ysautoml.network.zeroshot

Funtional Modules

• ysautoml.network.zeroshot.autoformer

• ysautoml.network.zeroshot.mobilenetv2

ysautoml.network.zeroshot.autoformer

ysautoml.network.zeroshot.autoformer.run_search_zeroshot

ysautoml.network.zeroshot.autoformer.run_search_zeroshot(**kwargs)

Run zero-shot architecture search using AutoFormer on ImageNet.

Parameters

• param_limits (float): Maximum parameter count limit (e.g., 6, 23, 54 for Tiny/Small/Base).

• min_param_limits (float): Minimum parameter count limit.

• cfg (str): Path or name of YAML search space (e.g., 'space-T.yaml').

• output_dir (str): Directory to save search results.

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

• population_num (int, default 10000): Number of architectures to sample.

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

• gp (bool, default True): Enable Gaussian process metric.

• relative_position (bool, default True): Enable relative positional embedding.

• change_qkv (bool, default True): Use QKV reparameterization.

• dist_eval (bool, default True): Enable distributed evaluation.

Returns

• None : Logs and search results (including best_arch.yaml and search.log) are saved under output_dir.

Examples

from ysautoml.network.zeroshot.autoformer import run_search_zeroshot

# Tiny Search
run_search_zeroshot(
    param_limits=6,
    min_param_limits=4,
    cfg="space-T.yaml",
    output_dir="./OUTPUT/search/AZ-NAS/Tiny"
)

---

ysautoml.network.zeroshot.autoformer.run_retrain_zeroshot

ysautoml.network.zeroshot.autoformer.run_retrain_zeroshot(**kwargs)

Retrain the best subnet architecture found by AZ-NAS using AutoFormer.

Parameters

• cfg (str): YAML configuration file path (absolute or relative).

• output_dir (str): Output directory for retraining results.

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

• epochs (int, default 500): Number of training epochs.

• warmup_epochs (int, default 20): Warm-up epochs.

• batch_size (int, default 256): Batch size per GPU.

• model_type (str, default 'AUTOFORMER'): Model type to train.

• mode (str, default 'retrain'): Training mode.

• relative_position (bool): Use relative positional embedding.

• change_qkv (bool): Use QKV reparameterization.

• gp (bool): Enable Gaussian process module.

• dist_eval (bool): Enable distributed evaluation.

• device (str, default '0,1,2,3,4,5,6,7'): Visible CUDA devices.

• nproc_per_node (int, default 8): Number of distributed processes per node.

• master_port (int, default 6666): Master port for distributed training.

Returns

• None : Trained model weights and logs are saved under output_dir.

Examples

from ysautoml.network.zeroshot.autoformer import run_retrain_zeroshot

run_retrain_zeroshot(
    cfg="./Tiny.yaml",
    output_dir="./OUTPUT/AZ-NAS/Tiny-bs256x8-use_subnet-500ep",
    epochs=500
)

---

ysautoml.network.zeroshot.mobilenetv2

ysautoml.network.zeroshot.mobilenetv2.run_search_zeroshot

ysautoml.network.zeroshot.mobilenetv2.run_search_zeroshot(**kwargs)

Run zero-shot evolution search for MobileNetV2 variants (AZ-NAS).

Parameters

• gpu (int, default 0): GPU ID to use.

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

• metric (str, default 'AZ_NAS'): Zero-shot score metric.

• population_size (int, default 1024): Population size for evolution search.

• evolution_max_iter (int, default 1e5): Maximum iterations for evolution search.

• resolution (int, default 224): Input image resolution.

• budget_flops (float, default 1e9): FLOPs constraint.

• max_layers (int, default 16): Maximum number of layers.

• batch_size (int, default 64): Batch size.

• data_path (str): Path to ImageNet dataset.

• num_classes (int, default 1000): Number of classes.

• search_space (str, default 'SearchSpace/search_space_IDW_fixfc.py'): Search space file.

Returns

• None – Outputs best architecture (best_structure.txt) and FLOPs/params summary in save_dir.

Examples

from ysautoml.network.zeroshot.mobilenetv2 import run_search_zeroshot

# Default 1G search
run_search_zeroshot(
    gpu=0,
    seed=123,
    metric="AZ_NAS",
    population_size=1024,
    evolution_max_iter=int(1e5),
    resolution=224,
    budget_flops=1e9,
    max_layers=16,
    batch_size=64,
    data_path="/dataset/ILSVRC2012/",
)

# Small (450M)
run_search_zeroshot(
    gpu=0,
    seed=123,
    budget_flops=450e6,
    max_layers=14,
)

# Medium (600M)
run_search_zeroshot(
    gpu=0,
    seed=123,
    budget_flops=600e6,
    max_layers=14,
)

# Large (1G)
run_search_zeroshot(
    gpu=0,
    seed=123,
    budget_flops=1000e6,
    max_layers=16,
)

---

ysautoml.network.zeroshot.mobilenetv2.run_retrain_zeroshot

ysautoml.network.zeroshot.mobilenetv2.run_retrain_zeroshot(**kwargs)

Retrain the searched MobileNetV2 architecture using AZ-NAS configuration (via Horovod).

Parameters

• gpu_devices (str, default '0,1,2,3,4,5,6,7'): Visible GPU device IDs.

• metric (str, default 'AZ_NAS'): Search metric name.

• population_size (int, default 1024): Population size used in search.

• evolution_max_iter (int, default 1e5): Number of evolution iterations used in search.

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

• num_workers (int, default 12): Number of data loader workers.

• init (str, default 'custom_kaiming'): Weight initialization method.

• epochs (int, default 150): Number of training epochs.

• resolution (int, default 224): Input image resolution.

• batch_size_per_gpu (int, default 64): Batch size per GPU.

• world_size (int, default 8): Number of distributed workers.

• data_path (str): Dataset root path.

• best_structure_path (str, optional): Path to best_structure.txt (absolute or relative to current working directory).

Returns

• Path – Directory of retraining outputs.

Examples

from ysautoml.network.zeroshot.mobilenetv2 import run_retrain_zeroshot

run_retrain_zeroshot(
    gpu_devices="0,1,2,3,4,5,6,7",
    epochs=150,
    init="custom_kaiming",
    best_structure_path="best_structure.txt"
)