"""Next Shift: few-shot adaptation of a digit recogniser to a new visual environment. Pieces shared by the notebook, inference.py and (ported line by line) web/core.js: * conv features: nlrt layer lists exported from PyTorch (BatchNorm folded); `conv_target` uses BatchNorm statistics re-estimated on *unlabeled* target images (AdaBN), `conv_source` the source statistics. * finetune_top: full-batch gradient descent on the top MLP (512 -> 64 -> 10) from the pretrained weights. * train_linear: logistic-regression probe from zero weights on frozen 64-d source features. All arithmetic is float64 and deterministic. """ import numpy as np from nlrt import forward def conv_features(layers, images): """images: (N, 16, 16) in [0, 1] -> (N, 512).""" return forward(layers, np.asarray(images, dtype=np.float64)[:, None, :, :]) def mlp_logits(p, F): h = np.maximum(F @ p["w1"].T + p["b1"], 0) return h @ p["w2"].T + p["b2"], h def softmax(z): z = z - z.max(axis=1, keepdims=True) e = np.exp(z) return e / e.sum(axis=1, keepdims=True) def finetune_top(p0, F, y, steps, lr): p = {k: np.array(v, dtype=np.float64) for k, v in p0.items()} n = len(y) for _ in range(steps): z, h = mlp_logits(p, F) P = softmax(z) P[np.arange(n), y] -= 1.0 P /= n gw2 = P.T @ h gb2 = P.sum(axis=0) dh = (P @ p["w2"]) * (h > 0) gw1 = dh.T @ F gb1 = dh.sum(axis=0) p["w1"] -= lr * gw1 p["b1"] -= lr * gb1 p["w2"] -= lr * gw2 p["b2"] -= lr * gb2 return p def train_linear(Z, y, steps, lr, n_classes=10): n, d = Z.shape W = np.zeros((n_classes, d)) b = np.zeros(n_classes) for _ in range(steps): P = softmax(Z @ W.T + b) P[np.arange(n), y] -= 1.0 P /= n W -= lr * (P.T @ Z) b -= lr * P.sum(axis=0) return W, b