C2_W2_Softmax assignment - behavior in model training

I finished the assignment.

After I went and updated epoch = 150 to see about removing the errors in the model. I noticed this in the loss calculations during the training:

Epoch 72/150
157/157 [==============================] - 0s 2ms/step - loss: 3.3577e-04
Epoch 73/150
157/157 [==============================] - 0s 2ms/step - loss: 3.2813e-04
Epoch 74/150
157/157 [==============================] - 0s 2ms/step - loss: 0.0553
Epoch 75/150
157/157 [==============================] - 0s 2ms/step - loss: 0.0109
Epoch 76/150
157/157 [==============================] - 0s 2ms/step - loss: 0.0030
Epoch 77/150
157/157 [==============================] - 0s 2ms/step - loss: 8.2575e-04
Epoch 78/150
157/157 [==============================] - 0s 2ms/step - loss: 6.1341e-04
Epoch 79/150
157/157 [==============================] - 0s 2ms/step - loss: 5.3338e-04

It happened a couple different times. Final model is ~e-05 and worked well. What I am looking for is insight into why these huge blips occurred at these various iterations. There really isn’t enough in the course on Adam for me to have good insight into this. I reduced learning rate by half and still see blips. Curious to learn more.

Thanks

Hey @Goertzel,

To make sure that i understand you well. Your question is about why the changing in loss e.g " Sometimes higher and then lower" right?

Hi,
I’ve not been able to repeat your results. I ran it several times, and always get a curve that looks like this (I added code to plot the loss history).
Did you make any other changes besides setting the epochs to 150?

image