whether to use validation set for subset selection
    v1=True,                   # GradMatch parameter v1
    lam=0.5,                   # Regularization parameter lam
    eps=1e-5                   # Convergence tolerance eps
)

dss_args = DotMap(dss_args)

dataloader = GradMatchDataLoader(
    trainloader,            # Training data loader
    valloader,              # Validation data loader
    dss_args,               # Parameter dictionary
    logger,                 # Logger
    batch_size=20,          # Batch size for the data loader
    shuffle=True,           # Whether to shuffle the data
    pin_memory=False        # Whether to use pin_memory for faster data loading
)

Specific Issue:

I am using the PerBatch subset selection method, but the accuracy still cannot exceed 82%, while the paper reports 90%.
Are there any key configurations or settings in dss_args or GradMatchDataLoader initialization that I may have missed?
I have checked all steps but still cannot achieve the expected results. Any guidance or recommendations for adjustments would be greatly appreciated.

Thank you for your time and help!

Best regards,