Can't reproduce pre-trained performance. accuracy does not improve after first epoch of training.

[bluesy7585]

Hi
thanks for your work.
im trying to train the model from begin.
but found that accuracy does not get better after first epoch of training.

here's command i used
python train.py --data-path .\dataset --dataset market --num-epoch 60

Epoch 1/60

step: (0/12936) | label loss: 0.6937
step: (3200/12936) | label loss: 0.3057
step: (6400/12936) | label loss: 0.3116
step: (9600/12936) | label loss: 0.2859
step: (12800/12936) | label loss: 0.3021
train Loss: 0.3216 Acc: 0.8732
step: (0/3368) | label loss: 0.3219
step: (3200/3368) | label loss: 0.2741
val Loss: 0.3034 Acc: 0.8799

...

Epoch 60/60

step: (0/12936) | label loss: 0.3060
step: (3200/12936) | label loss: 0.2891
step: (6400/12936) | label loss: 0.3144
step: (9600/12936) | label loss: 0.3067
step: (12800/12936) | label loss: 0.3105
train Loss: 0.3132 Acc: 0.8750
step: (0/3368) | label loss: 0.3128
step: (3200/3368) | label loss: 0.3078
val Loss: 0.3047 Acc: 0.8788

accuracy stuck at 0.87~0.88
while pre-trained model gives

Average accuracy: 0.9361
Average precision: 0.7306
Average recall: 0.6048
Average f1 score: 0.6485

any specific setting or flags should i use to reproduce your pre-trained model performance?
thanks