best_net_results_with_split

network with Xception as input plus 1 dense (37 outputs), LR with scheduler (starting from 1e-3), Adam optimizer and 'MSE' as loss function

We split training dataset into test set (20%) and train set (80%), using it 10% for validation and 90% for training, obtaining these following results:

RMSE on train = 0.0744, RMSE on test = 0.07298, Validation Loss = 0.00598

here is the NN used:

def CNN_galaxy(model):
  
	inp = model    
	model = Sequential()
	model.add(inp)

	model.add(Flatten())		
	model.add(Dropout(0.25))
	model.add(Dense(NUM_CLASSES))
	model.add(Activation('sigmoid'))

	for layer in model.layers:
	  layer.trainable = True

	print("compiling the model...")
	optimizer = tf.keras.optimizers.Adam(lr=LR, decay=WEIGHT_DECAY)
	model.compile(optimizer, loss='mse', metrics=[root_mean_squared_error])

	return model


img_shape = (224, 224, 3)
xception_model = tf.keras.applications.xception.Xception(include_top=False, input_shape=img_shape)
net = CNN_galaxy(xception_model)