20231102-20_epochs_layoutlmv2-base-uncased_finetuned_docvqa
This model was trained from scratch on the 1.2 Example dataset released by DocVQA.
It achieves the following results on the evaluation set:
Model description
This DocVQA model, built on the Layout LM v2 framework, represents an initial step in a series of experimental models aimed at document visual question answering. It's the "mini" version in a planned series, trained on a relatively small dataset of 1.2k samples (1,000 for training and 200 for testing) over 20 epochs. The training setup was modest, employing mixed precision (fp16), with manageable batch sizes and a focused approach to learning rate adjustment (warmup steps and weight decay). Notably, this model was trained without external reporting tools, emphasizing internal evaluation. As the first iteration in a progressive series that will later include medium (5k samples) and large (50k samples) models, this version serves as a foundational experiment, setting the stage for more extensive and complex models in the future.
Intended uses & limitations
Experimental Only
Training and evaluation data
Based on the sample 1.2 dataset released by DocVQA
Training procedure
Training hyperparameters
The following hyperparameters were used during training:
- learning_rate: 5e-05
- train_batch_size: 16
- eval_batch_size: 16
- seed: 42
- gradient_accumulation_steps: 2
- total_train_batch_size: 32
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- num_epochs: 20
Training results
| Training Loss | Epoch | Step | Validation Loss |
|---|
| 4.3689 | 3.51 | 100 | 3.7775 |
| 3.2761 | 7.02 | 200 | 3.3707 |
| 2.6415 | 10.53 | 300 | 3.0807 |
| 2.2233 | 14.04 | 400 | 3.0120 |
| 1.9586 | 17.54 | 500 | 2.9087 |
Framework versions
- Transformers 4.34.1
- Pytorch 2.0.1+cu118
- Datasets 2.10.1
- Tokenizers 0.14.1
