ysautoml.network.fewshot

Funtional Modules

• ysautoml.network.fewshot.mobilenet

ysautoml.network.fewshot.mobilenet

ysautoml.network.fewshot.mobilenet.train_supernet

ysautoml.network.fewshot.mobilenet.train_supernet(**kwargs)

Train a shared-weight SuperNet for few-shot MobileNet search using distributed or single-GPU execution.

Parameters

• tag (str): Experiment name for log and checkpoint files.

• seed (int, default -1): Random seed for reproducibility.

• thresholds (tuple, default (38, 40)): FLOPs-based sampling thresholds.

• data_path (str, default "/dataset/ILSVRC2012"): Dataset root directory.

• save_path (str, default "./SuperNet"): Directory to store logs and checkpoints. Automatically resolved relative to the caller’s working directory.

• search_space (str, default "proxyless"): Search space type (proxyless, spos, etc.).

• num_gpus (int, default 2): Number of GPUs to use for training.

• workers (int, default 4): Number of dataloader workers.

• max_epoch (int, default 120): Total training epochs.

• train_batch_size (int, default 1024): Training batch size.

• test_batch_size (int, default 256): Validation batch size.

• learning_rate (float, default 0.12): Initial learning rate.

• momentum (float, default 0.9): SGD momentum factor.

• weight_decay (float, default 4e-5): L2 regularization weight.

• lr_schedule_type (str, default "cosine"): Learning rate scheduling strategy.

• warmup (bool, default False): Enable LR warmup.

• log_to_tb (bool, default True): Write TensorBoard logs (./tb_logs/supernet/{tag}).

Returns

• None : Creates logs/{tag}-seed-{seed}.txt and checkpoint/{tag}-seed-{seed}.pt under the specified save_path.

Examples

from ysautoml.network.fewshot.mobilenet import train_supernet

train_supernet(
    tag="exp1",
    seed=42,
    thresholds=(38, 40),
    num_gpus=2,
    max_epoch=2,
    save_path="./save_dir"  # resolved relative to current file
)

---

ysautoml.network.fewshot.mobilenet.search_supernet

ysautoml.network.fewshot.mobilenet.search_supernet(**kwargs)

Run evolutionary architecture search on a pre-trained MobileNet SuperNet checkpoint.

Parameters

• ckpt (str): Checkpoint name to load (e.g., "baseline0-seed-0").

• seed (int, default 123): Random seed.

• gpu (int, default 0): GPU index.

• data_path (str, default "/dataset/ILSVRC2012"): Dataset root path.

• save_path (str, default "./Search"): Directory for search results. Automatically resolved relative to the caller’s working directory.

• search_space (str, default "proxyless"): Search space type.

• workers (int, default 4): Number of dataloader workers.

• run_calib (bool, default False): Enable BatchNorm calibration before evaluation.

Returns

• None : Evolutionary search results are logged and stored under save_path:

  • logs/ER-{ckpt}-runseed-{seed}.txt

  • checkpoint/ER-{ckpt}-runseed-{seed}.pt

Examples

from ysautoml.network.fewshot.mobilenet import search_supernet

search_supernet(
    ckpt="val6-2-seed-0",
    seed=0,
    gpu=0,
    run_calib=True,
    save_path="./save_dir"  # current working directory based
)