train 661676 lines · val 34622 lines charset: 393 characters (min freq 20) dropped 366 train / 16 val lines containing unspellable characters /home/danial/.cache/claude-worktrees/bagrat-fund-old-scripts/services/ocr-lab/venv/lib/python3.12/site-packages/torch/cuda/__init__.py:829: UserWarning: Can't initialize NVML warnings.warn("Can't initialize NVML") resumed from runs/rec-4script7/init.pt epoch 1 step 200/20665 loss 0.0386 (52s) epoch 1 step 400/20665 loss 0.0387 (104s) epoch 1 step 600/20665 loss 0.0412 (154s) epoch 1 step 800/20665 loss 0.0397 (204s) epoch 1 step 1000/20665 loss 0.0396 (254s) epoch 1 step 1200/20665 loss 0.0380 (303s) epoch 1 step 1400/20665 loss 0.0367 (352s) epoch 1 step 1600/20665 loss 0.0358 (402s) epoch 1 step 1800/20665 loss 0.0353 (452s) epoch 1 step 2000/20665 loss 0.0360 (500s) epoch 1 step 2200/20665 loss 0.0362 (550s) epoch 1 step 2400/20665 loss 0.0365 (599s) epoch 1 step 2600/20665 loss 0.0365 (649s) epoch 1 step 2800/20665 loss 0.0372 (697s) epoch 1 step 3000/20665 loss 0.0372 (746s) epoch 1 step 3200/20665 loss 0.0370 (795s) epoch 1 step 3400/20665 loss 0.0370 (843s) epoch 1 step 3600/20665 loss 0.0371 (892s) epoch 1 step 3800/20665 loss 0.0374 (941s) epoch 1 step 4000/20665 loss 0.0374 (990s) epoch 1 step 4200/20665 loss 0.0372 (1038s) epoch 1 step 4400/20665 loss 0.0373 (1087s) epoch 1 step 4600/20665 loss 0.0370 (1136s) epoch 1 step 4800/20665 loss 0.0371 (1184s) epoch 1 step 5000/20665 loss 0.0371 (1232s) epoch 1 step 5200/20665 loss 0.0370 (1282s) epoch 1 step 5400/20665 loss 0.0370 (1331s) epoch 1 step 5600/20665 loss 0.0371 (1380s) epoch 1 step 5800/20665 loss 0.0371 (1429s) epoch 1 step 6000/20665 loss 0.0370 (1478s) epoch 1 step 6200/20665 loss 0.0368 (1528s) epoch 1 step 6400/20665 loss 0.0367 (1576s) epoch 1 step 6600/20665 loss 0.0366 (1625s) epoch 1 step 6800/20665 loss 0.0366 (1673s) epoch 1 step 7000/20665 loss 0.0367 (1722s) epoch 1 step 7200/20665 loss 0.0367 (1771s) epoch 1 step 7400/20665 loss 0.0368 (1819s) epoch 1 step 7600/20665 loss 0.0367 (1868s) epoch 1 step 7800/20665 loss 0.0366 (1916s) epoch 1 step 8000/20665 loss 0.0366 (1964s) epoch 1 step 8200/20665 loss 0.0365 (2013s) epoch 1 step 8400/20665 loss 0.0365 (2061s) epoch 1 step 8600/20665 loss 0.0365 (2109s) epoch 1 step 8800/20665 loss 0.0365 (2157s) epoch 1 step 9000/20665 loss 0.0364 (2206s) epoch 1 step 9200/20665 loss 0.0363 (2254s) epoch 1 step 9400/20665 loss 0.0363 (2302s) epoch 1 step 9600/20665 loss 0.0362 (2351s) epoch 1 step 9800/20665 loss 0.0362 (2399s) epoch 1 step 10000/20665 loss 0.0361 (2448s) epoch 1 step 10200/20665 loss 0.0361 (2496s) epoch 1 step 10400/20665 loss 0.0362 (2544s) epoch 1 step 10600/20665 loss 0.0362 (2593s) epoch 1 step 10800/20665 loss 0.0363 (2641s) epoch 1 step 11000/20665 loss 0.0363 (2689s) epoch 1 step 11200/20665 loss 0.0363 (2737s) epoch 1 step 11400/20665 loss 0.0362 (2786s) epoch 1 step 11600/20665 loss 0.0364 (2834s) epoch 1 step 11800/20665 loss 0.0363 (2883s) epoch 1 step 12000/20665 loss 0.0363 (2931s) epoch 1 step 12200/20665 loss 0.0364 (2980s) epoch 1 step 12400/20665 loss 0.0365 (3028s) epoch 1 step 12600/20665 loss 0.0364 (3076s) epoch 1 step 12800/20665 loss 0.0364 (3125s) epoch 1 step 13000/20665 loss 0.0363 (3173s) epoch 1 step 13200/20665 loss 0.0363 (3222s) epoch 1 step 13400/20665 loss 0.0363 (3270s) epoch 1 step 13600/20665 loss 0.0363 (3318s) epoch 1 step 13800/20665 loss 0.0361 (3366s) epoch 1 step 14000/20665 loss 0.0363 (3414s) epoch 1 step 14200/20665 loss 0.0362 (3463s) epoch 1 step 14400/20665 loss 0.0362 (3511s) epoch 1 step 14600/20665 loss 0.0361 (3559s) epoch 1 step 14800/20665 loss 0.0362 (3608s) epoch 1 step 15000/20665 loss 0.0362 (3656s) epoch 1 step 15200/20665 loss 0.0362 (3704s) epoch 1 step 15400/20665 loss 0.0361 (3752s) epoch 1 step 15600/20665 loss 0.0362 (3800s) epoch 1 step 15800/20665 loss 0.0362 (3848s) epoch 1 step 16000/20665 loss 0.0361 (3897s) epoch 1 step 16200/20665 loss 0.0361 (3945s) epoch 1 step 16400/20665 loss 0.0360 (3993s) epoch 1 step 16600/20665 loss 0.0360 (4042s) epoch 1 step 16800/20665 loss 0.0360 (4090s) epoch 1 step 17000/20665 loss 0.0359 (4138s) epoch 1 step 17200/20665 loss 0.0359 (4186s) epoch 1 step 17400/20665 loss 0.0358 (4235s) epoch 1 step 17600/20665 loss 0.0358 (4283s) epoch 1 step 17800/20665 loss 0.0358 (4331s) epoch 1 step 18000/20665 loss 0.0358 (4380s) epoch 1 step 18200/20665 loss 0.0358 (4428s) epoch 1 step 18400/20665 loss 0.0357 (4476s) epoch 1 step 18600/20665 loss 0.0356 (4524s) epoch 1 step 18800/20665 loss 0.0356 (4572s) epoch 1 step 19000/20665 loss 0.0356 (4621s) epoch 1 step 19200/20665 loss 0.0357 (4669s) epoch 1 step 19400/20665 loss 0.0357 (4717s) epoch 1 step 19600/20665 loss 0.0357 (4765s) epoch 1 step 19800/20665 loss 0.0357 (4813s) epoch 1 step 20000/20665 loss 0.0356 (4861s) epoch 1 step 20200/20665 loss 0.0357 (4911s) epoch 1 step 20400/20665 loss 0.0356 (4960s) epoch 1 step 20600/20665 loss 0.0356 (5009s) epoch 1/2 loss 0.0356 CER verbatim 0.0106 normalized 0.0104 (5104s) ↑ best CER 0.0104 epoch 2 step 200/20665 loss 0.0284 (49s) epoch 2 step 400/20665 loss 0.0280 (98s) epoch 2 step 600/20665 loss 0.0297 (146s) epoch 2 step 800/20665 loss 0.0287 (194s) epoch 2 step 1000/20665 loss 0.0285 (243s) epoch 2 step 1200/20665 loss 0.0287 (291s) epoch 2 step 1400/20665 loss 0.0292 (339s) epoch 2 step 1600/20665 loss 0.0294 (388s) epoch 2 step 1800/20665 loss 0.0293 (436s) epoch 2 step 2000/20665 loss 0.0300 (484s) epoch 2 step 2200/20665 loss 0.0298 (533s) epoch 2 step 2400/20665 loss 0.0298 (581s) epoch 2 step 2600/20665 loss 0.0295 (629s) epoch 2 step 2800/20665 loss 0.0295 (678s) epoch 2 step 3000/20665 loss 0.0295 (726s) epoch 2 step 3200/20665 loss 0.0299 (774s) epoch 2 step 3400/20665 loss 0.0298 (823s) epoch 2 step 3600/20665 loss 0.0298 (871s) epoch 2 step 3800/20665 loss 0.0299 (920s) epoch 2 step 4000/20665 loss 0.0299 (968s) epoch 2 step 4200/20665 loss 0.0299 (1017s) epoch 2 step 4400/20665 loss 0.0300 (1065s) epoch 2 step 4600/20665 loss 0.0300 (1114s) epoch 2 step 4800/20665 loss 0.0298 (1162s) epoch 2 step 5000/20665 loss 0.0296 (1211s) epoch 2 step 5200/20665 loss 0.0293 (1259s) epoch 2 step 5400/20665 loss 0.0292 (1308s) epoch 2 step 5600/20665 loss 0.0293 (1356s) epoch 2 step 5800/20665 loss 0.0292 (1404s) epoch 2 step 6000/20665 loss 0.0292 (1453s) epoch 2 step 6200/20665 loss 0.0291 (1501s) epoch 2 step 6400/20665 loss 0.0290 (1550s) epoch 2 step 6600/20665 loss 0.0288 (1599s) epoch 2 step 6800/20665 loss 0.0289 (1647s) epoch 2 step 7000/20665 loss 0.0290 (1696s) epoch 2 step 7200/20665 loss 0.0290 (1744s) epoch 2 step 7400/20665 loss 0.0289 (1793s) epoch 2 step 7600/20665 loss 0.0291 (1841s) epoch 2 step 7800/20665 loss 0.0291 (1890s) epoch 2 step 8000/20665 loss 0.0291 (1938s) epoch 2 step 8200/20665 loss 0.0290 (1986s) epoch 2 step 8400/20665 loss 0.0290 (2035s) epoch 2 step 8600/20665 loss 0.0290 (2083s) epoch 2 step 8800/20665 loss 0.0289 (2131s) epoch 2 step 9000/20665 loss 0.0289 (2180s) epoch 2 step 9200/20665 loss 0.0288 (2228s) epoch 2 step 9400/20665 loss 0.0288 (2276s) epoch 2 step 9600/20665 loss 0.0287 (2325s) epoch 2 step 9800/20665 loss 0.0286 (2373s) epoch 2 step 10000/20665 loss 0.0286 (2422s) epoch 2 step 10200/20665 loss 0.0286 (2470s) epoch 2 step 10400/20665 loss 0.0286 (2518s) epoch 2 step 10600/20665 loss 0.0286 (2566s) epoch 2 step 10800/20665 loss 0.0285 (2615s) epoch 2 step 11000/20665 loss 0.0285 (2663s) epoch 2 step 11200/20665 loss 0.0284 (2712s) epoch 2 step 11400/20665 loss 0.0285 (2760s) epoch 2 step 11600/20665 loss 0.0285 (2809s) epoch 2 step 11800/20665 loss 0.0284 (2857s) epoch 2 step 12000/20665 loss 0.0284 (2906s) epoch 2 step 12200/20665 loss 0.0284 (2954s) epoch 2 step 12400/20665 loss 0.0285 (3002s) epoch 2 step 12600/20665 loss 0.0285 (3051s) epoch 2 step 12800/20665 loss 0.0284 (3099s) epoch 2 step 13000/20665 loss 0.0285 (3148s) epoch 2 step 13200/20665 loss 0.0284 (3196s) epoch 2 step 13400/20665 loss 0.0285 (3244s) epoch 2 step 13600/20665 loss 0.0284 (3293s) epoch 2 step 13800/20665 loss 0.0285 (3341s) epoch 2 step 14000/20665 loss 0.0284 (3391s) epoch 2 step 14200/20665 loss 0.0284 (3440s) epoch 2 step 14400/20665 loss 0.0284 (3490s) epoch 2 step 14600/20665 loss 0.0285 (3539s) epoch 2 step 14800/20665 loss 0.0285 (3588s) epoch 2 step 15000/20665 loss 0.0285 (3637s) epoch 2 step 15200/20665 loss 0.0284 (3686s) epoch 2 step 15400/20665 loss 0.0285 (3734s) epoch 2 step 15600/20665 loss 0.0284 (3783s) epoch 2 step 15800/20665 loss 0.0284 (3831s) epoch 2 step 16000/20665 loss 0.0284 (3880s) epoch 2 step 16200/20665 loss 0.0284 (3928s) epoch 2 step 16400/20665 loss 0.0283 (3976s) epoch 2 step 16600/20665 loss 0.0283 (4025s) epoch 2 step 16800/20665 loss 0.0283 (4073s) epoch 2 step 17000/20665 loss 0.0282 (4122s) epoch 2 step 17200/20665 loss 0.0283 (4171s) epoch 2 step 17400/20665 loss 0.0284 (4219s) epoch 2 step 17600/20665 loss 0.0284 (4267s) epoch 2 step 17800/20665 loss 0.0284 (4316s) epoch 2 step 18000/20665 loss 0.0284 (4364s) epoch 2 step 18200/20665 loss 0.0284 (4413s) epoch 2 step 18400/20665 loss 0.0284 (4462s) epoch 2 step 18600/20665 loss 0.0283 (4510s) epoch 2 step 18800/20665 loss 0.0283 (4558s) epoch 2 step 19000/20665 loss 0.0283 (4607s) epoch 2 step 19200/20665 loss 0.0284 (4655s) epoch 2 step 19400/20665 loss 0.0283 (4704s) epoch 2 step 19600/20665 loss 0.0283 (4753s) epoch 2 step 19800/20665 loss 0.0283 (4803s) epoch 2 step 20000/20665 loss 0.0283 (4852s) epoch 2 step 20200/20665 loss 0.0283 (4900s) epoch 2 step 20400/20665 loss 0.0282 (4949s) epoch 2 step 20600/20665 loss 0.0282 (4998s) epoch 2/2 loss 0.0282 CER verbatim 0.0102 normalized 0.0100 (5094s) ↑ best CER 0.0100 done. best normalized CER 0.0100 (gate: ≤ 0.10)