This website requires JavaScript.
ac9cdaf6f5
belka_gate_fusions: TODO: fix Incorrect fusion behaviour
2026-05-19 13:30:25 +03:00
a2729c065e
belka_gate_fusions: trainable version
2026-05-18 15:48:49 +03:00
b311913034
belka_gate_fusions: dataset path update to SSD_2_TB
2026-05-18 14:52:41 +03:00
b7542b27de
belka_gate_fusions: inject GateFusionResidual variants to asym and dual encoder
2026-05-18 11:24:58 +03:00
e2fd16e66c
belka_gate_fusions: add gate-variants for CVGL exp.
2026-05-15 16:30:11 +03:00
ecf8071220
checkpoint 15_05
2026-05-15 14:29:42 +03:00
d82fac9190
claude_refactor_v3: Added all preset-debug configs (launch.json in gitignore)
2026-05-08 13:45:53 +03:00
7b3acc633a
Update dataset path (SSD_2_TB)
2026-05-08 09:52:15 +03:00
455ae2e99f
checkpoint 07_05
2026-05-07 15:33:43 +03:00
1c03811af0
trainer_new logger small fix
2026-05-07 13:24:01 +03:00
676dea7932
Update log.md
2026-05-07 10:40:39 +03:00
c9083c89cc
claude_refactor_v3: Test real trainer loading before training loop + Add log.md
2026-05-07 10:32:30 +03:00
9878473419
checkpoint 06_05
2026-05-06 16:19:46 +03:00
2362ce0adb
claude_refactor_v3: Add and passed test on splited gin-configs loading but without weigts
2026-05-06 16:17:36 +03:00
911d2ce4e6
temp_step_pre_test
2026-05-06 10:06:38 +03:00
562a5e2e43
temp_step
2026-05-06 09:45:00 +03:00
4cb5ab97d8
claude_refactor_v3: fix extra lines in trainer_new
2026-05-05 12:37:20 +03:00
248bd331d2
claude_refactor_v3: Updated main (entry point), trainer_new (last version of train_gtauav), check: is extracted evluate() from train to evaluator.py correct in new context
2026-05-05 11:28:09 +03:00
4e148a29bb
checkpoint 04_05
2026-05-04 16:32:22 +03:00
c43b4c82b9
temp_step: verify actual final trainers scripts and remove obsolete
2026-05-04 15:49:41 +03:00
80d1806cee
temp_step: remove sofia branches from training
2026-05-04 15:26:16 +03:00
2bef5a4270
Update README.md: added authors info
2026-05-04 11:59:32 +03:00
7bbf67c465
claude_refactor_v3: Removed old conf with gins
2026-05-04 11:25:42 +03:00
89cb8ab0f7
claude_refactor_v3: New train_gtauav.py, added entry point main.py, added utils
2026-05-04 11:20:14 +03:00
08d328db09
checkpoint 30_04
2026-04-30 16:30:34 +03:00
4b441279e0
claude_refactor_v3: De-duplicate gin-configs, + "in/config_files" with common and cpecific experiment-presets gins
2026-04-30 16:22:04 +03:00
e8a0de7ad3
claude_refactor_v3: Added .py-confs and all presets (nx5 .gin files). TODO: common gins-mapping and prepare to next step
2026-04-30 12:02:15 +03:00
db2b5b32f4
checkpoint 29_04
2026-04-29 16:38:20 +03:00
af207d302c
claude_refactor_v3: Added models configs except sofiav71 - need extra presets for M, L, Tiny
2026-04-29 16:28:28 +03:00
ce9e3dc94b
claude_refactor_v3: Added plan.md, new conf dirictories
2026-04-29 15:50:36 +03:00
bf5e417094
Add new pre_refactor analysis.md
2026-04-29 15:15:40 +03:00
d42ef94821
clean up to baseline
2026-04-29 14:45:31 +03:00
6b928634f4
clean up
2026-04-29 14:43:42 +03:00
3275a41b36
claude_refactor_v2: remove extra configs additions
2026-04-29 14:40:40 +03:00
4bbe1dda1d
Added tmp extras before any code restructurig
2026-04-29 12:17:03 +03:00
5c48b6c8fd
claude_refactor_v2: temp changes before src changes
2026-04-29 11:55:38 +03:00
0d8c82acc3
add sofia models
2026-04-29 08:04:33 +03:00
a27b5a7357
Init refactor branch
2026-04-27 11:57:29 +03:00
c6fcd2222c
Add stripnet_freeze flag for full StripNet fine-tune mode
2026-04-25 16:04:24 +03:00
d4cb2dd300
Add StripNet backbone option with Conv-MONA adaptation
2026-04-25 14:34:53 +03:00
814586ce3b
Add g2q eval metrics and asymmetric MONA-24 configs
2026-04-25 13:47:02 +03:00
8f8cbb14dd
Diagnostic baseline v2: also disable MoCo queue
2026-04-25 05:57:14 +03:00
c25bd64357
Diagnostic baseline: disable DSS + hard mining, fresh output dir
2026-04-24 22:30:30 +03:00
9a7fbff743
Fix plot_combined: fallback from 'total' to 'train_loss'
2026-04-24 22:24:30 +03:00
70f1617317
Fix autograd in-place error: move memory-bank enqueue after backward
2026-04-24 16:40:58 +03:00
8197ab2407
Fix training loop: only pass positive_weights to WeightedInfoNCELoss
2026-04-24 16:38:16 +03:00
cb477f4b40
Simplify model: shared DINOv3 WEB + MONA in last 12/24 blocks
2026-04-24 16:26:17 +03:00
d98d853455
DSS pipeline: GPU kNN, LSH index, embedding cache
2026-04-24 16:20:48 +03:00
f8e0631210
Add DynamicSimilaritySampler — embedding-kNN batches with mutex constraint
2026-04-24 16:12:34 +03:00
c30726998b
Add per-query hard negative mining to InfoNCELoss
2026-04-24 16:05:32 +03:00
a499fcfd65
Fix GTA-UAV eval + training pipeline: full gallery, mutex sampler, per-sample mask
2026-04-24 15:58:27 +03:00
ce7892926f
Add Recall@K and AP panels to train_metrics.png
2026-04-24 13:01:51 +03:00
e0db8d2726
Rename 'total' to 'train_loss' in train.csv
2026-04-24 12:59:47 +03:00
536f904842
Remove eval_loss from train.csv (already in val.csv)
2026-04-24 12:57:37 +03:00
d9c795d9ae
Add train recall/AP to train.csv (merged with loss metrics per epoch)
2026-04-24 12:55:36 +03:00
b6dccbba7b
Fix GTA-UAV evaluation and loss (critical: false negatives + wrong R@K)
2026-04-24 12:40:10 +03:00
04d5307221
Asymmetric encoders + MONA all 24 blocks + 1024-dim + hard negatives
2026-04-24 08:47:33 +03:00
75a4350d18
Update README: AP metric, train/val eval, CSV/plot inventory
2026-04-22 08:27:41 +03:00
4a05d05ccd
Add Average Precision (AP) metric for train and val
2026-04-22 08:08:48 +03:00
df60e83ead
Add val/train loss to evaluation and plots
2026-04-22 08:05:12 +03:00
93ad66810d
Add train recall evaluation (R@K on train subset each epoch)
2026-04-22 08:01:09 +03:00
66c83469ef
Remove per-epoch train/val CSV files (redundant with cumulative CSVs)
2026-04-22 07:57:37 +03:00
1638b250a4
CSVLogger: deduplicate epochs on resume (replace, not append duplicates)
2026-04-22 07:54:33 +03:00
f382325ac6
CSVLogger: load previous epochs on resume for continuous plots
2026-04-22 07:41:32 +03:00
a32cb8ac6c
Fix Grad-CAM numpy() on tensor with grad (add detach)
2026-04-22 07:02:12 +03:00
b0e87c07ef
Fix Grad-CAM crash in shared encoder mode
2026-04-22 00:36:06 +03:00
7cc068e18c
Update README: full architecture refresh (shared DINOv3, MONA bf16 top-12, 512-dim)
2026-04-21 22:44:47 +03:00
e55734425c
MONA bfloat16: safe low-precision (gamma=1e-6 needs bf16 exponent range)
2026-04-21 22:13:35 +03:00
9200772bea
Fix NaN loss: revert MONA to fp32, fix loss logging
2026-04-21 22:09:49 +03:00
916854f124
MONA only on last 12 blocks (skip early low-level blocks)
2026-04-21 21:51:52 +03:00
4a0336e0ff
Update README: MONA vs LoRA rationale, projection 1024→512, model summary
2026-04-21 21:50:47 +03:00
471bc6444a
Restore conv7x7 in MONA (full 3x3+5x5+7x7 multi-scale)
2026-04-21 21:45:27 +03:00
11fc348ecd
Add projection 1024→512, revert MONA bottleneck to 64
2026-04-21 21:44:10 +03:00
d279a6e745
Optimize MONA: fp16, remove conv7x7, bottleneck 64→32
2026-04-21 21:40:23 +03:00
6b7bcae198
Add gradient checkpointing for DINOv3 and DGTRS-CLIP (bs 8→24)
2026-04-21 21:34:31 +03:00
da2d2ea90e
Switch to shared DINOv3 WEB encoder (saves ~4-5 GB VRAM)
2026-04-21 21:25:46 +03:00
46b1208891
Add gradient accumulation support
2026-04-21 21:21:56 +03:00
b5c8015616
Add pair formation and in-batch negative sampling docs to README
2026-04-21 21:14:07 +03:00
8d8d556093
Update README architecture diagram with adapters, loss, and diagnostics pipeline
2026-04-21 20:58:07 +03:00
7b13a4c4db
Write per-batch CSV immediately (append mode, no buffering)
2026-04-21 20:54:48 +03:00
69857e0ade
Fix gin-config ambiguity: pin module name, remove redundant import
2026-04-21 20:45:32 +03:00
29a09349e7
Add ML diagnostics tooling (W&B, TensorBoard, Grad-CAM, profiler) and gin configs
2026-04-21 20:30:50 +03:00
83ce04150d
Add seaborn/matplotlib metric plots, auto-generated after each eval
2026-04-21 19:54:18 +03:00
aee8212454
Add CSV logging with pandas (train.csv, val.csv, per-epoch files)
2026-04-21 19:46:07 +03:00
2db3dff819
Set default epochs to 20
2026-04-21 19:40:00 +03:00
b72c433870
Reduce default batch_size 64→8 (MONA adapters need ~17GB at bs=8)
2026-04-21 19:38:08 +03:00
e3ecb09687
Add VRAM cleanup (gc + empty_cache) before training start
2026-04-21 19:36:25 +03:00
858718431b
Increase default epochs from 10 to 50
2026-04-21 19:32:35 +03:00
4c862b6ec7
Fix README formulas: use ```math code fence for Gitea rendering
2026-04-21 19:28:08 +03:00
082f82f138
Update docs: MONA/LoRA architecture, LaTeX formulas, param summary
2026-04-21 19:26:01 +03:00
a39f8a9655
Add MONA adapters for DINOv3 + LoRA for DGTRS-CLIP text encoder
2026-04-21 19:24:01 +03:00
0c41c1f017
Remove projections (1024 native), add satellite text, dual GatedFusion
2026-04-21 19:01:30 +03:00
219bb779eb
Update docs: full architecture with tensor shapes, formulas, optimizer details
2026-04-21 18:47:38 +03:00
bcb01bcb6d
Fix NaN: compute loss in fp32 outside AMP autocast
2026-04-21 18:42:29 +03:00
f41a0f27fe
Fix epoch display: show 1/10 instead of 0/9
2026-04-21 18:37:05 +03:00
a47dd6308e
Add architecture diagram to README, update docs for DGTRS text encoder
2026-04-21 18:36:24 +03:00
433fa40ed6
Replace broken LRSCLIPTextEncoder with official DGTRS architecture
2026-04-21 18:35:41 +03:00
a214320d81
Suppress open_clip 'no pretrained weights' warning for LRSCLIP init
2026-04-21 18:26:59 +03:00
fa32b2e67f
Suppress spurious lr_scheduler.step() warning from PyTorch
2026-04-21 18:24:54 +03:00
517b87d3d8
Fix scheduler warning: use last_epoch instead of step() loop on resume
2026-04-21 18:23:17 +03:00
pikaliov