layoutlmv3-finetuned-wildreceipt
This model is a fine-tuned version of microsoft/layoutlmv3-base on the wild_receipt dataset. It achieves the following results on the evaluation set:
- Loss: 0.3108
- Precision: 0.8772
- Recall: 0.8799
- F1: 0.8785
- Accuracy: 0.9249
Model description
More information needed
Intended uses & limitations
More information needed
Training and evaluation data
The WildReceipt dataset consists of 1740 receipt images, and contains 25 key information categories, and a total of about 69000 text boxes. 1268 and 472 images are used for training and testing respectively to train the LayoutLMv3 model for Key Information Extraction.
Training procedure
The training code: https://github.com/Theivaprakasham/layoutlmv3/blob/main/training_codes/LayoutLMv3_training_WildReceipts_dataset.ipynb
Training hyperparameters
The following hyperparameters were used during training:
- learning_rate: 1e-05
- train_batch_size: 4
- eval_batch_size: 4
- seed: 42
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- training_steps: 4000
Training results
Training Loss | Epoch | Step | Validation Loss | Precision | Recall | F1 | Accuracy |
---|---|---|---|---|---|---|---|
No log | 0.32 | 100 | 1.3143 | 0.6709 | 0.2679 | 0.3829 | 0.6700 |
No log | 0.63 | 200 | 0.8814 | 0.6478 | 0.5195 | 0.5766 | 0.7786 |
No log | 0.95 | 300 | 0.6568 | 0.7205 | 0.6491 | 0.6829 | 0.8303 |
No log | 1.26 | 400 | 0.5618 | 0.7544 | 0.7072 | 0.7300 | 0.8519 |
1.0284 | 1.58 | 500 | 0.5003 | 0.7802 | 0.7566 | 0.7682 | 0.8687 |
1.0284 | 1.89 | 600 | 0.4454 | 0.7941 | 0.7679 | 0.7807 | 0.8748 |
1.0284 | 2.21 | 700 | 0.4314 | 0.8142 | 0.7928 | 0.8033 | 0.8852 |
1.0284 | 2.52 | 800 | 0.3870 | 0.8172 | 0.8200 | 0.8186 | 0.8953 |
1.0284 | 2.84 | 900 | 0.3629 | 0.8288 | 0.8369 | 0.8329 | 0.9025 |
0.4167 | 3.15 | 1000 | 0.3537 | 0.8540 | 0.8200 | 0.8366 | 0.9052 |
0.4167 | 3.47 | 1100 | 0.3383 | 0.8438 | 0.8285 | 0.8361 | 0.9063 |
0.4167 | 3.79 | 1200 | 0.3403 | 0.8297 | 0.8493 | 0.8394 | 0.9062 |
0.4167 | 4.1 | 1300 | 0.3271 | 0.8428 | 0.8545 | 0.8487 | 0.9110 |
0.4167 | 4.42 | 1400 | 0.3182 | 0.8491 | 0.8518 | 0.8504 | 0.9131 |
0.2766 | 4.73 | 1500 | 0.3111 | 0.8491 | 0.8539 | 0.8515 | 0.9129 |
0.2766 | 5.05 | 1600 | 0.3177 | 0.8397 | 0.8620 | 0.8507 | 0.9124 |
0.2766 | 5.36 | 1700 | 0.3091 | 0.8676 | 0.8548 | 0.8612 | 0.9191 |
0.2766 | 5.68 | 1800 | 0.3080 | 0.8508 | 0.8645 | 0.8576 | 0.9162 |
0.2766 | 5.99 | 1900 | 0.3059 | 0.8492 | 0.8662 | 0.8576 | 0.9163 |
0.2114 | 6.31 | 2000 | 0.3184 | 0.8536 | 0.8657 | 0.8596 | 0.9147 |
0.2114 | 6.62 | 2100 | 0.3161 | 0.8583 | 0.8713 | 0.8648 | 0.9184 |
0.2114 | 6.94 | 2200 | 0.3055 | 0.8707 | 0.8682 | 0.8694 | 0.9220 |
0.2114 | 7.26 | 2300 | 0.3004 | 0.8689 | 0.8745 | 0.8717 | 0.9219 |
0.2114 | 7.57 | 2400 | 0.3111 | 0.8701 | 0.8720 | 0.8711 | 0.9211 |
0.174 | 7.89 | 2500 | 0.3130 | 0.8599 | 0.8741 | 0.8669 | 0.9198 |
0.174 | 8.2 | 2600 | 0.3034 | 0.8661 | 0.8748 | 0.8704 | 0.9219 |
0.174 | 8.52 | 2700 | 0.3005 | 0.8799 | 0.8673 | 0.8736 | 0.9225 |
0.174 | 8.83 | 2800 | 0.3043 | 0.8687 | 0.8804 | 0.8745 | 0.9240 |
0.174 | 9.15 | 2900 | 0.3121 | 0.8776 | 0.8704 | 0.8740 | 0.9242 |
0.1412 | 9.46 | 3000 | 0.3131 | 0.8631 | 0.8755 | 0.8692 | 0.9204 |
0.1412 | 9.78 | 3100 | 0.3067 | 0.8715 | 0.8773 | 0.8744 | 0.9233 |
0.1412 | 10.09 | 3200 | 0.3021 | 0.8751 | 0.8812 | 0.8782 | 0.9248 |
0.1412 | 10.41 | 3300 | 0.3092 | 0.8651 | 0.8808 | 0.8729 | 0.9228 |
0.1412 | 10.73 | 3400 | 0.3084 | 0.8776 | 0.8749 | 0.8762 | 0.9237 |
0.1254 | 11.04 | 3500 | 0.3156 | 0.8738 | 0.8785 | 0.8761 | 0.9237 |
0.1254 | 11.36 | 3600 | 0.3131 | 0.8723 | 0.8818 | 0.8770 | 0.9244 |
0.1254 | 11.67 | 3700 | 0.3108 | 0.8778 | 0.8781 | 0.8780 | 0.9250 |
0.1254 | 11.99 | 3800 | 0.3097 | 0.8778 | 0.8771 | 0.8775 | 0.9239 |
0.1254 | 12.3 | 3900 | 0.3115 | 0.8785 | 0.8801 | 0.8793 | 0.9251 |
0.111 | 12.62 | 4000 | 0.3108 | 0.8772 | 0.8799 | 0.8785 | 0.9249 |
Framework versions
- Transformers 4.20.0.dev0
- Pytorch 1.11.0+cu113
- Datasets 2.2.2
- Tokenizers 0.12.1
- Downloads last month
- 75
This model does not have enough activity to be deployed to Inference API (serverless) yet. Increase its social
visibility and check back later, or deploy to Inference Endpoints (dedicated)
instead.
Spaces using Theivaprakasham/layoutlmv3-finetuned-wildreceipt 2
Evaluation results
- Precision on wild_receiptself-reported0.877
- Recall on wild_receiptself-reported0.880
- F1 on wild_receiptself-reported0.879
- Accuracy on wild_receiptself-reported0.925