chrmc40
9fd6498ba6
- Tag-builder: port ~1,477 v1 characters into the v2 chip format so the character picker shows 1,486 characters (was 9); 1,390 carry rich per-chip descriptions restructured from each character's own bio. Adds scripts/port_v1_characters_to_v2.py. - Ideogram: rewrite caption to the real Ideogram-4 schema + broaden refusal detection; sampler updates. - Pose-studio, regional-detailer, attention-couple: assorted fixes. - Add validate_templates / test_regions / test_crop_windows scripts.
95 lines
4.0 KiB
Python
95 lines
4.0 KiB
Python
#!/usr/bin/env python3
|
|
"""Unit tests for the crop-window remap (core/crop_windows.py).
|
|
|
|
Locks the invariants of the most-regressed geometry in the codebase: output
|
|
dims, the nearest-vs-bilinear mask choice (the confirmed region-edge bleed fix),
|
|
full-canvas identity, and that distinct windows select distinct source regions
|
|
(the mask_win != ctrl_win moved-content mechanism). torch-only; run anywhere.
|
|
"""
|
|
|
|
from __future__ import annotations
|
|
|
|
import importlib.util
|
|
import os
|
|
import sys
|
|
|
|
import torch
|
|
|
|
_HERE = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
|
|
_spec = importlib.util.spec_from_file_location(
|
|
"crop_windows", os.path.join(_HERE, "core", "crop_windows.py"))
|
|
_cw = importlib.util.module_from_spec(_spec)
|
|
_spec.loader.exec_module(_cw)
|
|
crop = _cw.crop_hint_to_window
|
|
|
|
_failures = []
|
|
|
|
|
|
def check(name, cond):
|
|
print(f" {'PASS' if cond else 'FAIL'} {name}")
|
|
if not cond:
|
|
_failures.append(name)
|
|
|
|
|
|
def main() -> int:
|
|
H, W = 120, 80 # non-square canvas (portrait), the case aspect bugs hide in
|
|
|
|
# 1. Output dims are always (sh, sw) regardless of input size or window.
|
|
src = torch.rand(1, 1, 64, 48)
|
|
out = crop(src, (10, 20, 70, 100), H, W, 32, 24)
|
|
check("output dims == (sh, sw)", tuple(out.shape[-2:]) == (32, 24))
|
|
|
|
# 2. nearest-exact keeps a binary mask binary (no soft bleed band).
|
|
mask = torch.zeros(1, 1, H, W)
|
|
mask[..., 40:80, 20:60] = 1.0
|
|
out_n = crop(mask, (0, 0, W, H), H, W, 60, 40, mode="nearest-exact")
|
|
uniq = set(torch.unique(out_n).tolist())
|
|
check("nearest mask stays binary {0,1}", uniq.issubset({0.0, 1.0}))
|
|
|
|
# 3. bilinear DOES introduce an intermediate edge value — proves the mode
|
|
# actually changes behaviour (so the mask fix is meaningful, not a no-op).
|
|
# Upscale a hard edge so it must ramp (an exact-2x downsample with the edge
|
|
# on an even boundary would sample exactly 0/1 and hide the difference).
|
|
small = torch.zeros(1, 1, 8, 8)
|
|
small[..., 4:, :] = 1.0
|
|
up_b = crop(small, (0, 0, 8, 8), 8, 8, 32, 32, mode="bilinear")
|
|
up_n = crop(small, (0, 0, 8, 8), 8, 8, 32, 32, mode="nearest-exact")
|
|
check("bilinear softens the edge (mode matters)",
|
|
bool(((up_b > 0.01) & (up_b < 0.99)).any()))
|
|
check("nearest stays hard on the same upscale",
|
|
set(torch.unique(up_n).tolist()).issubset({0.0, 1.0}))
|
|
|
|
# 4. Full-canvas window is identity-ish: a uniform field stays uniform.
|
|
flat = torch.full((1, 1, H, W), 0.5)
|
|
out_f = crop(flat, (0, 0, W, H), H, W, H, W)
|
|
check("full-canvas uniform field preserved", torch.allclose(out_f, flat, atol=1e-5))
|
|
|
|
# 5. A window over an all-ones sub-region returns ~all-ones; a disjoint
|
|
# window returns ~all-zeros. This is what makes a tile sample the right
|
|
# region — and the basis of mask_win != ctrl_win moved-content selection.
|
|
field = torch.zeros(1, 1, H, W)
|
|
field[..., 0:40, 0:40] = 1.0 # top-left block is the "content"
|
|
inside = crop(field, (0, 0, 40, 40), H, W, 16, 16, mode="nearest-exact")
|
|
outside = crop(field, (40, 60, 80, 120), H, W, 16, 16, mode="nearest-exact")
|
|
check("window over content -> all ones", bool((inside > 0.5).all()))
|
|
check("disjoint window -> all zeros", bool((outside < 0.5).all()))
|
|
|
|
# 6. Distinct windows select DISTINCT regions of the same source (the
|
|
# mask_win vs ctrl_win divergence relies on this).
|
|
grad = torch.linspace(0, 1, W).view(1, 1, 1, W).expand(1, 1, H, W).contiguous()
|
|
left = crop(grad, (0, 0, 20, H), H, W, 8, 8)
|
|
right = crop(grad, (60, 0, 80, H), H, W, 8, 8)
|
|
check("distinct windows -> distinct content", float(left.mean()) < float(right.mean()))
|
|
|
|
# 7. Input already at canvas size skips the upscale but still crops+resizes.
|
|
at_canvas = torch.rand(1, 1, H, W)
|
|
out_ac = crop(at_canvas, (10, 10, 50, 50), H, W, 20, 20)
|
|
check("pre-sized input crops to (sh,sw)", tuple(out_ac.shape[-2:]) == (20, 20))
|
|
|
|
print(f"\n{'ALL PASS' if not _failures else str(len(_failures)) + ' FAILURE(S)'}")
|
|
return 1 if _failures else 0
|
|
|
|
|
|
if __name__ == "__main__":
|
|
sys.exit(main())
|