Pikaliov/caption-test
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity
34 Commits 4 Branches 0 Tags
6b7bcae198ec013b49c1d084bfd9d2f0fcf4dfdc
Commit Graph

9 Commits

Author SHA1 Message Date
pikaliov
6b7bcae198 Add gradient checkpointing for DINOv3 and DGTRS-CLIP (bs 8→24) 
- DINOv3: checkpoint each of 24 transformer blocks (recompute on backward)
- DGTRS-CLIP: checkpoint each of 12 transformer blocks
- Enables batch_size=24 on RTX 4090 (was 8 without checkpointing)
- Peak VRAM: 20.3 GB at bs=24 (was OOM at bs=16 before)
- ~20-30% slower per step, but 3x more in-batch negatives (23 vs 7)
- Enabled by default (gradient_checkpointing=True in config)
- Update README with VRAM benchmarks and checkpointing docs

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
 2026-04-21 21:34:31 +03:00
pikaliov
da2d2ea90e Switch to shared DINOv3 WEB encoder (saves ~4-5 GB VRAM) 
- Single DINOv3 WEB for both drone and satellite branches (shared_encoder=True default)
- One set of MONA adapters instead of two: 7M trainable vs 14M
- Total params: 438M (was 748M), trainable: 10.6M (was 17.6M)
- Asymmetric mode still available via shared_encoder=False
- Add gradient accumulation (grad_accum_steps, --grad-accum CLI flag)
- Update model summary in README

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
 2026-04-21 21:25:46 +03:00
pikaliov
a39f8a9655 Add MONA adapters for DINOv3 + LoRA for DGTRS-CLIP text encoder 
MONA (Multi-Cognitive One-Shot Nested Adaptation, CVPR 2025):
- 2 adapters per DINOv3 block (after MSA, after MLP) × 24 layers × 2 encoders
- MonaOp: parallel DWConv 3×3/5×5/7×7 + 1×1 projector
- ScaledLayerNorm + down(1024→64) + MonaOp + GELU + up(64→1024)
- 7M params per encoder, 14M total for drone+sat

LoRA (Low-Rank Adaptation):
- Q and V projections in all 12 DGTRS-CLIP transformer blocks
- rank=4, ~147K params total
- Replaces partial unfreeze (was ~7.6M for last block)

Both adapters run in fp32 (torch.amp.autocast disabled) to avoid
AMP gradient overflow. Frozen backbone layers still run in fp16.

Total trainable: 17.6M / 748M (2.35%)
  MONA (2×DINOv3): 14.0M
  LoRA (DGTRS):    147K
  TextFusionMLP:   3.4M
  Gates + logit:   3

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
 2026-04-21 19:24:01 +03:00
pikaliov
0c41c1f017 Remove projections (1024 native), add satellite text, dual GatedFusion 
Architecture changes:
- Removed proj_drone/proj_sat (1024→512): retrieval space is now
  DINOv3 native 1024-dim, no information loss from projection
- TextFusionMLP: 2304→1024→1024 (was 2304→768→512), shared between branches
- Gallery branch now uses satellite captions (L1/L2/L3) via shared TextFusionMLP
- Two separate GatedFusion gates: α_q (query) and α_g (gallery)
- For sat images without captions (~57%): gate passes image features through

Dataset changes:
- GTAUAVDataset now loads satellite captions from caption index
- collate_gtauav_batch includes sat_caption_l1/l2/l3

Training loop:
- Passes satellite captions to model forward
- Logs both gate_q and gate_g values

11.1M trainable / 734M total (1.51%)

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
 2026-04-21 19:01:30 +03:00
pikaliov
433fa40ed6 Replace broken LRSCLIPTextEncoder with official DGTRS architecture 
Root cause of NaN: our open_clip wrapper had 3 bugs:
1. Positional embeddings summed for all positions instead of masked
   (official: mask1 for pos 0-19, mask2 for pos 20-247)
2. open_clip uses batch-first transformer, DGTRS uses sequence-first
   (LND format with nn.MultiheadAttention)
3. open_clip tokenizer truncates to 77 tokens, DGTRS needs 248

Fix: copied official DGTRS text encoder architecture from
github.com/MitsuiChen14/DGTRS (Apache-2.0):
- src/models/dgtrs/model.py: DGTRSTextEncoder, build_model,
  load_dgtrs_text_encoder, tokenize_dgtrs
- src/models/dgtrs/simple_tokenizer.py: BPE tokenizer (248 tokens)
- src/models/dgtrs/bpe_simple_vocab_16e6.txt.gz: vocabulary

Removed: LRSCLIPTextEncoder class, open_clip dependency for text encoding

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
 2026-04-21 18:35:41 +03:00
pikaliov
a214320d81 Suppress open_clip 'no pretrained weights' warning for LRSCLIP init 
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
 2026-04-21 18:26:59 +03:00
pikaliov
44bce3096c Add model save/load and --resume for training continuation 
- AsymmetricEncoder.save_checkpoint(): saves model_state + metadata
- AsymmetricEncoder.load_checkpoint(): rebuilds model with frozen backbones,
  then loads trainable weights from checkpoint
- --resume flag restores optimizer, loss (learnable tau), and scheduler state
- Training continues from the saved epoch

Usage:
  python -m src.training.train_gtauav --resume out/gtauav/with_text/ckpt_epoch004.pt

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
 2026-04-21 18:14:54 +03:00
pikaliov
998d52cb57 Improve training: learnable temperature, per-group LR, warmup, augmentations 
Loss:
- Learnable temperature (CLIP-style logit_scale) with clamp [0.01, 0.5]
- Replaces fixed cosine schedule (still available via --no-learnable-temp)
- Default tau_init=0.07

Optimizer:
- Per-group LR: projections 1e-4, text encoder 1e-5 (10x lower)
- Learnable temperature included in projection param group

Scheduler:
- Linear warmup (2 epochs default) + cosine annealing
- Per-step scheduling (not per-epoch)

Augmentations (separate drone/satellite):
- Drone: RandomResizedCrop(0.7-1.0), HFlip, Rotation(15), ColorJitter,
  RandomGrayscale(0.05), GaussianBlur
- Satellite: RandomResizedCrop(0.7-1.0), HFlip, ColorJitter, RandomGrayscale
- Eval: clean Resize+CenterCrop (no augmentation)

Dataset: supports separate drone_transform/sat_transform args

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
 2026-04-21 18:07:17 +03:00
pikaliov
6ad9c4d149 Add GTA-UAV experiment: asymmetric DINOv3 + LRSCLIP text encoder 
V3 architecture for CVGL caption validation on GTA-UAV-LR dataset:
- AsymmetricEncoder: DINOv3 ViT-L/16 (LVD drone + SAT satellite, frozen)
  + LRSCLIP/DGTRS-CLIP ViT-L-14 text encoder (248 tok, partial unfreeze)
- L1/L2/L3 hierarchical captions from VLM-generated descriptions
- TextFusionMLP (concat 3x768 -> MLP -> 512) + GatedFusion
- Segmentation filter: exclude images with >=90% background+water
- 10.9M trainable / 733M total params, 256x256 input
- coloredlogs + tqdm + emoji for training UX
- Baseline mode (--baseline): image-only, no text encoder loaded

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
 2026-04-21 17:54:27 +03:00