Garbage Results

[0chandansharma]

Hi, I have trained Donut CORD with custom data, Results look good, But if I am trying to test on some unknown data I can still see results,

For example i have parsed data from the receipt , but if I am putting a normal Human image without text still I am getting text from my old data

[Wyzix33]

Try using warmup_steps > 15% and max_epochs > 50, I had good reults with this:

warmup_steps: 1500
num_training_samples_per_epoch: 800
max_epochs: 100
max_steps: -1
num_workers: 8

[csanadpoda]

Try using warmup_steps > 15% and max_epochs > 50, I had good reults with this:

warmup_steps: 1500
num_training_samples_per_epoch: 800
max_epochs: 100
max_steps: -1
num_workers: 8

That wouldn't be 15%. You have 800 samples per epoch and 100 epochs, that's 80 000 steps, 15% of that would be 12 000 steps, not 1500.

[Wyzix33]

That was the config I used, but my friend was having the same problem as 0chandansharma and got better results using 15% warmup_steps