why the result are not good ?

[PythonImageDeveloper]

Hi , I'vd trained the nasnet-large on my own dataset that include 13400 samples for training with 12 classes , i trained this network until 30k step but the results are very bad may top-1 is 9% . why ? but when i train the inception-v4 only 7k step the result is 98 % , Where am i wrong ? i retrain all layers , with this commad :

python3 train_image_classifier.py \
    --train_dir=./training_nasnet_large_plants \
    --dataset_dir=./flowers \
    --dataset_name=flowers \
    --dataset_split_name=train \
    --model_name=nasnet_large \
    --checkpoint_path=./nasnet-a_large_04_10_2017/model.ckpt \
    --max_number_of_steps=30000 \
    --checkpoint_exclude_scopes=final_layer,aux_11