One Hundred Layers Tiramisu


(Matthew Kleinsmith) #41

I think I deleted everything on your website. Can you reset it?


(kelvin) #42

I ran it for a while longer. I couldn’t replicate the results in the paper but I got it to: 0.89869.

Here’s the normal run where it got to lr=1e-5:

Epoch 400 : [test : 100%]Average cost test = 0.44551 | jacc test = 0.66192 | acc_test = 0.89572 
poch 401 took 82+25 sec. loss = 0.11973 | jacc = 0.89054 | acc = 0.97155 || loss = 0.23997 | jacc = 0.77234 | acc = 0.94306
poch 402 took 82+25 sec. loss = 0.11927 | jacc = 0.89013 | acc = 0.97166 || loss = 0.23649 | jacc = 0.77555 | acc = 0.94115
poch 403 took 82+25 sec. loss = 0.11985 | jacc = 0.89109 | acc = 0.97148 || loss = 0.26417 | jacc = 0.76499 | acc = 0.93459
poch 404 took 82+25 sec. loss = 0.11890 | jacc = 0.88950 | acc = 0.97170 || loss = 0.26945 | jacc = 0.75154 | acc = 0.93333
poch 405 took 82+25 sec. loss = 0.12935 | jacc = 0.88567 | acc = 0.96967 || loss = 0.28136 | jacc = 0.74973 | acc = 0.92782
poch 406 took 82+26 sec. loss = 0.11787 | jacc = 0.89221 | acc = 0.97205 || loss = 0.22469 | jacc = 0.77774 | acc = 0.94502
poch 407 took 85+25 sec. loss = 0.11889 | jacc = 0.88976 | acc = 0.97159 || loss = 0.26393 | jacc = 0.75832 | acc = 0.93290
poch 408 took 83+25 sec. loss = 0.11745 | jacc = 0.89205 | acc = 0.97224 || loss = 0.24525 | jacc = 0.77972 | acc = 0.93991
poch 409 took 82+25 sec. loss = 0.11855 | jacc = 0.89185 | acc = 0.97160 || loss = 0.26345 | jacc = 0.75329 | acc = 0.93774
poch 410 took 82+25 sec. loss = 0.11986 | jacc = 0.89018 | acc = 0.97140 || loss = 0.26760 | jacc = 0.74310 | acc = 0.93221
poch 411 took 83+25 sec. loss = 0.11563 | jacc = 0.89137 | acc = 0.97276 || loss = 0.24849 | jacc = 0.75993 | acc = 0.93763
poch 412 took 83+24 sec. loss = 0.11678 | jacc = 0.88962 | acc = 0.97258 || loss = 0.24521 | jacc = 0.75927 | acc = 0.94002
poch 413 took 82+25 sec. loss = 0.11797 | jacc = 0.89058 | acc = 0.97187 || loss = 0.23040 | jacc = 0.76507 | acc = 0.94355
poch 414 took 83+25 sec. loss = 0.11659 | jacc = 0.89123 | acc = 0.97231 || loss = 0.23458 | jacc = 0.76906 | acc = 0.94316
poch 415 took 83+25 sec. loss = 0.11489 | jacc = 0.89150 | acc = 0.97286 || loss = 0.23989 | jacc = 0.75547 | acc = 0.94098
poch 416 took 82+25 sec. loss = 0.12547 | jacc = 0.88921 | acc = 0.97163 || loss = 0.23973 | jacc = 0.75282 | acc = 0.94166
poch 417 took 86+25 sec. loss = 0.11506 | jacc = 0.89441 | acc = 0.97281 || loss = 0.26942 | jacc = 0.75133 | acc = 0.93395
poch 418 took 82+25 sec. loss = 0.11449 | jacc = 0.89490 | acc = 0.97313 || loss = 0.25337 | jacc = 0.74915 | acc = 0.93870
poch 419 took 85+25 sec. loss = 0.12026 | jacc = 0.88637 | acc = 0.97145 || loss = 0.23404 | jacc = 0.77759 | acc = 0.94029
poch 420 took 82+25 sec. loss = 0.11444 | jacc = 0.89101 | acc = 0.97293 || loss = 0.25301 | jacc = 0.74736 | acc = 0.93732
poch 421 took 82+25 sec. loss = 0.11824 | jacc = 0.89021 | acc = 0.97134 || loss = 0.22841 | jacc = 0.77869 | acc = 0.94276
poch 422 took 82+25 sec. loss = 0.11593 | jacc = 0.89509 | acc = 0.97250 || loss = 0.23906 | jacc = 0.75243 | acc = 0.93903
poch 423 took 82+25 sec. loss = 0.11677 | jacc = 0.89188 | acc = 0.97246 || loss = 0.23834 | jacc = 0.77156 | acc = 0.94132
poch 424 took 82+25 sec. loss = 0.11761 | jacc = 0.89181 | acc = 0.97173 || loss = 0.29662 | jacc = 0.73143 | acc = 0.92898
poch 425 took 82+25 sec. loss = 0.11273 | jacc = 0.89615 | acc = 0.97347 || loss = 0.27497 | jacc = 0.74034 | acc = 0.93133
poch 426 took 82+25 sec. loss = 0.11621 | jacc = 0.88929 | acc = 0.97195 || loss = 0.24624 | jacc = 0.76099 | acc = 0.93988
poch 427 took 82+25 sec. loss = 0.11508 | jacc = 0.89506 | acc = 0.97261 || loss = 0.26933 | jacc = 0.74811 | acc = 0.93440
poch 428 took 82+25 sec. loss = 0.11462 | jacc = 0.89279 | acc = 0.97279 || loss = 0.23778 | jacc = 0.76040 | acc = 0.94288
poch 429 took 82+25 sec. loss = 0.11853 | jacc = 0.89181 | acc = 0.97119 || loss = 0.24963 | jacc = 0.75770 | acc = 0.93353
poch 430 took 82+25 sec. loss = 0.11514 | jacc = 0.89545 | acc = 0.97259 || loss = 0.23486 | jacc = 0.77042 | acc = 0.94369
poch 431 took 82+25 sec. loss = 0.11392 | jacc = 0.89393 | acc = 0.97299 || loss = 0.26474 | jacc = 0.74743 | acc = 0.93302
poch 432 took 82+25 sec. loss = 0.11553 | jacc = 0.89329 | acc = 0.97219 || loss = 0.24203 | jacc = 0.76679 | acc = 0.93940
poch 433 took 83+25 sec. loss = 0.11351 | jacc = 0.89498 | acc = 0.97311 || loss = 0.27473 | jacc = 0.74366 | acc = 0.93487
poch 434 took 82+25 sec. loss = 0.11429 | jacc = 0.89327 | acc = 0.97272 || loss = 0.25902 | jacc = 0.76335 | acc = 0.93730
poch 435 took 82+25 sec. loss = 0.11501 | jacc = 0.89519 | acc = 0.97231 || loss = 0.27569 | jacc = 0.74799 | acc = 0.93455
poch 436 took 83+25 sec. loss = 0.11103 | jacc = 0.89601 | acc = 0.97371 || loss = 0.33712 | jacc = 0.73761 | acc = 0.92427
poch 437 took 82+25 sec. loss = 0.11639 | jacc = 0.89243 | acc = 0.97196 || loss = 0.29298 | jacc = 0.73720 | acc = 0.92870
poch 438 took 83+25 sec. loss = 0.11264 | jacc = 0.89325 | acc = 0.97306 || loss = 0.29553 | jacc = 0.74728 | acc = 0.92933
poch 439 took 83+25 sec. loss = 0.11321 | jacc = 0.89197 | acc = 0.97276 || loss = 0.24109 | jacc = 0.76869 | acc = 0.94170
poch 440 took 86+25 sec. loss = 0.11046 | jacc = 0.89633 | acc = 0.97395 || loss = 0.26787 | jacc = 0.73956 | acc = 0.93926
poch 441 took 82+25 sec. loss = 0.11163 | jacc = 0.89559 | acc = 0.97339 || loss = 0.27125 | jacc = 0.74363 | acc = 0.93701
poch 442 took 82+25 sec. loss = 0.11218 | jacc = 0.89580 | acc = 0.97331 || loss = 0.25748 | jacc = 0.76202 | acc = 0.93663
poch 443 took 82+25 sec. loss = 0.11136 | jacc = 0.89820 | acc = 0.97347 || loss = 0.24674 | jacc = 0.77040 | acc = 0.93779
poch 444 took 83+25 sec. loss = 0.11079 | jacc = 0.90072 | acc = 0.97362 || loss = 0.23859 | jacc = 0.77370 | acc = 0.94326
poch 445 took 82+26 sec. loss = 0.11462 | jacc = 0.89409 | acc = 0.97218 || loss = 0.24741 | jacc = 0.75840 | acc = 0.94038
poch 446 took 82+25 sec. loss = 0.11599 | jacc = 0.89347 | acc = 0.97187 || loss = 0.26342 | jacc = 0.76131 | acc = 0.93892
poch 447 took 82+25 sec. loss = 0.11324 | jacc = 0.89225 | acc = 0.97295 || loss = 0.25148 | jacc = 0.75535 | acc = 0.93894
poch 448 took 82+25 sec. loss = 0.11048 | jacc = 0.89414 | acc = 0.97353 || loss = 0.23327 | jacc = 0.77955 | acc = 0.94641
poch 449 took 82+25 sec. loss = 0.11203 | jacc = 0.89563 | acc = 0.97314 || loss = 0.26307 | jacc = 0.74764 | acc = 0.93377
poch 450 took 82+26 sec. loss = 0.11160 | jacc = 0.89743 | acc = 0.97333 || loss = 0.28975 | jacc = 0.75140 | acc = 0.93216
poch 451 took 82+25 sec. loss = 0.10995 | jacc = 0.89711 | acc = 0.97377 || loss = 0.24661 | jacc = 0.76891 | acc = 0.94153
poch 452 took 86+25 sec. loss = 0.11034 | jacc = 0.89712 | acc = 0.97367 || loss = 0.23632 | jacc = 0.77147 | acc = 0.94301
poch 453 took 83+25 sec. loss = 0.10992 | jacc = 0.89839 | acc = 0.97387 || loss = 0.24479 | jacc = 0.75545 | acc = 0.94101
poch 454 took 82+25 sec. loss = 0.11264 | jacc = 0.89340 | acc = 0.97285 || loss = 0.24719 | jacc = 0.75742 | acc = 0.94118
poch 455 took 82+26 sec. loss = 0.10833 | jacc = 0.89628 | acc = 0.97442 || loss = 0.24705 | jacc = 0.76254 | acc = 0.94073
poch 456 took 82+25 sec. loss = 0.10871 | jacc = 0.89581 | acc = 0.97407 || loss = 0.21408 | jacc = 0.78399 | acc = 0.94876
poch 457 took 83+25 sec. loss = 0.11039 | jacc = 0.89631 | acc = 0.97335 || loss = 0.22362 | jacc = 0.78561 | acc = 0.94702
poch 458 took 82+25 sec. loss = 0.11027 | jacc = 0.89717 | acc = 0.97353 || loss = 0.22458 | jacc = 0.77480 | acc = 0.94668
poch 459 took 82+25 sec. loss = 0.10861 | jacc = 0.89732 | acc = 0.97416 || loss = 0.24762 | jacc = 0.77305 | acc = 0.94176
poch 460 took 83+25 sec. loss = 0.11066 | jacc = 0.89379 | acc = 0.97323 || loss = 0.25111 | jacc = 0.75966 | acc = 0.94122
poch 461 took 82+25 sec. loss = 0.11021 | jacc = 0.89741 | acc = 0.97354 || loss = 0.23351 | jacc = 0.76731 | acc = 0.94389
poch 462 took 81+25 sec. loss = 0.11145 | jacc = 0.89533 | acc = 0.97296 || loss = 0.23733 | jacc = 0.76419 | acc = 0.94452
poch 463 took 82+25 sec. loss = 0.11108 | jacc = 0.89508 | acc = 0.97324 || loss = 0.25298 | jacc = 0.75648 | acc = 0.93792
poch 464 took 82+25 sec. loss = 0.10795 | jacc = 0.89642 | acc = 0.97426 || loss = 0.25442 | jacc = 0.76069 | acc = 0.93763
poch 465 took 85+25 sec. loss = 0.10888 | jacc = 0.89942 | acc = 0.97396 || loss = 0.23947 | jacc = 0.76814 | acc = 0.93916
poch 466 took 82+25 sec. loss = 0.10902 | jacc = 0.89771 | acc = 0.97382 || loss = 0.23234 | jacc = 0.76390 | acc = 0.94158
poch 467 took 82+25 sec. loss = 0.10734 | jacc = 0.89996 | acc = 0.97443 || loss = 0.25591 | jacc = 0.75830 | acc = 0.93560
poch 468 took 82+25 sec. loss = 0.11036 | jacc = 0.89828 | acc = 0.97324 || loss = 0.25573 | jacc = 0.75614 | acc = 0.93752
poch 469 took 82+25 sec. loss = 0.12138 | jacc = 0.89028 | acc = 0.97077 || loss = 0.23595 | jacc = 0.77278 | acc = 0.94028
poch 470 took 84+25 sec. loss = 0.10809 | jacc = 0.89840 | acc = 0.97405 || loss = 0.26369 | jacc = 0.77248 | acc = 0.93661
poch 471 took 82+25 sec. loss = 0.10642 | jacc = 0.89961 | acc = 0.97468 || loss = 0.26021 | jacc = 0.74446 | acc = 0.93933
poch 472 took 82+25 sec. loss = 0.10954 | jacc = 0.89927 | acc = 0.97347 || loss = 0.25040 | jacc = 0.76289 | acc = 0.93823
poch 473 took 82+25 sec. loss = 0.10788 | jacc = 0.89954 | acc = 0.97423 || loss = 0.23172 | jacc = 0.76876 | acc = 0.94112
poch 474 took 82+25 sec. loss = 0.10721 | jacc = 0.89636 | acc = 0.97433 || loss = 0.24835 | jacc = 0.76390 | acc = 0.93861
poch 475 took 81+26 sec. loss = 0.10639 | jacc = 0.90093 | acc = 0.97462 || loss = 0.23708 | jacc = 0.76916 | acc = 0.94322
poch 476 took 82+25 sec. loss = 0.10873 | jacc = 0.89869 | acc = 0.97361 || loss = 0.25154 | jacc = 0.77560 | acc = 0.93878
poch 477 took 81+25 sec. loss = 0.10817 | jacc = 0.89557 | acc = 0.97406 || loss = 0.25649 | jacc = 0.74962 | acc = 0.93672
poch 478 took 82+25 sec. loss = 0.10538 | jacc = 0.90038 | acc = 0.97481 || loss = 0.24708 | jacc = 0.74899 | acc = 0.93947
poch 479 took 85+26 sec. loss = 0.11027 | jacc = 0.89843 | acc = 0.97306 || loss = 0.25935 | jacc = 0.74481 | acc = 0.93540
poch 480 took 82+25 sec. loss = 0.10694 | jacc = 0.89890 | acc = 0.97426 || loss = 0.25588 | jacc = 0.74729 | acc = 0.93653
poch 481 took 82+25 sec. loss = 0.10723 | jacc = 0.90034 | acc = 0.97399 || loss = 0.24782 | jacc = 0.75376 | acc = 0.93752
poch 482 took 81+25 sec. loss = 0.10773 | jacc = 0.89900 | acc = 0.97417 || loss = 0.23471 | jacc = 0.76332 | acc = 0.94229
poch 483 took 82+26 sec. loss = 0.10734 | jacc = 0.90115 | acc = 0.97394 || loss = 0.26635 | jacc = 0.76355 | acc = 0.93315
poch 484 took 82+25 sec. loss = 0.10886 | jacc = 0.90031 | acc = 0.97363 || loss = 0.23132 | jacc = 0.77647 | acc = 0.94394
poch 485 took 82+25 sec. loss = 0.10805 | jacc = 0.89846 | acc = 0.97395 || loss = 0.24093 | jacc = 0.76847 | acc = 0.93999
poch 486 took 82+25 sec. loss = 0.10694 | jacc = 0.89849 | acc = 0.97409 || loss = 0.28335 | jacc = 0.74985 | acc = 0.92917
poch 487 took 85+25 sec. loss = 0.10787 | jacc = 0.89991 | acc = 0.97385 || loss = 0.23524 | jacc = 0.76888 | acc = 0.94132
poch 488 took 82+25 sec. loss = 0.10792 | jacc = 0.90199 | acc = 0.97377 || loss = 0.22309 | jacc = 0.77804 | acc = 0.94439
poch 489 took 82+25 sec. loss = 0.10600 | jacc = 0.89957 | acc = 0.97446 || loss = 0.23979 | jacc = 0.76619 | acc = 0.94078
poch 490 took 82+25 sec. loss = 0.10690 | jacc = 0.89972 | acc = 0.97417 || loss = 0.23890 | jacc = 0.75983 | acc = 0.94077
poch 491 took 82+25 sec. loss = 0.10616 | jacc = 0.89949 | acc = 0.97446 || loss = 0.24208 | jacc = 0.77011 | acc = 0.94171
poch 492 took 85+25 sec. loss = 0.10682 | jacc = 0.90047 | acc = 0.97397 || loss = 0.25706 | jacc = 0.75976 | acc = 0.93777
poch 493 took 81+26 sec. loss = 0.10655 | jacc = 0.89992 | acc = 0.97407 || loss = 0.24872 | jacc = 0.76031 | acc = 0.93913
poch 494 took 81+26 sec. loss = 0.10491 | jacc = 0.90199 | acc = 0.97475 || loss = 0.25842 | jacc = 0.76532 | acc = 0.93852
poch 495 took 82+25 sec. loss = 0.10377 | jacc = 0.89982 | acc = 0.97527 || loss = 0.24542 | jacc = 0.75673 | acc = 0.93997
poch 496 took 82+25 sec. loss = 0.10506 | jacc = 0.90250 | acc = 0.97457 || loss = 0.24679 | jacc = 0.76496 | acc = 0.94126
poch 497 took 82+25 sec. loss = 0.10735 | jacc = 0.89933 | acc = 0.97367 || loss = 0.26098 | jacc = 0.76634 | acc = 0.93604
poch 498 took 81+25 sec. loss = 0.10559 | jacc = 0.89858 | acc = 0.97449 || loss = 0.26146 | jacc = 0.76774 | acc = 0.93409
poch 499 took 82+25 sec. loss = 0.10469 | jacc = 0.90138 | acc = 0.97465 || loss = 0.24331 | jacc = 0.77963 | acc = 0.93763
poch 500 took 82+25 sec. loss = 0.10356 | jacc = 0.89884 | acc = 0.97524 || loss = 0.24491 | jacc = 0.76169 | acc = 0.93914
poch 501 took 82+25 sec. loss = 0.10546 | jacc = 0.89894 | acc = 0.97446 || loss = 0.24925 | jacc = 0.76253 | acc = 0.93860
poch 502 took 82+25 sec. loss = 0.10660 | jacc = 0.89638 | acc = 0.97408 || loss = 0.28833 | jacc = 0.73400 | acc = 0.93093
poch 503 took 82+25 sec. loss = 0.10394 | jacc = 0.89895 | acc = 0.97503 || loss = 0.24790 | jacc = 0.77866 | acc = 0.93880
poch 504 took 82+25 sec. loss = 0.10699 | jacc = 0.90258 | acc = 0.97401 || loss = 0.23594 | jacc = 0.77132 | acc = 0.94353
poch 505 took 85+25 sec. loss = 0.10610 | jacc = 0.89921 | acc = 0.97428 || loss = 0.25685 | jacc = 0.74812 | acc = 0.93534
poch 506 took 82+25 sec. loss = 0.10404 | jacc = 0.90237 | acc = 0.97492 || loss = 0.25267 | jacc = 0.74799 | acc = 0.93850
poch 507 took 82+25 sec. loss = 0.10817 | jacc = 0.90184 | acc = 0.97369 || loss = 0.22824 | jacc = 0.77676 | acc = 0.94359
poch 508 took 82+25 sec. loss = 0.10571 | jacc = 0.90109 | acc = 0.97429 || loss = 0.26493 | jacc = 0.76246 | acc = 0.93700
poch 509 took 83+25 sec. loss = 0.10522 | jacc = 0.89821 | acc = 0.97449 || loss = 0.24795 | jacc = 0.76297 | acc = 0.93944
poch 510 took 83+25 sec. loss = 0.10386 | jacc = 0.90406 | acc = 0.97482 || loss = 0.23007 | jacc = 0.77989 | acc = 0.94419
poch 511 took 82+25 sec. loss = 0.10477 | jacc = 0.90102 | acc = 0.97440 || loss = 0.23268 | jacc = 0.77491 | acc = 0.94237
poch 512 took 82+25 sec. loss = 0.10446 | jacc = 0.90323 | acc = 0.97469 || loss = 0.23026 | jacc = 0.77756 | acc = 0.94376
poch 513 took 82+25 sec. loss = 0.10586 | jacc = 0.90302 | acc = 0.97414 || loss = 0.23188 | jacc = 0.76283 | acc = 0.94278
poch 514 took 82+25 sec. loss = 0.10296 | jacc = 0.89971 | acc = 0.97510 || loss = 0.24981 | jacc = 0.74167 | acc = 0.94044
poch 515 took 82+27 sec. loss = 0.10407 | jacc = 0.90185 | acc = 0.97474 || loss = 0.22517 | jacc = 0.78083 | acc = 0.94520
poch 516 took 91+27 sec. loss = 0.10373 | jacc = 0.90165 | acc = 0.97509 || loss = 0.23864 | jacc = 0.76183 | acc = 0.94357
poch 517 took 88+27 sec. loss = 0.10450 | jacc = 0.90003 | acc = 0.97458 || loss = 0.22222 | jacc = 0.77901 | acc = 0.94670
poch 518 took 88+27 sec. loss = 0.10310 | jacc = 0.90104 | acc = 0.97497 || loss = 0.24351 | jacc = 0.76136 | acc = 0.94255
poch 519 took 88+27 sec. loss = 0.10303 | jacc = 0.90234 | acc = 0.97494 || loss = 0.23871 | jacc = 0.75701 | acc = 0.94291
poch 520 took 88+27 sec. loss = 0.10245 | jacc = 0.90210 | acc = 0.97523 || loss = 0.24773 | jacc = 0.76166 | acc = 0.94338
poch 521 took 88+27 sec. loss = 0.10352 | jacc = 0.90255 | acc = 0.97475 || loss = 0.21457 | jacc = 0.78102 | acc = 0.94919
poch 522 took 88+27 sec. loss = 0.10017 | jacc = 0.90321 | acc = 0.97592 || loss = 0.22458 | jacc = 0.76841 | acc = 0.94672
poch 523 took 88+27 sec. loss = 0.10379 | jacc = 0.90225 | acc = 0.97466 || loss = 0.23441 | jacc = 0.77610 | acc = 0.94520
poch 524 took 89+27 sec. loss = 0.10383 | jacc = 0.90050 | acc = 0.97477 || loss = 0.28286 | jacc = 0.72969 | acc = 0.93612
poch 525 took 88+27 sec. loss = 0.10180 | jacc = 0.90202 | acc = 0.97546 || loss = 0.21285 | jacc = 0.78490 | acc = 0.94874
poch 526 took 88+27 sec. loss = 0.10287 | jacc = 0.90288 | acc = 0.97488 || loss = 0.22418 | jacc = 0.77670 | acc = 0.94533
poch 527 took 88+27 sec. loss = 0.10149 | jacc = 0.90330 | acc = 0.97547 || loss = 0.21352 | jacc = 0.79449 | acc = 0.94696 (BEST)
Epoch 527 : [test : 100%]Average cost test = 0.44401 | jacc test = 0.66204 | acc_test = 0.89867

So I tried to lower the lr to 1e-9 and it was improving really slowly so I killed it.

poch 0 took 88+27 sec. loss = 0.10183 | jacc = 0.90184 | acc = 0.97515 || loss = 0.21352 | jacc = 0.79449 | acc = 0.94697 (BEST)
Epoch 0 : [test : 100%]Average cost test = 0.44401 | jacc test = 0.66205 | acc_test = 0.89867 
poch 1 took 88+27 sec. loss = 0.10246 | jacc = 0.90308 | acc = 0.97509 || loss = 0.21351 | jacc = 0.79449 | acc = 0.94697 (BEST)
Epoch 1 : [test : 100%]Average cost test = 0.44400 | jacc test = 0.66205 | acc_test = 0.89867 
poch 2 took 92+27 sec. loss = 0.10220 | jacc = 0.90131 | acc = 0.97513 || loss = 0.21351 | jacc = 0.79449 | acc = 0.94697 (BEST)
Epoch 2 : [test : 100%]Average cost test = 0.44400 | jacc test = 0.66205 | acc_test = 0.89867 
poch 3 took 88+27 sec. loss = 0.10415 | jacc = 0.90096 | acc = 0.97444 || loss = 0.21351 | jacc = 0.79449 | acc = 0.94697 (BEST)
Epoch 3 : [test : 100%]Average cost test = 0.44400 | jacc test = 0.66205 | acc_test = 0.89867 
poch 4 took 88+27 sec. loss = 0.10159 | jacc = 0.90278 | acc = 0.97525 || loss = 0.21351 | jacc = 0.79449 | acc = 0.94697 (BEST)
Epoch 4 : [test : 100%]Average cost test = 0.44399 | jacc test = 0.66206 | acc_test = 0.89867 
poch 5 took 88+27 sec. loss = 0.10237 | jacc = 0.90362 | acc = 0.97522 || loss = 0.21351 | jacc = 0.79449 | acc = 0.94697
poch 6 took 88+27 sec. loss = 0.10132 | jacc = 0.90428 | acc = 0.97544 || loss = 0.21351 | jacc = 0.79449 | acc = 0.94697
poch 7 took 88+27 sec. loss = 0.10195 | jacc = 0.89980 | acc = 0.97518 || loss = 0.21351 | jacc = 0.79449 | acc = 0.94697
poch 8 took 89+27 sec. loss = 0.10228 | jacc = 0.90374 | acc = 0.97505 || loss = 0.21351 | jacc = 0.79449 | acc = 0.94697
poch 9 took 89+27 sec. loss = 0.10233 | jacc = 0.90323 | acc = 0.97508 || loss = 0.21350 | jacc = 0.79449 | acc = 0.94697
poch 10 took 88+27 sec. loss = 0.10304 | jacc = 0.90112 | acc = 0.97472 || loss = 0.21350 | jacc = 0.79449 | acc = 0.94697
poch 11 took 88+27 sec. loss = 0.10277 | jacc = 0.90206 | acc = 0.97488 || loss = 0.21350 | jacc = 0.79449 | acc = 0.94697
poch 12 took 92+27 sec. loss = 0.10709 | jacc = 0.90022 | acc = 0.97402 || loss = 0.21350 | jacc = 0.79449 | acc = 0.94697
poch 13 took 88+27 sec. loss = 0.10120 | jacc = 0.90164 | acc = 0.97556 || loss = 0.21349 | jacc = 0.79449 | acc = 0.94697
poch 14 took 88+27 sec. loss = 0.10340 | jacc = 0.90420 | acc = 0.97491 || loss = 0.21349 | jacc = 0.79449 | acc = 0.94697
poch 15 took 88+27 sec. loss = 0.10417 | jacc = 0.90222 | acc = 0.97441 || loss = 0.21349 | jacc = 0.79449 | acc = 0.94697
poch 16 took 88+27 sec. loss = 0.10172 | jacc = 0.90342 | acc = 0.97539 || loss = 0.21349 | jacc = 0.79449 | acc = 0.94697
poch 17 took 88+27 sec. loss = 0.10237 | jacc = 0.90243 | acc = 0.97505 || loss = 0.21349 | jacc = 0.79449 | acc = 0.94697 (BEST)
Epoch 17 : [test : 100%]Average cost test = 0.44395 | jacc test = 0.66209 | acc_test = 0.89868 
poch 18 took 88+27 sec. loss = 0.10304 | jacc = 0.90404 | acc = 0.97501 || loss = 0.21349 | jacc = 0.79449 | acc = 0.94697 (BEST)
Epoch 18 : [test : 100%]Average cost test = 0.44395 | jacc test = 0.66209 | acc_test = 0.89869

(Brendan Fortuner) #43

Fine-tuning stopped early (using max patience of 50 based on validation loss). Accuracy on the test set did not improve. But perhaps I should train longer and not use the validation set to early stop? The training set error does look like it was coming down… Perhaps I can also exclude the random horizontal flipping during fine-tuning…

Metrics - Epoch 970 (Loss, Error)
Validation – 0.178, 7.17
Train – 0.183, 8.90
Test – 0.441, 13.4

What are good next steps? I was going to try running it on Pascal VOC or MSCOCO.


(Jeremy Howard) #44

Pascal VOC and MSCOCO would be interesting!


(Brendan Fortuner) #45

Can you loan me some GPUs :wink:


(Jeremy Howard) #46

Win some!


(Even Oldridge) #47

Facebook has a recent blog post where they’ve open sourced some of their research on segmentation, treating it as a pixel level classification problem. It’s not exactly what you’re looking for, but it might stimulate some ideas by digging around their source code.


(Jeremy Howard) #48

@brendan I figured out why we’re not replicating the paper’s accuracy. It’s because they remove the ‘void’ category from their accuracy measurement. Once I remove that, I can get around 89%. Still not quite as good as the paper, but as good as @kelvin was getting.


(Brendan Fortuner) #49

Interesting! I think the real test is to benchmark performance on MSCoCo or Pascal.


(Erdem Basegmez) #50

What does +80 refer to next to DB in the photo taken from Lecture 14? Does DenseBlock add 80 filters each time?


(Constantin) #51

Looks to me like the number of filters you add with the respective densebox. TD is for transition, so no new filters there. How many filters are added is controlled through the growth factor parameter.


(Brendan Fortuner) #52

Yep! If the growth factor is 16, each dense layer appends 16 filters to the volume it receives and passes it on to the next layer. In the example above, each Dense Block has 4 dense layers.


(Erdem Basegmez) #53

@brendan but how does that make 80 then?


(Brendan Fortuner) #54

Apologies! In that particular example each Dense Block had 5 Dense Layers. But the # of dense layers per block is a parameter you can play with–4, 5, 10, etc.


(Erdem Basegmez) #55

@brendan thank you for the clarification :blush:


#56

@brendan Hi, is there any new progress now? I am running your pytorch-tiramisu, I would like to know do you replicate it successfully, thank you!


(Brendan Fortuner) #57

Haven’t touched it since class. I was not able to replicate the authors result although we got close. Jeremy suggested it may have to do with including/ excluding the background class.


(Akshay) #58

Hello Guys …

I was Trying to Convert this Tiramisu Model into a Protobuff File to run this Model on Android …

def export_model(saver, model, input_node_names, output_node_name):
    MODEL_NAME='tiramisu'
    tf.train.write_graph(K.get_session().graph_def, 'out', \
        MODEL_NAME + '_graph.pbtxt')

    saver.save(K.get_session(), 'out/' + MODEL_NAME + '.chkp')

    freeze_graph.freeze_graph('out/' + MODEL_NAME + '_graph.pbtxt', None, \
        False, 'out/' + MODEL_NAME + '.chkp', output_node_name, \
        "save/restore_all", "save/Const:0", \
        'out/frozen_' + MODEL_NAME + '.pb', True, "")

    input_graph_def = tf.GraphDef()
    with tf.gfile.Open('out/frozen_' + MODEL_NAME + '.pb', "rb") as f:
        input_graph_def.ParseFromString(f.read())

    output_graph_def = optimize_for_inference_lib.optimize_for_inference(
            input_graph_def, input_node_names, [output_node_name],
            tf.float32.as_datatype_enum)

    with tf.gfile.FastGFile('out/opt_' + MODEL_NAME + '.pb', "wb") as f:
        f.write(output_graph_def.SerializeToString())

    print("graph saved!")

Tried Calling this function >>

export_model(tf.train.Saver(), model, [model.input.name], model.output.name)

It gives the below error >>
INFO:tensorflow:Restoring parameters from out/tiramisu.chkp

---------------------------------------------------------------------------
AssertionError                            Traceback (most recent call last)
<ipython-input-80-ce4abb242ea5> in <module>()
----> 1 export_model(tf.train.Saver(), model, [model.input.name], model.output.name)

<ipython-input-71-81c8aebbb95e> in export_model(saver, model, input_node_names, output_node_name)
      5     saver.save(K.get_session(), 'out/' + MODEL_NAME + '.chkp')
      6 
----> 7     freeze_graph.freeze_graph('out/' + MODEL_NAME + '_graph.pbtxt', None,         False, 'out/' + MODEL_NAME + '.chkp', output_node_name,         "save/restore_all", "save/Const:0",         'out/frozen_' + MODEL_NAME + '.pb', True, "")
      8 
      9     input_graph_def = tf.GraphDef()

/usr/local/lib/python3.5/dist-packages/tensorflow/python/tools/freeze_graph.py in freeze_graph(input_graph, input_saver, input_binary, input_checkpoint, output_node_names, restore_op_name, filename_tensor_name, output_graph, clear_devices, initializer_nodes, variable_names_blacklist)
    177       clear_devices,
    178       initializer_nodes,
--> 179       variable_names_blacklist)
    180 
    181 

/usr/local/lib/python3.5/dist-packages/tensorflow/python/tools/freeze_graph.py in freeze_graph_with_def_protos(***failed resolving arguments***)
    114         input_graph_def,
    115         output_node_names.split(","),
--> 116         variable_names_blacklist=variable_names_blacklist)
    117 
    118   with gfile.GFile(output_graph, "wb") as f:

/usr/local/lib/python3.5/dist-packages/tensorflow/python/framework/graph_util_impl.py in convert_variables_to_constants(sess, input_graph_def, output_node_names, variable_names_whitelist, variable_names_blacklist)
    202   # This graph only includes the nodes needed to evaluate the output nodes, and
    203   # removes unneeded nodes like those involved in saving and assignment.
--> 204   inference_graph = extract_sub_graph(input_graph_def, output_node_names)
    205 
    206   found_variables = {}

/usr/local/lib/python3.5/dist-packages/tensorflow/python/framework/graph_util_impl.py in extract_sub_graph(graph_def, dest_nodes)
    139 
    140   for d in dest_nodes:
--> 141     assert d in name_to_node_map, "%s is not in graph" % d
    142 
    143   nodes_to_keep = set()

AssertionError: truediv:0 is not in graph

Below is the Model.summary ()

https://pastebin.com/72gUrRr0

Can Some one Suggest…


Implementing Mask R-CNN
(Ben) #59

Hey, guys

When I try to replicate Jeremy’s results, I keep get these at the beginning of training. Is that normal?

Epoch 1/100
232s - loss: nan - acc: 2.0441e-06 - val_loss: nan - val_acc: 2.9938e-04
Epoch 2/100
232s - loss: nan - acc: 2.4274e-06 - val_loss: nan - val_acc: 2.9938e-04


(Ben) #61

Uh, figured out it is class unmatch problem