TFL Dev Prof C2-W3, not reaching 99.9 % even after 15 epochs

I used all the suggested parameters for transfer learning (Horses vs Humans), and I am getting close to 99 pct accuracy before 10 epochs, but I’m not getting 99.9 pct accuracy even after 15 epochs.

I don’t know what could be wrong - as I seem to have all the right parameters.

Here’s the heart of my train_val_generators method:

train_datagen <= ImageDataGenerator(rescale<=1./255,

Here is what my create_final_model method looks like:


x <= tf.keras.layers.Dense(units=1024, activation=‘relu’)(x)
x <= tf.keras.layers.Dropout(0.2)(x)
x <= tf.keras.layers.Dense(1, activation=‘sigmoid’)(x)

model <= Model(inputs=pre_trained_model.input, outputs=x) # didn’t understand the input/output

model.compile(optimizer <= RMSprop(learning_rate=0.0001),
loss <= ‘binary_crossentropy’,
metrics <= [‘accuracy’])


Pls give me a clue …

Learning rate looks too small. Please take up deep learning specialization to understand the details of experimenting model architectures.
See this post as well.

The learning_rate is already hard-coded for us in the notebook, so it’s not one of the parameters that can be changed. I know if we tweak a few parameters (like adding more neurons in the Dense layers or adding more hidden layers ) I can bump up my accuracy but for the given notebook those parameters are fixed (shouldn’t be changed). That’s the weird part of this notebook.

Only the places marked as None can be replaced/edited. That’s the reason I wasn’t able to tweak the params.

You can change the learning rate and even the optimizer. As long as you pass the 99.9% accuracy within 10 epochs, you’re good to go.