feat(trackers): Implement training ReIDModel using Cross-entropy loss

[soumik12345]

Description

• Implement an identity dataloader that can be used to train a ReIDModel using cross-entropy loss.
• Make Market1501 dataset work with the identity dataloader.
• Implement PK-sampling: randomly sample P distinct IDs from the dataset and then randomly sample K unique images from each ID if possible, duplicate if that ID has < K images. (reference: section 3.1 of ICE: Inter-instance Contrastive Encoding for Unsupervised Person Re-identification)
• Refactor triplet training logic into TripletsTrainer, executed only when the dataset is a TripletsDataset.
• Implement fine-tuning with a cross-entropy loss with a projection head + (log?)softmax activation.
• Experiment with optimal hyperparameters and note results integrating into the updated DeepSORT implementation from feat(trackers): Implementation of Kalman Gating for DeepSORTTracker #17.

Type of change

Please delete options that are not relevant.

• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• This change requires a documentation update

Example code snippet

from torch.utils.data import DataLoader
from torch.utils.data.sampler import RandomSampler

from trackers import ReIDModel
from trackers.core.reid.dataset.market_1501 import get_market1501_triplets_dataset

train_dataset, val_dataset = get_market1501_triplets_dataset(
    data_dir="datasets/Market-1501-v15.09.15/bounding_box_train",
    dataset_type="identity",
    split_ratio=0.8,
)

train_loader = DataLoader(
    dataset=train_dataset,
    batch_size=32,
    sampler=RandomSampler(train_dataset),
    drop_last=True,
)

val_loader = DataLoader(
    dataset=val_dataset,
    batch_size=32,
    drop_last=True,
)
reid_model = ReIDModel.from_timm("resnetv2_50.a1h_in1k")

reid_model.train(
    train_loader=train_loader,
    epochs=5,
    validation_loader=val_loader,
    projection_dimension=len(train_dataset),
    freeze_backbone=True,
    learning_rate=5e-4,
    weight_decay=1e-2,
    log_to_wandb=True,
)