Pikaliov/caption-test
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Add per-query hard negative mining to InfoNCELoss

Browse Source 
New `hard_mining_k` parameter on InfoNCELoss. When >0 and queue is non-empty,
each query row keeps only its K highest-similarity queue entries (via
`torch.topk`) as negatives, instead of the full queue. Fully vectorized —
no Python loop, no extra forward pass.

Rationale: the memory bank grows to 4096 detached gallery embeddings, but
most are easy negatives that contribute almost nothing to the gradient.
Hard mining focuses compute on the small subset that actually shapes the
decision boundary. +2-3% R@1 in similar contrastive setups.

Edge cases:
- K=0: mining disabled, full queue used (original behavior).
- K >= queue size: falls back to full queue (e.g. warmup when queue is small).
- Queue empty: in-batch only, no changes.

Default in `gtauav_balanced.gin`: K=512 (1/8 of queue). Smoke-train updated
to exercise the full memory-bank + mining path.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This commit is contained in:
pikaliov
2026-04-24 16:05:32 +03:00
parent a499fcfd65
commit c30726998b
4 changed files with 36 additions and 11 deletions
Download Patch File Download Diff File Expand all files Collapse all files

3
CLAUDE.md
View File

@@ -101,7 +101,7 @@ Eval: Resize(256) + CenterCrop(256) + ImageNet normalization.
| `src/models/asymmetric_encoder.py` | DINOv3ViT + TextFusionMLP + AsymmetricEncoder + GatedFusion + encode_query/encode_gallery (per-sample caption mask) |
| `src/datasets/gtauav_dataset.py` | GTA-UAV-LR loader + L1/L2/L3 captions + GTAUAVSatGallery/GTAUAVDroneQuery (full retrieval eval) |
| `src/datasets/mutually_exclusive_sampler.py` | BatchSampler: drone'ы в батче не делят sat_candidates (no false negatives) |
| `src/losses/multi_infonce.py` | **Primary:** SymmetricInfoNCE + MoCo queue, learnable τ clamp [0.01, 0.1], weights q2g=0.6 g2q=0.4 |
| `src/losses/multi_infonce.py` | **Primary:** SymmetricInfoNCE + MoCo queue, learnable τ clamp [0.01, 0.1], weights q2g=0.6 g2q=0.4, `hard_mining_k` для top-K hardest negatives |
| `src/losses/weighted_infonce.py` | Alternative: per-sample adaptive label smoothing (активируется `loss_type="weighted"`) |
| `src/losses/hard_negatives.py` | NegativeMemoryBank (MoCo-style FIFO queue 4096 × 1024) |
| `src/training/train_gtauav.py` | Training loop: full-gallery `_evaluate`, mutex sampler wiring, loss_type switch |
@@ -212,6 +212,7 @@ Meta-файл `meta/seg_filter.json`: исключение изображени
- **Optimizer:** AdamW, per-group LR: proj=1e-4, text=1e-5 (10x lower)
- **Scheduler:** linear warmup (2 epochs) + cosine annealing (per-step)
- **Loss:** SymmetricInfoNCE (q2g=0.6, g2q=0.4) с learnable τ (init=0.07, clamp [0.01, 0.1])
- **Hard mining:** top-K=512 hardest negatives per query из MoCo queue (размер 4096); `hard_mining_k=0` отключает
- **Batch sampler:** MutuallyExclusiveSampler — batches disjoint по sat_candidates (на bs=8 сохраняет 100% entries)
- **Eval:** full satellite gallery (~2684 unique tiles для test_20) с multi-match R@K (учитывает все positive/semi-positive)
- **Augmentations:**

1
conf/gtauav_balanced.gin
View File

@@ -56,3 +56,4 @@ InfoNCELoss.weight_q2g = 0.6
InfoNCELoss.weight_g2q = 0.4
InfoNCELoss.tau_min = 0.01
InfoNCELoss.tau_max = 0.1
InfoNCELoss.hard_mining_k = 512 # 0 = use whole queue (disable mining)

15
scripts/smoke_train.py
View File

@@ -8,6 +8,7 @@ from torch.utils.data import DataLoader
from src.datasets.gtauav_dataset import GTAUAVDataset, collate_gtauav_batch
from src.datasets.mutually_exclusive_sampler import MutuallyExclusiveSampler
from src.losses.hard_negatives import NegativeMemoryBank
from src.losses.multi_infonce import InfoNCELoss
from src.models.asymmetric_encoder import AsymmetricEncoder, get_dino_transform
@@ -40,13 +41,15 @@ def main() -> None:
temperature_init=0.07, learnable_temperature=True,
label_smoothing=0.1, weight_q2g=0.6, weight_g2q=0.4,
tau_min=0.01, tau_max=0.1,
hard_mining_k=512,
).to("cuda")
neg_bank = NegativeMemoryBank(size=4096, dim=model.embed_dim).to("cuda")
trainable = [p for p in model.trainable_parameters()] + list(loss_fn.parameters())
opt = torch.optim.AdamW(trainable, lr=1e-4)
it = iter(loader)
for step in range(2):
for step in range(3):
batch = next(it)
opt.zero_grad()
emb = model(
@@ -55,9 +58,11 @@ def main() -> None:
caption_l1=batch["caption_l1"], caption_l2=batch["caption_l2"], caption_l3=batch["caption_l3"],
sat_caption_l1=batch["sat_caption_l1"], sat_caption_l2=batch["sat_caption_l2"], sat_caption_l3=batch["sat_caption_l3"],
)
out = loss_fn(emb, epoch=0, total_epochs=10)
queue = neg_bank.get_queue()
out = loss_fn(emb, epoch=0, total_epochs=10, queue_negatives=queue)
out["total"].backward()
opt.step()
neg_bank.enqueue(emb["gallery"].detach())
# Verify mutual exclusion in batch
batch_sats = [set(ds.entries[i]["sat_candidates"]) for i in batch.get("__indices__", range(8))]
@@ -67,12 +72,12 @@ def main() -> None:
f" step {step}: loss={out['total'].item():.4f} "
f"tau={out['temperature'].item():.4f} "
f"gate_q={out['gate_q'].item():.3f} gate_g={out['gate_g'].item():.3f} "
f"n_drone_caps={sum(1 for t in batch['caption_l1'] if t)} "
f"n_sat_caps={sum(1 for t in batch['sat_caption_l1'] if t)}"
f"queue_size={queue.shape[0] if queue is not None else 0} "
f"mining_k={loss_fn.hard_mining_k if queue is not None and queue.shape[0] > loss_fn.hard_mining_k else 'full'}"
)
assert torch.isfinite(out["total"]).all(), "Loss not finite!"
print("OK: 2 train steps completed with finite loss")
print("OK: 3 train steps completed with finite loss (hard mining K=512)")
if __name__ == "__main__":

28
src/losses/multi_infonce.py
View File

@@ -24,6 +24,7 @@ def _symmetric_info_nce(
weight_a2b: float = 0.5,
weight_b2a: float = 0.5,
queue_negatives: torch.Tensor | None = None,
hard_mining_k: int = 0,
) -> torch.Tensor:
"""Weighted symmetric InfoNCE with optional hard negative queue.
@@ -31,20 +32,31 @@ def _symmetric_info_nce(
emb_a: Query embeddings [B, D].
emb_b: Gallery embeddings [B, D]. Positives on diagonal.
queue_negatives: Extra gallery negatives [Q, D] from memory bank.
hard_mining_k: If > 0 and queue is non-empty, use only the top-K
hardest (highest-similarity) queue entries per query instead
of the full queue. Per-query selection — each row gets its
own K negatives gathered via `topk`.
"""
batch_size = emb_a.size(0)
emb_a_f = emb_a.float()
emb_b_f = emb_b.float()
if queue_negatives is not None and queue_negatives.shape[0] > 0:
# a→b: query sees B in-batch + Q queue negatives.
all_b = torch.cat([emb_b_f, queue_negatives.float()], dim=0) # [B+Q, D]
logits_a2b = emb_a_f @ all_b.t() / temperature # [B, B+Q]
queue_f = queue_negatives.float()
sim_inbatch = emb_a_f @ emb_b_f.t() / temperature # [B, B]
sim_queue = emb_a_f @ queue_f.t() / temperature # [B, Q]
if hard_mining_k > 0 and hard_mining_k < queue_f.shape[0]:
# Per-row top-K — each query gets its own hardest negatives.
sim_queue, _ = sim_queue.topk(k=hard_mining_k, dim=1) # [B, K]
# a→b: [B, B + (Q or K)]. Positive at column `i` for row `i`.
logits_a2b = torch.cat([sim_inbatch, sim_queue], dim=1)
targets_a = torch.arange(batch_size, device=emb_a.device)
loss_a2b = F.cross_entropy(logits_a2b, targets_a, label_smoothing=label_smoothing)
# b→a: gallery sees B in-batch queries (no queue for this direction).
logits_b2a = emb_b_f @ emb_a_f.t() / temperature # [B, B]
# b→a: gallery sees B in-batch queries (queue is gallery-side, irrelevant here).
logits_b2a = sim_inbatch.t() # [B, B]
targets_b = torch.arange(batch_size, device=emb_a.device)
loss_b2a = F.cross_entropy(logits_b2a, targets_b, label_smoothing=label_smoothing)
else:
@@ -83,6 +95,9 @@ class InfoNCELoss(nn.Module):
(CLIP-style logit_scale). If False, uses cosine schedule.
tau_min: Minimum clamp for learnable temperature.
tau_max: Maximum clamp for learnable temperature.
hard_mining_k: If > 0, mine top-K hardest negatives per query from
the memory bank queue instead of using the full queue. 0 disables
mining (queue used whole). Typical values: 256-1024 for queue=4096.
"""
def __init__(
@@ -95,6 +110,7 @@ class InfoNCELoss(nn.Module):
learnable_temperature: bool = True,
tau_min: float = 0.01,
tau_max: float = 0.1,
hard_mining_k: int = 0,
) -> None:
super().__init__()
self.temperature_init = temperature_init
@@ -105,6 +121,7 @@ class InfoNCELoss(nn.Module):
self.learnable_temperature = learnable_temperature
self.tau_min = tau_min
self.tau_max = tau_max
self.hard_mining_k = hard_mining_k
if learnable_temperature:
# Store as log(1/tau) like CLIP's logit_scale.
@@ -167,6 +184,7 @@ class InfoNCELoss(nn.Module):
weight_a2b=self.weight_q2g,
weight_b2a=self.weight_g2q,
queue_negatives=queue_negatives,
hard_mining_k=self.hard_mining_k,
)
gate_q = embeddings.get("gate_q", embeddings.get("gate", 1.0))