Next-Shift / README.md
Lamapi's picture
Update README.md
5519776 verified
|
Raw History Blame Contribute Delete
3.08 kB
metadata
license: cc-by-4.0
library_name: pytorch
tags:
  - domain-adaptation
  - few-shot
  - image-classification
  - next-lab

shift

Next Shift

Next Shift adapts a digit recogniser trained on clean scans to a new visual environment (grey, low-contrast, textured, rotated, blurred captures) using only a handful of labeled target images. It first re-estimates BatchNorm statistics on unlabeled target images (AdaBN), then fine-tunes the top MLP on the labels. The target environment is a synthetic transform of real UCI digits.

Space: https://huggingface.co/spaces/thelamapi/Next-Shift — label a few target images; the top layers fine-tune in your browser and predictions update.

Results: target test accuracy (297 images, 5 seeds, mean ± std)

Method k=0 k=1 k=5 k=10 (labels per class)
Next Shift (AdaBN + fine-tune top) 0.394 ± 0.000 0.489 ± 0.033 0.588 ± 0.021 0.661 ± 0.010
Ablation: fine-tune top, no AdaBN 0.401 ± 0.000 0.459 ± 0.042 0.589 ± 0.014 0.673 ± 0.009
Frozen source features + linear head 0.401 ± 0.000 0.358 ± 0.031 0.558 ± 0.019 0.613 ± 0.022
Random init, same labels and steps 0.099 ± 0.023 0.211 ± 0.022 0.467 ± 0.026 0.636 ± 0.050

Source-only accuracy on the source validation split: 0.973; on the target test split: 0.401. For each k and seed, every method sees the same labeled images. Steps and learning rates were chosen per method on a separate 100-image target development split (seeds 100–102), never on the test split.

Accuracy vs labels

Method

  • Encoder: 2 conv blocks (16, 32 channels, BatchNorm, ReLU, max-pool) → 512 features → fc 64 → 10 classes; pre-trained on 900 source images.
  • Next Shift: AdaBN on 350 unlabeled target pool images, then 50 full-batch GD steps (lr 0.05) on the top MLP.
  • Frozen representation: linear head on fixed 64-d source features (200 steps, lr 0.1).
  • Random init: the whole network from scratch with Adam (100 steps, lr 0.001).
  • Ablation without AdaBN: same fine-tuning on source-normalised features.
  • Browser: conv stacks run through nlrt.js and the fine-tuning is a line-by-line port; Node parity max abs difference 5.412337245047638e-16.

Usage

from inference import load, adapt, predict
spec = load()                               # NumPy only
top = adapt(spec, target_images, labels)    # images: (N,16,16) in [0,1]
probs = predict(spec, new_images, top)

source_model.safetensors holds the PyTorch source weights (architecture in the notebook).

Limitations

  • The domain shift is synthetic and the images are tiny; real shifts (sensors, lighting, populations) can be harder.
  • AdaBN needs a batch of unlabeled target images; the released conv_target was estimated on the target pool.
  • 96 of 297 target test images are still wrong for Next Shift at k=10 (seed 0); see evaluation.json.
  • Trained on NVIDIA L4; profile full.