ysautoml.optimization.mtl

Funtional Modules

• ysautoml.optimization.mtl.examples.nyusp.train_mtl_nyusp

ysautoml.optimization.mtl.examples.nyusp.train_mtl_nyusp

ysautoml.optimization.mtl.examples.nyusp.train_mtl_nyusp(**kwargs)

Run Multi-Task Learning (MTL) training on the NYUv2 dataset using the LibMTL framework integrated within YSAutoML. This function reproduces the original run.sh script from Multi-Task-Learning/examples/nyusp, allowing you to launch MTL experiments (e.g., GeMTL, GradNorm, UW, DWA, etc.) directly through the YSAutoML API.

Parameters

• gpu_id (int, default 0): Index of the GPU to use for training (e.g., 0, 1, ...).

• seed (int, default 0): Random seed for reproducibility of task sampling, optimizer initialization, and dataset splits.

• weighting (str, default "GeMTL"): Multi-task weighting strategy. Available options include: Arithmetic, GLS, UW, DWA, RLW, GradNorm, SI, IMTL_L, LSBwD, LSBwoD, AMTL, GeMTL.

• arch (str, default "HPS"): Model architecture type for task-specific decoders and shared backbone. Supported architectures include: HPS, Cross_stitch, MTAN, CGC, PLE, MMoE, DSelect_k, DIY, LTB.

• dataset_path (str, default "/dataset/nyuv2"): Root path of the NYUv2 dataset containing images and labels.

• scheduler (str, default "step"): Learning rate scheduling policy for the optimizer. Common options: "step", "cos", "exp".

• mode (str, default "train"): Operation mode for the run.

  • "train": Train model with train/val/test splits

  • "test": Evaluate a pretrained model only

• save_dir (str, default "./logs/nyusp_exp1"): Directory where experiment logs and checkpoints will be saved.

Returns

• None : All logs, TensorBoard events, and model checkpoints are saved automatically in save_dir.

  • Training logs: ${save_dir}/train.log

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

  • Model checkpoints: ${save_dir}/checkpoints/epoch_xx.pth

Examples

from ysautoml.optimization.mtl.examples.nyusp import train_mtl_nyusp

train_mtl_nyusp(
    gpu_id=0,
    seed=42,
    weighting="GeMTL",
    arch="HPS",
    dataset_path="/dataset/nyuv2",
    scheduler="step",
    mode="train",
    save_dir="./logs/nyusp_exp1"
)