Training

Train, sweep, or search FPS configs from selected H5 views.

Training YAML is sectioned and explicit: choose the bank, domains, role-specific views, optimizer, schedule, normalization, losses, and evaluation settings.

CLI

Train or sweep

Train

PYTHONPATH=src python -m fps_uda.cli train \
  --config configs/training/office31/amazon_to_webcam/vit.yaml \
  --out runs/office31/amazon_to_webcam/vit

Sweep

PYTHONPATH=src python -m fps_uda.cli sweep \
  --config configs/training/office31/amazon_to_webcam/vit.yaml \
  --alpha-grid 0.6,0.8,1.0 \
  --beta-grid 0.4,0.6 \
  --out runs/sweeps/office31_aw_vit

Direct CLI

Train without a config

For quick experiments, specify the feature bank, domains, and view roles directly from the command line.

PYTHONPATH=src python -m fps_uda.cli train \
  --feature-bank fps_h5cache/banks/office31_vit.h5 \
  --source-domain amazon \
  --target-domain webcam \
  --src-view pad_resize256_input224_center_orig_clean \
  --entropy-view pad_resize256_input224_center_orig_clean \
  --cr-view1 pad_resize256_input224_center_orig_pool_a \
  --cr-view2 pad_resize256_input224_center_orig_pool_b \
  --eval-view pad_resize256_input224_center_orig_clean \
  --feature-transform none \
  --device cuda:0 \
  --base-lr 0.000375 \
  --iter-num 1000 \
  --out runs/debug/cli_only

Configuration

Sectioned YAML

`io` and `views`

Select the H5 bank, source/target domains, feature transform, and role-specific view keys. Multiple views can be stacked or mean-combined.

`optimization` and `schedule`

Configure SGD or AdamW, base_lr, LR schedule, update count, dynamic alpha/beta, and sample ratios.

`losses`

Enable LSE, LCE, LCR, LDelta, sparse density weights, and pseudo-margin weights. margin_start_step makes the margin warmup explicit.

io:
  feature_bank: fps_h5cache/banks/office_home_vit.h5
  source_domain: "Real World"
  target_domain: Clipart
  feature_transform: none
  num_classes: 65

views:
  src: {key: pad_resize256_input224_center_orig_clean, combine: stack}
  entropy: {key: pad_resize256_input224_center_orig_clean, combine: mean}
  cr:
    view1: {key: pad_resize256_input224_center_orig_pool_a, combine: stack}
    view2: {key: pad_resize256_input224_center_orig_pool_b, combine: stack}
  eval: {key: pad_resize256_input224_center_orig_clean, combine: mean}

Autosearch

Search reusable training configs

The search helper selects candidate views, evaluates LR candidates, searches alpha/beta, lambda_lcr, and margin settings, then writes sectioned YAML files.

PYTHONPATH=src BACKBONES="resnet vit siglip2" DEVICE=cuda:0 \
  bash scripts/search_training_configs.sh

Python API

Train from NumPy, Torch, or selected H5 views

from fps_uda import FeatureSet, FPSConfig, train_fps

features = FeatureSet(
    src_features=src_x,
    src_labels=src_y,
    entropy_features=tgt_x_eval,
    entropy_labels=tgt_y_eval,
    cr_features_1=tgt_x_view1,
    cr_features_2=tgt_x_view2,
    cr_labels=tgt_y_eval,
    eval_features=tgt_x_eval,
    eval_labels=tgt_y_eval,
)

config = FPSConfig(
    num_classes=31,
    feature_dim=src_x.shape[1],
    device="cuda:0",
    base_lr=0.000375,
    iter_num=3000,
)

result = train_fps(features, config, output_dir="runs/python_api")