hi i ask question

[netilovefm1]

[1/1000] Train loss: 0.02626, Valid loss: 3.31002, Elapsed_time: 13.65418
Current_accuracy : 9.091, Current_norm_ED : 0.52
Best_accuracy : 9.091, Best_norm_ED : 0.52

Ground Truth | Prediction | Confidence Score & T/F

1650/1838,1242/1318 | 1287/1838,1124/1318 | 0.0516 False
1011/1011,84/548 | 1287/1838,1124/1318 | 0.0532 False
1493/1838,1108/1318 | 1287/1838,1124/1318 | 0.0479 False
980/1011,75/548 | 1287/1838,1124/1318 | 0.0499 False
1801/1978,869/1442 | 1287/1838,1124/1318 | 0.0498 False

[1/1000] Train loss: 0.03482, Valid loss: 3.67872, Elapsed_time: 19.08283
Current_accuracy : 0.000, Current_norm_ED : 0.52
Best_accuracy : 0.000, Best_norm_ED : 0.52

Ground Truth | Prediction | Confidence Score & T/F

1650/1838,1242/1318 | 1287/1838,11244/1318 | 0.0098 False
1011/1011,84/548 | 1287/1838,11244/1318 | 0.0098 False
1493/1838,1108/1318 | 1287/1838,11244/1318 | 0.0089 False
980/1011,75/548 | 1287/1838,11244/1318 | 0.0092 False
1801/1978,869/1442 | 1287/1838,11244/1318 | 0.0092 False

[200/1000] Train loss: 0.01555, Valid loss: 3.88356, Elapsed_time: 210.72053
Current_accuracy : 0.000, Current_norm_ED : 0.43
Best_accuracy : 0.000, Best_norm_ED : 0.52

Ground Truth | Prediction | Confidence Score & T/F

980/1011,75/548 | 1287/1838,112441114 | 0.0239 False
1011/1011,548/548 | 1287/1838,112441114 | 0.0250 False
1011/1011,84/548 | 1287/1838,112441114 | 0.0245 False
1011/1011,84/548 | 1287/1838,112441114 | 0.0245 False
1287/1838,1124/1318 | 1287/1838,112441114 | 0.0231 False

python train.py --train_data dataset/train --valid_data dataset/valid --Transformation TPS --FeatureExtraction ResNet --SequenceModeling BiLSTM --Prediction Attn --saved_model best_accuracy.pth

I tested simple pic datas about hp bar in game

and lmdb not work so I changed source a little just dataset.text file to image and label

but it looks like train not bad but prediction is always same... I don't know why this happend

I guess some layer input is corrupted or if it's not that reason it's very similar small hp bar pics so it can't distinguish

can u guess what is problem?....

sorry too ambiguous explaination