| { | |
| "architecture": "DenseNet121-Transformer", | |
| "model_type": "custom", | |
| "task": "handwritten-text-recognition", | |
| "language": "Arabic", | |
| "language_code": "ar", | |
| "script": "Arabic", | |
| "dataset": "KHATT", | |
| "cnn_backbone": "densenet121", | |
| "use_upsample": true, | |
| "hidden_size": 256, | |
| "num_encoder_layers": 3, | |
| "num_decoder_layers": 6, | |
| "num_attention_heads": 8, | |
| "feed_forward_dim": 2048, | |
| "dropout_pretrain": 0.3, | |
| "dropout_finetune": 0.2, | |
| "vocab_size": 143, | |
| "max_sequence_length": 555, | |
| "image_height": 600, | |
| "image_width": 1235, | |
| "total_parameters": 22760778, | |
| "training": { | |
| "best_epoch": 65, | |
| "best_val_cer": 0.11814673883916608, | |
| "best_val_loss": null, | |
| "pretrain_optimizer": "AdamW", | |
| "pretrain_lr": 0.0001, | |
| "finetune_optimizer": "AdamW", | |
| "finetune_lr": 5e-05, | |
| "pretrain_scheduler": "StepLR (step=15, gamma=0.5)", | |
| "finetune_scheduler": "ReduceLROnPlateau (patience=5, factor=0.5)", | |
| "pretrain_batch_size": 16, | |
| "finetune_batch_size": 16, | |
| "pretrain_epochs": 80, | |
| "finetune_epochs": 80, | |
| "curriculum_learning": true, | |
| "teacher_forcing_noise_pretrain": 0.15, | |
| "teacher_forcing_noise_finetune": 0.05, | |
| "encoder_freeze_epochs": 10, | |
| "encoder_lr_multiplier": 0.1 | |
| }, | |
| "performance": { | |
| "test_cer": 0.1394, | |
| "test_wer": 0.5075, | |
| "test_cer_with_lm": null | |
| } | |
| } |