duplicate of https://huggingface.co/spaces/to-be/invoice_document_headers_extraction_with_donut/discussions/2

Hello @to-be .Thank you for the demo. Since the dataset is confidential, I have 2 questions regarding the training that you've done please :

• What hyperparameters did you choose for your training (I have a similar dataset -- as I've seen int your 3 test invoices --) ?
• How were you able to use a different input size (I get an error indicating a mismatch when I change the default input size used in Donut) ?

Thank you in advance.

1. train_batch_sizes:
   • 1
     val_batch_sizes:
   • 2
     input_size:
   • 1600
   • 1280
     max_length: 256
     align_long_axis: False
     num_nodes: 1
     seed: 2022
     lr: 3e-05
     warmup_steps: 300
     num_training_samples_per_epoch: 1200
     max_epochs: 100
     max_steps: -1
     num_workers: 4
     val_check_interval: 1.0
     check_val_every_n_epoch: 3
     gradient_clip_val: 1.0
2. from transformers import VisionEncoderDecoderConfig

max_length = 768
image_size = [1920, 1280]
#image_size = [1280, 960]

update image_size of the encoder

during pre-training, a larger image size was used

config = VisionEncoderDecoderConfig.from_pretrained("naver-clova-ix/donut-base")
config.encoder.image_size = image_size # (height, width)

update max_length of the decoder (for generation)

config.decoder.max_length = max_length