detection-heads / target_cache.py
phanerozoic's picture
update repository
74e3c01
"""Option 4: Precomputed FCOS target cache.
The FCOS target assignment for each image is deterministic given
(spatial_features, boxes, labels) and a fixed level layout (strides + sizes).
Our 5-scale layout is fixed, so we can precompute targets once per image and
cache them alongside the spatial features in each shard. Training then loads
targets directly instead of recomputing on every forward pass.
Specific to our backbone configuration: 640px input, 40x40 stride-16 spatial
output, 5 prediction levels at strides [8, 16, 32, 64, 128] with FCOS standard
size ranges. Any architecture change to scale count, strides, or size ranges
invalidates the cache.
Includes a thorough self-test: builds a synthetic shard via the mock backbone,
precomputes targets, runs the same data through the original
assign_targets_batched, and asserts bitwise equivalence of all target tensors.
"""
import json
import os
import sys
import time
import torch
SCRIPT_DIR = os.path.dirname(os.path.abspath(__file__))
sys.path.insert(0, SCRIPT_DIR)
# Fixed level layout for the 5-scale split-tower head this cache targets.
STRIDES = [8, 16, 32, 64, 128]
SIZE_RANGES = [(-1, 32), (32, 64), (64, 128), (128, 256), (256, float("inf"))]
RESOLUTION = 640
H = RESOLUTION // 16 # 40 — base patch grid
def make_locations(feature_sizes, strides, device):
locs = []
for (h, w), s in zip(feature_sizes, strides):
ys = (torch.arange(h, device=device, dtype=torch.float32) + 0.5) * s
xs = (torch.arange(w, device=device, dtype=torch.float32) + 0.5) * s
gy, gx = torch.meshgrid(ys, xs, indexing="ij")
locs.append(torch.stack([gx.flatten(), gy.flatten()], -1))
return locs
def precompute_targets_for_image(boxes, labels, locs, level_ranges, device):
"""Compute FCOS targets for one image. Mirrors assign_targets_batched but
operates on a single image (B=1 implicit) so we can store per-image targets.
boxes: (M, 4) in (x1, y1, x2, y2)
labels: (M,) int
locs: concatenated (N_total, 2) of (cx, cy)
level_ranges: list of (start, end, stride, size_lo, size_hi)
Returns:
tgt_cls: (N_total,) class index or -1
tgt_reg: (N_total, 4) ltrb distances (only valid where tgt_cls >= 0)
tgt_ctr: (N_total,) centerness (only valid where tgt_cls >= 0)
"""
N = locs.shape[0]
tgt_cls = torch.full((N,), -1, dtype=torch.long, device=device)
tgt_reg = torch.zeros(N, 4, device=device)
tgt_ctr = torch.zeros(N, device=device)
if boxes.numel() == 0:
return tgt_cls, tgt_reg, tgt_ctr
areas = (boxes[:, 2] - boxes[:, 0]) * (boxes[:, 3] - boxes[:, 1])
M = boxes.shape[0]
for lo, hi, stride, slo, shi in level_ranges:
n = hi - lo
loc = locs[lo:hi]
l = loc[:, None, 0] - boxes[None, :, 0]
t = loc[:, None, 1] - boxes[None, :, 1]
r = boxes[None, :, 2] - loc[:, None, 0]
b = boxes[None, :, 3] - loc[:, None, 1]
ltrb = torch.stack([l, t, r, b], dim=-1)
in_box = ltrb.min(dim=-1).values > 0
cx = (boxes[:, 0] + boxes[:, 2]) / 2
cy = (boxes[:, 1] + boxes[:, 3]) / 2
rad = stride * 1.5
in_center = ((loc[:, None, 0] >= cx - rad) & (loc[:, None, 0] <= cx + rad) &
(loc[:, None, 1] >= cy - rad) & (loc[:, None, 1] <= cy + rad))
max_d = ltrb.max(dim=-1).values
in_level = (max_d >= slo) & (max_d <= shi)
pos = in_box & in_center & in_level
a = areas[None, :].expand_as(pos).clone()
a[~pos] = float("inf")
matched = a.argmin(dim=-1)
is_pos = a.gather(1, matched[:, None]).squeeze(1) < float("inf")
if is_pos.any():
tgt_cls[lo:hi][is_pos] = labels[matched[is_pos]]
arange_n = torch.arange(n, device=device)[is_pos]
ltrb_pos = ltrb[arange_n, matched[is_pos]]
tgt_reg[lo:hi][is_pos] = ltrb_pos
lp, tp, rp, bp = ltrb_pos.unbind(-1)
tgt_ctr[lo:hi][is_pos] = torch.sqrt(
(torch.minimum(lp, rp) / torch.maximum(lp, rp).clamp(min=1e-6)) *
(torch.minimum(tp, bp) / torch.maximum(tp, bp).clamp(min=1e-6)))
return tgt_cls, tgt_reg, tgt_ctr
def precompute_shard_targets(shard, device="cuda"):
"""Add precomputed (tgt_cls, tgt_reg, tgt_ctr) to each entry in a shard.
Modifies shard in place. Each entry gains three keys:
tgt_cls: (N_total,) int8 — stored compactly; -1 for negatives.
tgt_reg: (N_total, 4) float16 — only meaningful where tgt_cls >= 0.
tgt_ctr: (N_total,) float16 — only meaningful where tgt_cls >= 0.
"""
feat_sizes = [(H * 2, H * 2), (H, H), (H // 2, H // 2),
(H // 4, H // 4), (H // 8, H // 8)]
locs_per_level = make_locations(feat_sizes, STRIDES, torch.device(device))
all_locs = torch.cat(locs_per_level, 0)
n_per_level = [loc.shape[0] for loc in locs_per_level]
level_ranges = []
cumsum = 0
for i, n in enumerate(n_per_level):
lo, hi = SIZE_RANGES[i]
level_ranges.append((cumsum, cumsum + n, STRIDES[i], lo, hi))
cumsum += n
for entry in shard:
boxes = entry["boxes"].to(device).float()
labels = entry["labels"].to(device).long()
tcls, treg, tctr = precompute_targets_for_image(
boxes, labels, all_locs, level_ranges, device)
# Store compactly: int16 for cls (saves 4×), fp16 for reg/ctr
entry["tgt_cls"] = tcls.to(torch.int16).cpu()
entry["tgt_reg"] = treg.to(torch.float16).cpu()
entry["tgt_ctr"] = tctr.to(torch.float16).cpu()
return shard
def precompute_loss_with_cache(cls_per, reg_per, ctr_per, batch_tgt_cls, batch_tgt_reg, batch_tgt_ctr,
num_classes=80):
"""Compute FCOS loss using PRECOMPUTED targets — replaces the assignment
step with cache lookup. The classification, regression, and centerness
losses themselves are unchanged from the in-line version.
cls_per/reg_per/ctr_per: lists of per-level prediction tensors (B, C, H, W)
batch_tgt_cls/reg/ctr: per-batch precomputed targets (B, N_total) and (B, N_total, 4)
"""
import torch.nn.functional as F
B = cls_per[0].shape[0]
device = cls_per[0].device
flat_cls = torch.cat([c.permute(0, 2, 3, 1).reshape(B, -1, num_classes) for c in cls_per], 1)
flat_reg = torch.cat([r.permute(0, 2, 3, 1).reshape(B, -1, 4) for r in reg_per], 1)
flat_ctr = torch.cat([c.permute(0, 2, 3, 1).reshape(B, -1) for c in ctr_per], 1)
pos = batch_tgt_cls >= 0
npos = max(pos.sum().item(), 1)
oh = torch.zeros_like(flat_cls)
pi = pos.nonzero(as_tuple=True)
oh[pi[0], pi[1], batch_tgt_cls[pos].long()] = 1.0
# Focal loss (matches existing implementation)
p = torch.sigmoid(flat_cls)
ce = F.binary_cross_entropy_with_logits(flat_cls, oh, reduction="none")
pt = p * oh + (1 - p) * (1 - oh)
at = 0.25 * oh + 0.75 * (1 - oh)
loss_cls = (at * (1 - pt) ** 2 * ce).sum() / npos
if pos.any():
# Decode locations on the fly (still cheaper than full assignment)
feat_sizes = [(H * 2, H * 2), (H, H), (H // 2, H // 2),
(H // 4, H // 4), (H // 8, H // 8)]
all_locs = torch.cat(make_locations(feat_sizes, STRIDES, device), 0)
pl = all_locs[None].expand(B, -1, -1)[pos]
pp = flat_reg[pos]
tp = batch_tgt_reg[pos].float()
pb = torch.stack([pl[:, 0] - pp[:, 0], pl[:, 1] - pp[:, 1],
pl[:, 0] + pp[:, 2], pl[:, 1] + pp[:, 3]], -1)
tb = torch.stack([pl[:, 0] - tp[:, 0], pl[:, 1] - tp[:, 1],
pl[:, 0] + tp[:, 2], pl[:, 1] + tp[:, 3]], -1)
from torchvision.ops import generalized_box_iou
giou = generalized_box_iou(pb, tb)
loss_reg = (1 - giou.diagonal()).sum() / npos
loss_ctr = F.binary_cross_entropy_with_logits(
flat_ctr[pos], batch_tgt_ctr[pos].float(), reduction="sum") / npos
else:
loss_reg = torch.tensor(0.0, device=device)
loss_ctr = torch.tensor(0.0, device=device)
return loss_cls + loss_reg + loss_ctr
# ============================================================
# Self-test using the mock backbone
# ============================================================
if __name__ == "__main__":
from mock_eupe_backbone import make_mock_features, make_mock_boxes
device = "cuda" if torch.cuda.is_available() else "cpu"
print(f"Self-test on {device}")
print("=" * 60)
B = 4
boxes_list, labels_list = make_mock_boxes(B=B, n_boxes_per_image=8, device=device, seed=0)
# Build a synthetic shard
print("\n1. Building synthetic shard via mock features + boxes...")
shard = []
for i in range(B):
feats = make_mock_features(B=1, device=device, seed=i)[0].half()
shard.append({
"img_id": i,
"spatial": feats,
"boxes": boxes_list[i].cpu(),
"labels": labels_list[i].cpu(),
"scale": 1.0,
})
print(f" shard with {len(shard)} entries")
# Precompute targets for each image
print("\n2. Precomputing targets for each image...")
t0 = time.time()
shard = precompute_shard_targets(shard, device=device)
t_precompute = time.time() - t0
print(f" precompute time: {t_precompute*1000:.1f} ms ({t_precompute*1000/B:.1f} ms/image)")
for i, e in enumerate(shard):
n_pos = (e["tgt_cls"] >= 0).sum().item()
print(f" img {i}: tgt_cls shape {e['tgt_cls'].shape}, {n_pos} positives")
# Verify equivalence with in-line assign_targets_batched
print("\n3. Verifying equivalence with in-line assign_targets_batched...")
from cache_and_train_fast import assign_targets_batched
feat_sizes = [(H * 2, H * 2), (H, H), (H // 2, H // 2), (H // 4, H // 4), (H // 8, H // 8)]
locs_per_level = make_locations(feat_sizes, STRIDES, torch.device(device))
all_locs = torch.cat(locs_per_level, 0)
n_per_level = [loc.shape[0] for loc in locs_per_level]
level_ranges = []
cumsum = 0
strides_per_loc = torch.zeros(all_locs.shape[0], device=device)
for i, n in enumerate(n_per_level):
lo, hi = SIZE_RANGES[i]
level_ranges.append((cumsum, cumsum + n, STRIDES[i], lo, hi))
strides_per_loc[cumsum:cumsum + n] = STRIDES[i]
cumsum += n
max_m = max(b.shape[0] for b in boxes_list)
boxes_padded = torch.zeros(B, max_m, 4, device=device)
labels_padded = torch.zeros(B, max_m, dtype=torch.long, device=device)
valid_mask = torch.zeros(B, max_m, dtype=torch.bool, device=device)
for i in range(B):
m = boxes_list[i].shape[0]
boxes_padded[i, :m] = boxes_list[i]
labels_padded[i, :m] = labels_list[i]
valid_mask[i, :m] = True
inline_cls, inline_reg, inline_ctr = assign_targets_batched(
all_locs, level_ranges, boxes_padded, labels_padded, valid_mask, strides_per_loc)
cached_cls = torch.stack([e["tgt_cls"].to(device).long() for e in shard])
cached_reg = torch.stack([e["tgt_reg"].to(device).float() for e in shard])
cached_ctr = torch.stack([e["tgt_ctr"].to(device).float() for e in shard])
cls_match = torch.equal(cached_cls, inline_cls)
reg_diff = (cached_reg - inline_reg)[inline_cls >= 0].abs().max().item() if (inline_cls >= 0).any() else 0
ctr_diff = (cached_ctr - inline_ctr)[inline_cls >= 0].abs().max().item() if (inline_cls >= 0).any() else 0
print(f" cls equal: {cls_match}")
print(f" reg max abs diff (positives only, fp16 precision): {reg_diff:.6f}")
print(f" ctr max abs diff (positives only, fp16 precision): {ctr_diff:.6f}")
if not cls_match:
n_diff = (cached_cls != inline_cls).sum().item()
print(f" WARNING: {n_diff} cls mismatches")
sys.exit(1)
if reg_diff > 0.5 or ctr_diff > 0.01:
print(f" WARNING: reg/ctr drift exceeds fp16 tolerance")
sys.exit(1)
print("\n4. Benchmarking loss computation: cached vs in-line...")
# Build mock predictions
cls_per = [torch.randn(B, 80, h, w, device=device) for (h, w) in feat_sizes]
reg_per = [torch.rand(B, 4, h, w, device=device) * 30 for (h, w) in feat_sizes]
ctr_per = [torch.randn(B, 1, h, w, device=device) for (h, w) in feat_sizes]
from cache_and_train_fast import compute_loss
# Warmup
for _ in range(3):
_ = compute_loss(cls_per, reg_per, ctr_per, locs_per_level, boxes_list, labels_list)
_ = precompute_loss_with_cache(cls_per, reg_per, ctr_per, cached_cls, cached_reg, cached_ctr)
torch.cuda.synchronize() if device == "cuda" else None
N_ITERS = 100
t0 = time.time()
for _ in range(N_ITERS):
_ = compute_loss(cls_per, reg_per, ctr_per, locs_per_level, boxes_list, labels_list)
if device == "cuda": torch.cuda.synchronize()
inline_time = (time.time() - t0) / N_ITERS
t0 = time.time()
for _ in range(N_ITERS):
_ = precompute_loss_with_cache(cls_per, reg_per, ctr_per, cached_cls, cached_reg, cached_ctr)
if device == "cuda": torch.cuda.synchronize()
cached_time = (time.time() - t0) / N_ITERS
print(f" in-line compute_loss: {inline_time*1000:.2f} ms/iter")
print(f" cached compute_loss: {cached_time*1000:.2f} ms/iter")
print(f" speedup: {inline_time / cached_time:.2f}x")
print("\nAll Option 4 tests passed.")