Early stopping frequency not honored

[mikegerber]

I am using Calamari 1.0.1:

calamari-cross-fold-train \
  --files \
  "$TMPDIR/*/*/*.png" \
  --best_models_dir $outdir \
  --early_stopping_frequency=0.25 \
  --early_stopping_nbest=5 \
  --batch_size=128 \
  --n_folds=5 \
  --max_parallel_models=1 \
  --display=0.01 \
  2>&1 | tee $outdir/train.`date -Iminutes`.log

I would expect early stopping to be checked every 0.25 epochs, however it actually happens roughly every 0.05 epochs now. This was not the case with Calamari 0.3.5.

Log excerpt:

Early stopping: 100%|██████████| 490/490 [03:01<00:00,  2.69it/s]
FOLD 0 | Found better model with accuracy of 90.495146%
FOLD 0 | Storing checkpoint to '/home/mike.gerber/devel/qurator-mono-repo/calamari-models/train-calamari-gt4histocr/data/calamari-models/GT4HistOCR/0.ckpt'
FOLD 0 | #0.320388: loss=16.61115993 ler=0.43991868 dt=3.12265495s
FOLD 0 |   PRED: '‪Zweck das Konfirmationsbiüchlein wieder einmal zur‬'
FOLD 0 |   TRUE: '‪Zweck das Konfirmationsbüchlein wieder einmal zur‬'
FOLD 0 | #0.330097: loss=15.82271417 ler=0.42927558 dt=3.11410474s
FOLD 0 |   PRED: '‪Die lUfer faſen den Jubel der Ströme nicht,‬'
FOLD 0 |   TRUE: '‪Die Ufer faſſen den Jubel der Stroͤme nicht,‬'
FOLD 0 | #0.339806: loss=15.27738711 ler=0.41900858 dt=3.09278003s
FOLD 0 |   PRED: '‪voruͤber, Studenten in Schnͤrrbͤcken und ſilber—‬'
FOLD 0 |   TRUE: '‪voruͤber, Studenten in Schnuͤrroͤcken und ſilber—‬'
FOLD 0 | #0.349515: loss=14.70154876 ler=0.40963159 dt=3.09945018s
FOLD 0 |   PRED: '‪nur l bis 1 Perſonen intereſſnren, grade dieſen‬'
FOLD 0 |   TRUE: '‪nur 50 bis 100 Perſonen intereſſiren, grade dieſen‬'
FOLD 0 | #0.359223: loss=14.19648499 ler=0.40086900 dt=3.10238453s
FOLD 0 |   PRED: '‪immer verging. Wenn Jhr es hören wollt, ſo will ich‬'
FOLD 0 |   TRUE: '‪immer verging. Wenn Jhr es hören wollt, ſo will ich‬'
FOLD 0 | #0.368932: loss=13.71040510 ler=0.39230464 dt=1.28325271s
FOLD 0 |   PRED: '‪hend die noch mehr fͤr mich gerhan, vielleicht‬'
FOLD 0 |   TRUE: '‪hend) die noch mehr fuͤr mich gethan, vielleicht‬'
FOLD 0 | Storing checkpoint to '/tmp/calamaribnc4tmc8/fold_0/model_00000732.ckpt'
FOLD 0 | Checking early stopping model
Early stopping: 100%|██████████| 490/490 [03:03<00:00,  2.67it/s]
FOLD 0 | Found better model with accuracy of 92.196010%
FOLD 0 | Storing checkpoint to '/home/mike.gerber/devel/qurator-mono-repo/calamari-models/train-calamari-gt4histocr/data/calamari-models/GT4HistOCR/0.ckpt'
FOLD 0 | #0.378641: loss=13.24382450 ler=0.38367690 dt=3.12489480s
FOLD 0 |   PRED: '‪uns einer Wirklichkeit bewußt werden, in ihrem eigenen‬'
FOLD 0 |   TRUE: '‪uns einer Wirklichkeit bewußt werden, in ihrem eigenen‬'
FOLD 0 | #0.388350: loss=12.86139683 ler=0.37567462 dt=3.13790299s
FOLD 0 |   PRED: '‪endlich vorüber, und er mußte wieder nah‬'
FOLD 0 |   TRUE: '‪endlich vorüber, und er mußte wieder nach‬'
FOLD 0 | #0.398058: loss=12.63035151 ler=0.36839195 dt=3.12761085s
FOLD 0 |   PRED: '‪der Schud des Böſen iſt das Bewußtſein, mit der poſitien‬'
FOLD 0 |   TRUE: '‪der Schuld des Böſen iſt das Bewußtſein, mit der poſitiven‬'
FOLD 0 | #0.407767: loss=12.44438200 ler=0.36114099 dt=3.11889884s
FOLD 0 |   PRED: '‪on dannen kamen wir ym die kyrch zů den heylgen engeln genant‬'
FOLD 0 |   TRUE: '‪¶ Von dannen kamen wir yn die kyrch zů den heyligen engeln genant ·‬'
FOLD 0 | #0.417476: loss=12.18084080 ler=0.35440814 dt=3.10386337s
FOLD 0 |   PRED: '‪Sie ſchadet ſonſt der Geſundheit.‬'
FOLD 0 |   TRUE: '‪Sie ſchadet ſonſt der Geſundheit.‬'
FOLD 0 | #0.427184: loss=11.85094854 ler=0.34774322 dt=1.25803710s
FOLD 0 |   PRED: '‪tiſſin iſt meiner, eines grauſamen Todes ge⸗‬'
FOLD 0 |   TRUE: '‪tiſſin iſt meiner, eines grauſamen Todes ge—‬'
FOLD 0 | Storing checkpoint to '/tmp/calamaribnc4tmc8/fold_0/model_00000854.ckpt'
FOLD 0 | Checking early stopping model

[mikegerber]

I suspect the problem to be here: https://github.com/Calamari-OCR/calamari/blob/master/calamari_ocr/ocr/backends/tensorflow_backend/tensorflow_model.py#L320 – Calculating of the callback parameter steps_per_epoch is based on the validation set and I think that should be just the same value as 2 lines above. But please review for yourself, I might not understand the code correctly.

[ChWick]

You are absolutely right, I already fixed this 5 days ago but forgot to push it (4837402).