de-htr-web / results /m2 /seq2seq_reference.json
naeyn's picture
Publish DE·HTR browser release
e1d7189 verified
Raw
History Blame Contribute Delete
4.69 kB
{
"model": "microsoft/trocr-small-handwritten",
"date": "2026-07-29",
"architecture": "encoder-decoder",
"role": "accuracy_reference_not_for_release",
"seed": 42,
"notes": "Encoder-decoder accuracy reference for the CER-gap comparison; CTC remains the release artifact.",
"hf_revision": "b4648cfa171985a6745f37ddd637e98c0da958ac",
"parameter_count": 61596672,
"processor_load_seconds": 3.2194992679997085,
"model_load_seconds": 1.1561599249998835,
"train_seconds": 681.4541459979973,
"best_epoch": 12,
"best_val_cer": 0.11095395253811095,
"test_cer": 0.11507950752001149,
"test_wer": 0.3416644509438979,
"test_cer_ci95": [
0.10609754239875374,
0.12452546445905048
],
"test_wer_ci95": [
0.3234093074471375,
0.3607701922432293
],
"n_test_lines": 601,
"checkpoint": "checkpoints/m2/trocr-small-reference/best",
"config": {
"model_id": "microsoft/trocr-small-handwritten",
"architecture": "encoder-decoder",
"role": "accuracy_reference_not_for_release",
"role_ref": "encoder-decoder CER gap",
"split_manifest": "data/split_manifest.json",
"seed": 42,
"epochs": 12,
"batch_size": 12,
"workers": 8,
"lr": 5e-05,
"patience": 3,
"max_label_length": 128,
"grad_clip_norm": 1.0,
"optimizer": "AdamW",
"scheduler": "CosineAnnealingLR",
"amp": {
"enabled": true,
"dtype": "bfloat16"
},
"train_data": "real_only",
"synthetic_ratio": 0.0,
"early_stopped": false,
"resumed": false
},
"generation": {
"max_new_tokens": 128,
"num_beams": 1,
"do_sample": false,
"decoding": "deterministic_greedy"
},
"environment": {
"python": "3.12.3",
"platform": "Linux-6.8.0-134-generic-x86_64-with-glibc2.39",
"torch": "2.11.0+cu128",
"transformers": "5.14.1",
"device": "cuda",
"cuda_available": true,
"cudnn_enabled": true,
"cudnn_version": 91900,
"cuda_device_name": "NVIDIA RTX A5000",
"cuda_device_count": 1
},
"epoch_history": [
{
"epoch": 1,
"train_loss": 4.4270035982131954,
"val_cer": 0.26995913877101996,
"val_wer": 0.5919075144508671,
"lr": 4.914814565722671e-05,
"epoch_seconds": 60.28538311400007
},
{
"epoch": 2,
"train_loss": 1.558124960385836,
"val_cer": 0.21004243281471005,
"val_wer": 0.4979768786127168,
"lr": 4.665063509461098e-05,
"epoch_seconds": 57.15164858999924
},
{
"epoch": 3,
"train_loss": 1.1930473674566318,
"val_cer": 0.20524909633820526,
"val_wer": 0.49942196531791905,
"lr": 4.267766952966369e-05,
"epoch_seconds": 55.6579075919999
},
{
"epoch": 4,
"train_loss": 0.9639835356137691,
"val_cer": 0.15833726229765835,
"val_wer": 0.4184971098265896,
"lr": 3.7500000000000003e-05,
"epoch_seconds": 57.48923320799986
},
{
"epoch": 5,
"train_loss": 0.7886032088444783,
"val_cer": 0.15047933364765048,
"val_wer": 0.4028901734104046,
"lr": 3.1470476127563024e-05,
"epoch_seconds": 57.11078792199987
},
{
"epoch": 6,
"train_loss": 0.614951229095459,
"val_cer": 0.14112839855414114,
"val_wer": 0.38497109826589593,
"lr": 2.5000000000000005e-05,
"epoch_seconds": 57.38486632099921
},
{
"epoch": 7,
"train_loss": 0.48801431422814345,
"val_cer": 0.1395961024673896,
"val_wer": 0.3829479768786127,
"lr": 1.852952387243699e-05,
"epoch_seconds": 57.561237477999384
},
{
"epoch": 8,
"train_loss": 0.3661836621088859,
"val_cer": 0.12831997485462832,
"val_wer": 0.353757225433526,
"lr": 1.2500000000000009e-05,
"epoch_seconds": 54.417468428000575
},
{
"epoch": 9,
"train_loss": 0.2723297044443779,
"val_cer": 0.12006914977212006,
"val_wer": 0.3242774566473988,
"lr": 7.322330470336316e-06,
"epoch_seconds": 54.96981200200025
},
{
"epoch": 10,
"train_loss": 0.2096752209541125,
"val_cer": 0.1131934622033632,
"val_wer": 0.319364161849711,
"lr": 3.3493649053890335e-06,
"epoch_seconds": 56.168583769998804
},
{
"epoch": 11,
"train_loss": 0.1640623667683357,
"val_cer": 0.11287914505736288,
"val_wer": 0.3196531791907514,
"lr": 8.518543427732954e-07,
"epoch_seconds": 57.256369542001266
},
{
"epoch": 12,
"train_loss": 0.1423123875967203,
"val_cer": 0.11095395253811095,
"val_wer": 0.3173410404624277,
"lr": 0.0,
"epoch_seconds": 56.000848030998895
}
]
}