ysautoml.optimization.losssearch

Funtional Modules

• ysautoml.optimization.losssearch.train_losssearch

• ysautoml.optimization.losssearch.custom_loss

Run Loss Function Search (LFS) using a custom, learnable loss function integrated into the YSAutoML optimization system. This module supports two main functionalities:

1. Launching a complete end-to-end training pipeline (train_losssearch) where both model and loss function are jointly optimized.

2. Directly accessing and using the learnable criterion (custom_loss) in custom training or evaluation scripts.

ysautoml.optimization.losssearch.train_losssearch

ysautoml.optimization.losssearch.train_losssearch(**kwargs)

Launching a complete end-to-end training pipeline (train_losssearch) where both model and loss function are jointly optimized.

Parameters

• epochs (int, default 100): Total number of training epochs.

• lr_model (float, default 0.1): Learning rate for the model optimizer.

• lr_loss (float, default 0.0001): Learning rate for the optimizer that updates the custom loss parameters.

• momentum (float, default 0.9): Momentum factor for model and loss optimizers.

• weight_decay (float, default 0.0005): Weight decay coefficient for L2 regularization in model optimization.

• save_dir (str, default "./logs/losssearch"): Directory to save logs, checkpoints, and learned loss parameters.

• device (str, default "cuda:0"): Device identifier (e.g., "cuda:0", "cuda:1", or "cpu").

Returns

• None : All training logs, learned loss parameter values, and model checkpoints are saved under the specified save_dir.

  • Logs: ${save_dir}/training.log

  • TensorBoard events: ${save_dir}/events.out.tfevents.*

  • Checkpoints: ${save_dir}/model_epoch_XX.pth

  • Learned loss parameters: saved in model state dict as criterion.theta

Examples

from ysautoml.optimization.losssearch import train_losssearch

train_losssearch(
    epochs=50,
    lr_model=0.05,
    lr_loss=0.0005,
    momentum=0.9,
    weight_decay=0.0001,
    save_dir="./logs/losssearch_exp1",
    device="cuda:0"
)

---

ysautoml.optimization.losssearch.custom_loss

ysautoml.optimization.losssearch.custom_loss( )

Directly accessing and using the learnable criterion (custom_loss) in custom training or evaluation scripts.

Parameters

• None

Examples

from ysautoml.optimization.losssearch import custom_loss
import torch

criterion = custom_loss().cuda()
preds = torch.randn(8, 10).cuda()
targets = torch.randint(0, 10, (8,)).cuda()
loss = criterion(preds, targets)
loss.backward()