deep-text-recognition-benchmark icon indicating copy to clipboard operation
deep-text-recognition-benchmark copied to clipboard

Published 20 hours ago • verified publisher icon clovaai

100% accuracy with poor predictions on finetuned model

Open sultanovazamat opened this issue 7 months ago • 0 comments

Hi!

Has anyone faced the issue of ideal accuracy already on the 20th iteration of FT, but with poor results?

Image

This is my training command:

!python train.py --train_data /content/drive/MyDrive/lmbd_output_train --valid_data /content/drive/MyDrive/lmbd_output_val --select_data "/" --batch_ratio 1 --Transformation TPS --FeatureExtraction ResNet --SequenceModeling BiLSTM --Prediction Attn --batch_size 256 --data_filtering_off --workers 0 --batch_max_length 80 --num_iter 1000 --valInterval 10 --sensitive --saved_model /content/drive/MyDrive/TPS-ResNet-BiLSTM-Attn-case-sensitive.pth --data_filtering_off --FT --adam --imgW 180 --PAD

I am using small datasets: train - 3000 imgs and val - 1500 imgs to set up the training correctly and then run the training on the huge dataset.

sultanovazamat avatar Mar 03 '25 06:03 sultanovazamat