layoutlmv3-base-finetuned-rvlcdip

layoutlmv3-base-finetuned-rvlcdip

This model is a fine-tuned version of microsoft/layoutlmv3-base on the RVL-CDIP dataset processed using Amazon OCR. The following metrics were computed on the evaluation set after the final optimization step:

• Evaluation Loss: 0.1856316477060318
• Evaluation Accuracy: 0.9519237980949524
• Evaluation Weighted F1: 0.9518911690649716
• Evaluation Micro F1: 0.9519237980949524
• Evaluation Macro F1: 0.9518042570370386
• Evaluation Weighted Recall: 0.9519237980949524
• Evaluation Micro Recall: 0.9519237980949524
• Evaluation Macro Recall: 0.9518171728908463
• Evaluation Weighted Precision: 0.9519094862975979
• Evaluation Micro Precision: 0.9519237980949524
• Evaluation Macro Precision: 0.9518423447239385
• Evaluation Runtime (seconds): 514.7031
• Evaluation Samples per Second: 77.713
• Evaluation Steps per Second: 1.214

Training logs

See wandb report: https://api.wandb.ai/links/gordon-lim/lokqu7ok

Training arguments

The following arguments were provided to Trainer:

• Output Directory: ./results
• Maximum Steps: 20000
• Per Device Train Batch Size: 32 (due to CUDA memory constraints; paper uses 64, trained using 2 GPUs so 32 * 2 effective batch size)
• Per Device Evaluation Batch Size: 32 (due to CUDA memory constraints)
• Warmup Steps: 0 (not specified in paper, but warmup ratio is used for DocVQA, hence assumed default)
• Weight Decay: 0 (not specified in paper for RVL-CDIP, but 0.05 for PubLayNet, hence assumed default)
• Evaluation Strategy: steps
• Evaluation Steps: 1000
• Evaluate on Start: True
• Save Strategy: steps
• Save Steps: 1000
• Save Total Limit: 5
• Learning Rate: 2e-5
• Load Best Model at End: True
• Metric for Best Model: accuracy
• Greater is Better: True
• Report to: wandb (log to Weights & Biases)
• Logging Steps: 1000
• Logging First Step: True
• Learning Rate Scheduler Type: cosine (not mentioned in paper, but PubLayNet GitHub example uses 'cosine')
• FP16: True (due to CUDA memory constraints)
• Dataloader Number of Workers: 4 (number of subprocesses to use for data loading)
• DDP Find Unused Parameters: True

Framework versions

• Transformers 4.42.3
• Pytorch 2.2.0+cu121
• Datasets 2.14.0
• Tokenizers 0.19.1