We use models from transformers. https://huggingface.co/models
will be released soon.
Anaconda env
| Backbone Model | Size | Dataset | F1 |
|---|---|---|---|
| LayoutLMv3 | BASE | FUNSD | 93.76 |
| LayoutLMv3 | LARGE | FUNSD | 93.52 |
| LayoutLMv3 | BASE | CORD | 97.23 |
| LayoutLMv3 | LARGE | CORD | 94.49 |
python -m torch.distributed.launch --nproc_per_node=1
--master_port 44398
examples/run_funsd.py
--model_name_or_path microsoft/layoutlm-base-uncased
--output_dir output/
--do_train
--do_predict
--max_steps 5000
--warmup_ratio 0.1
--fp16
--per_device_train_batch_size 4python -m torch.distributed.launch --nproc_per_node=1
--master_port 24398
examples/run_funsd.py
--model_name_or_path microsoft/layoutlmv2-large-uncased
--output_dir output/
--do_train
--do_predict
--max_steps 2000
--warmup_ratio 0.1
--fp16
--overwrite_output_dir
--per_device_train_batch_size 4python -m torch.distributed.launch
--nproc_per_node=1
--master_port 4398
examples/run_funsd_cord.py
--dataset_name [funsd or cord]
--do_train
--do_eval
--model_name_or_path microsoft/layoutlmv3-base
--output_dir output/
--segment_level_layout 1
--visual_embed 1
--input_size 224
--max_steps 1000
--save_steps -1
--evaluation_strategy steps
--eval_steps 1000
--learning_rate 1e-5
--per_device_train_batch_size 8
--gradient_accumulation_steps 1
--dataloader_num_workers 1
--overwrite_output_dir python run_classification.py
--data_dir [datasetPath]
--model_type layoutlm
--model_name_or_path ~/dev/Models/LayoutLM/layoutlm-base-uncased
--output_dir output/
--do_lower_case
--max_seq_length 512
--do_train
--do_eval
--num_train_epochs 40.0
--logging_steps 5000
--save_steps 5000
--per_gpu_train_batch_size 16
--per_gpu_eval_batch_size 16
--evaluate_during_training
--fp16 --data_level 1python run_classification.py
--data_dir [datasetPath]
--model_type v3
--model_name_or_path microsoft/layoutlmv3-base
--do_lower_case
--max_seq_length 512
--do_train
--num_train_epochs 40.0
--logging_steps 5000
--save_steps 5000
--per_gpu_train_batch_size 2
--per_gpu_eval_batch_size 2
--evaluate_during_trainingWe use some codes from https://github.com/microsoft/unilm.