Advice on fine tuning to improve accuracy on digits and special characters

[itsainii]

I'm trying to fine tune EasyOCR to improve accuracy with digits and special characters. I found this dataset on Kaggle and applied the method outlined in this tutorial.

I tested with 40 images just for testing, and here are the results I obtained:

exp_name: TPS-ResNet-BiLSTM-Attn-Seed1111
train_data: lmbd_archive2/
valid_data: lmbd_archive2/
manualSeed: 1111
workers: 0
batch_size: 1
num_iter: 10
valInterval: 5
saved_model: TPS-ResNet-BiLSTM-Attn.pth
FT: False
adam: False
lr: 1
beta1: 0.9
rho: 0.95
eps: 1e-08
grad_clip: 5
baiduCTC: False
select_data: ['/']
batch_ratio: ['0.5']
total_data_usage_ratio: 1.0
batch_max_length: 80
imgH: 32
imgW: 100
rgb: False
character: 0123456789abcdefghijklmnopqrstuvwxyz
sensitive: False
PAD: False
data_filtering_off: True
Transformation: TPS
FeatureExtraction: ResNet
SequenceModeling: BiLSTM
Prediction: Attn
num_fiducial: 20
input_channel: 1
output_channel: 512
hidden_size: 256
num_gpu: 1
num_class: 38

[1/10] Train loss: 8.93992, Valid loss: 7.16805, Elapsed_time: 1.62669
Current_accuracy : 0.000, Current_norm_ED : 0.00
Best_accuracy : 0.000, Best_norm_ED : 0.00

Ground Truth | Prediction | Confidence Score & T/F

                      | sichasism                 | 0.0064	False

---

[5/10] Train loss: 4.55050, Valid loss: 1.03087, Elapsed_time: 6.90535
Current_accuracy : 65.854, Current_norm_ED : 0.00
Best_accuracy : 65.854, Best_norm_ED : 0.00

Ground Truth | Prediction | Confidence Score & T/F

                      | iisile                    | 0.0647	False

---

[10/10] Train loss: 0.01275, Valid loss: 0.22137, Elapsed_time: 10.19817
Current_accuracy : 95.122, Current_norm_ED : 0.00
Best_accuracy : 95.122, Best_norm_ED : 0.00

Ground Truth | Prediction | Confidence Score & T/F

                      |                           | 0.0000	True

---

end the training

Any advice or suggestions on where I should go from here?