hello everyone!
I am a newbie in machine learning, and I decided to start off with fitting a 3D function z= -x^2 + y^2. First, I created a grid and evaluated the function at each point:
dataset=[(x,y,-x**2+y**2) for x in range(-50,50) for y in range(-50,50)]
coord=dataset[:,0:2]
z=dataset[:,2:]
The model architecture:
opt= tf.keras.optimizers.Adam(learning_rate=0.01)
model = tf.keras.models.Sequential([
Dense(256, activation='relu',input_dim=2),
Dense(128, activation='relu'),
Dense(64, activation='relu'),
Dense(10,activation='relu'),
Dense(1)
])
model.compile(loss='mae', optimizer=opt)
history=model.fit(coord, z, epochs=30,batch_size=15, verbose=1)
At this point I tried to tweak the architecture, tried different loss functions, optimizers and batch sizes. But my loss function, however, does not seem to improve. It follows a good trend, but it gets stuck at pretty high values.
What would you do to get closer to zero? Thank you!
You should definitely use more than 30 epochs. I recommend about 300-600 epochs. This could be what is causing your problem.
