ysautoml.network.oneshot

Funtional Modules

• ysautoml.network.oneshot.train_dynas

• ysautoml.network.oneshot.quantize_dynas

ysautoml.network.oneshot.train_dynas

ysautoml.network.oneshot.train_dynas(**kwargs)

Run DYNAS (Subnet-Aware Dynamic Supernet Training for Neural Architecture Search) or SPOS baseline training using the original train_spos.py script, integrated into the YSAutoML API system.

Parameters

• log_dir (str, default logs/spos_dynamic): Directory path to store TensorBoard and log outputs.

• file_name (str, default spos_dynamic): Base name for experiment log and validation pickle files.

• seed (int, default 0): Random seed for reproducibility.

• epochs (int, default 250): Total number of training epochs.

• lr (float, default 0.025): Initial learning rate for SGD optimizer.

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

• wd (float, default 0.0005): Weight decay coefficient for L2 regularization.

• nesterov (bool, default True): Whether to enable Nesterov momentum in SGD.

• train_batch_size (int, default 64): Batch size used for training dataset.

• val_batch_size (int, default 256): Batch size used for validation dataset.

• method (str, default dynas): Training mode. Use baseline for SPOS baseline, or dynas for dynamic architecture training with adaptive learning rate scheduling.

• max_coeff (float, default 4.0): Maximum coefficient γ_max controlling adaptive learning rate scaling for DYNAS.

Returns

• None : All training logs, pickled validation results, and TensorBoard summaries are saved under log_dir.

  • Logs: ${log_dir}/spos_dynamic.log

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

  • Validation accuracy pickle: exps/NAS-Bench-201-algos/valid_accs/{file_name}.pkl

  • Kendall correlation report (CIFAR10/100, ImageNet)

Examples

from ysautoml.network.oneshot import train_dynas

# Run DYNAS (Dynamic NAS)
train_dynas(
    log_dir="./logs/dynas_exp1",
    file_name="dynas_c10",
    seed=42,
    epochs=5,
    method="dynas",          # DYNAS adaptive LR mode
    max_coeff=4.0
)

# Run SPOS Baseline
train_dynas(
    log_dir="./logs/spos_exp1",
    file_name="spos_baseline",
    seed=0,
    epochs=5,
    method="baseline"        # vanilla SPOS
)

---

ysautoml.network.oneshot.quantize_dynas

ysautoml.network.oneshot.quantize_dynas(**kwargs)

Run post-training quantization for a searched DYNAS/SPOS subnet using the original comparison.py script, integrated into the YSAutoML API system. This function loads the searched architecture from a text file, restores a trained checkpoint, applies uniform k-bit quantization, and saves the quantized checkpoint automatically.

Parameters

• arch_file (str, default best_structure.txt): Path to the text file containing the searched architecture configuration. The file must contain comma-separated integers (e.g., 2, 2, 6, 0, ...).

• ckpt (str, default epoch-241.pt): Path to the trained floating-point checkpoint file to be quantized.

• quant_bits (int, default 8): Bit-width for uniform weight quantization (e.g., 8-bit, 6-bit).

• log_name (int, default quantize_output): Tag name for quantization process logging.

Returns

• None :

  All quantization results are saved automatically as a new checkpoint file.

  • Quantized checkpoint: ${ckpt_dir}/${ckpt_name}_quant{quant_bits}.pt

  • Console output:

    • Loaded architecture tuple

    • Quantization status

    • Final saved checkpoint path

Examples

from ysautoml.network.oneshot import quantize_dynas

# Run 8-bit post-training quantization
quantize_dynas(
    arch_file="best_architecture.txt",
    ckpt="epoch-241.pt",
    quant_bits=8
)

# Run lower-bit quantization (e.g., 6-bit)
quantize_dynas(
    arch_file="./outputs/best_architecture.txt",
    ckpt="./checkpoints/epoch-250.pt",
    quant_bits=6,
    log_name="quant6_test"
)