"""KNP input files for the one-time fresh-test evaluation (600 images, test 0700–0999 per class).

D: learned features (K 512 model from make_knp_features.py, same gain), top-256 per image; training part =
   the first 18000 rows of kmeans_top256_validation.bin (9000 images + mirror copies).
E: v2 sparse HOG (cats-dogs/features.py, descriptor(sparse=True)); training part = the first 4000 rows of
   features-v2/sparse_hog_validation.bin (v2 training set).

usage: ~/anvenv/bin/python prepare_fresh_knp.py
"""
import shutil
import sys

import numpy as np

sys.path.insert(0, "/mnt/c/knp/objects")
sys.path.insert(0, "/mnt/c/knp/cats-dogs")
import kmeans_features as kf  # noqa: E402
from features import descriptor  # noqa: E402

root = "/home/int/run/cats-dogs"
fresh = np.load(f"{root}/data-v3/fresh_test.npz")
labels = "\n".join(map(str, fresh["labels"])) + "\n"


def write(path, train_bin, train_target, n_train, width, test):
    x = np.fromfile(train_bin, np.uint8).reshape(-1, width)[:n_train]
    y = open(train_target).read().split()[:n_train]
    np.concatenate([x, test]).tofile(path + ".bin")
    open(path + ".target", "w").write("\n".join(y) + "\n" + labels)
    print(path, n_train, "+", len(test))


# D
kf.CONTRAST_DIVIDE = False
m = dict(np.load(f"{root}/features-v3/kmeans_model.npz"))
f = kf.encode(fresh["rgb"], m, alpha=0.25)
q = np.clip(np.rint(f / m["gain"] * 255), 0, 255).astype(np.uint8)
keep = np.argpartition(q, -256, axis=1)[:, -256:]
s = np.zeros_like(q)
np.put_along_axis(s, keep, np.take_along_axis(q, keep, 1), 1)
write(f"{root}/features-v3/kmeans_top256_fresh", f"{root}/features-v3/kmeans_top256_validation.bin",
      f"{root}/features-v3/kmeans_top256_validation.target", 18000, 2048, s)

# E
h = descriptor(fresh["rgb"].reshape(-1, 3, 1024), sparse=True)
write(f"{root}/features-v2/sparse_hog_fresh", f"{root}/features-v2/sparse_hog_validation.bin",
      f"{root}/features-v2/sparse_hog_validation.target", 4000, 2048, h)