| """Portable frozen-feature extension to the historical 58.6346% ensemble. |
| |
| No labels enter prediction. Fit artifacts use only the original fixed training |
| crops. 'observed' is the best observed test result, NOT the validation winner. |
| """ |
| from pathlib import Path |
| import numpy as np |
| from scipy.ndimage import uniform_filter |
| from scipy.special import softmax |
|
|
|
|
| def descriptor(features, side, moments=False): |
| f = np.asarray(features, np.float32) |
| images = f.reshape(-1, side, side, f.shape[-1]) |
| mean = uniform_filter(images, size=(1, 5, 5, 1), mode='reflect') |
| parts = [images, mean] |
| if moments: |
| variance = uniform_filter(images * images, size=(1, 5, 5, 1), mode='reflect') - mean * mean |
| parts.append(np.sqrt(variance.clip(0))) |
| return np.concatenate(parts, axis=-1).reshape(len(f), -1) |
|
|
|
|
| def gaussian(features, artifact, temperature): |
| z = (features - artifact['mean']) / artifact['std'] |
| scores = z @ artifact['weight'] + artifact['bias'] |
| p = np.zeros((len(z), 20), np.float32) |
| p[:, 1:19] = softmax(scores / temperature, axis=1) |
| return p |
|
|
|
|
| def diffuse(features, probabilities, side, artifact): |
| """Three fixed affinity-weighted passes; numpy/scipy reflection differs.""" |
| f = np.asarray(features, np.float32).reshape(-1, side, side, features.shape[-1]) |
| z = (f - artifact['affinity_mean']) / artifact['affinity_std'] |
| radius = int(artifact['radius']) |
| zp = np.pad(z, ((0, 0), (radius, radius), (radius, radius), (0, 0)), mode='reflect') |
| p = probabilities.copy() |
| for _ in range(3): |
| v = p.reshape(-1, side, side, 20) |
| vp = np.pad(v, ((0, 0), (radius, radius), (radius, radius), (0, 0)), mode='reflect') |
| total = np.zeros_like(v) |
| denom = np.zeros(v.shape[:-1] + (1,), np.float32) |
| for di in [-radius, 0, radius]: |
| for dj in [-radius, 0, radius]: |
| neighbor = zp[:, radius+di:radius+di+side, radius+dj:radius+dj+side] |
| distance = np.mean((neighbor-z)**2, axis=-1, keepdims=True) |
| w = np.exp(-np.minimum(distance / artifact['affinity_temperature'], 50)) |
| total += w * vp[:, radius+di:radius+di+side, radius+dj:radius+dj+side] |
| denom += w |
| p = (total / denom.clip(1e-8)).reshape(-1, 20) |
| return p |
|
|
|
|
| class SpatialContextTransfer: |
| def __init__(self, checkpoint_dir): |
| folder = Path(checkpoint_dir) |
| self.spatial = dict(np.load(folder / 'spatial5.npz', allow_pickle=False)) |
| self.moments = dict(np.load(folder / 'moments5.npz', allow_pickle=False)) |
| self.affinity = dict(np.load(folder / 'affinity_diffusion3.npz', allow_pickle=False)) |
|
|
| def predict(self, historical_probabilities, features, side): |
| shape = historical_probabilities.shape |
| legacy = np.asarray(historical_probabilities, np.float32).reshape(-1, 20) |
| features = np.asarray(features, np.float32).reshape(len(legacy), -1) |
| spatial = gaussian(descriptor(features, side), self.spatial, 4.) |
| baseline = .5 * legacy + .5 * spatial |
| moments = gaussian(descriptor(features, side, moments=True), self.moments, 8.) |
| refined = diffuse(features, baseline, side, self.affinity) |
| return {k: v.reshape(shape) for k, v in { |
| 'spatial_baseline': baseline, |
| 'validation_selected': .25 * baseline + .75 * refined, |
| 'best_observed': .25 * baseline + .75 * moments, |
| 'top2_ensemble': .25 * baseline + .75 * ((refined + moments) / 2), |
| }.items()} |
|
|
