Low accuracy of LRCN model

[shashi438]

Hello,

I'm trying to recreate the results you've mentioned in your medium blog post. I've split the dataset using your scripts, used the default hyper parameters-
1000 epochs
lr = 1e-5, decay - 1e-6
Adam optimizer

But after 60 epochs or so, the results are as follows - (very low accuracy)
131/131 [==============================] - 1315s 10s/step - loss: 3.8444 - acc: 0.1359 - top_k_categorical_accuracy: 0.3930 - val_loss: 4.3090 - val_acc: 0.0922 - val_top_k_categorical_accuracy: 0.2879

Can you please clarify if I'm missing something, or I'm supposed to run for 500+ epochs? thanks

[self-ms]

If your data is embedded and labels are available, you can use the following repository:
https://github.com/ms-unlimit/Transformer-Based-Machine-Learning-Framework