# Set models according to settings
if FLAGS.mode == TrainMode.WIDE:
models = [linear_model]
elif FLAGS.mode == TrainMode.DEEP:
models = [mlp_model]
elif FLAGS.mode == TrainMode.WIDE_AND_DEEP:
models = [linear_model, mlp_model]
else:
raise ValueError('Unsupported option in Boston training %s'
% FLAGS.mode)
model_metrics = [
metrics.AbsError(),
metrics.MeanPredError(),
metrics.MedianPredError(),
metrics.Mean(),
metrics.Std(),
metrics.AbsErrorInterval(cuts=[10, 15, 20]),
metrics.RMSEInterval(cuts=[10, 15, 20])
]
# Create model
joint = JointMDN(mixture=GaussianMixture(components=FLAGS.K),
model_dir=utils.get_default_output(),
models=models,
outputs=dataset.get_num_classes(),
l1_reg=FLAGS.l1_regularization,
l2_reg=FLAGS.l2_regularization,
clip_gradient=FLAGS.gradient_clip)
if FLAGS.training:
# Create model
joint = JointRegressor(ut.get_default_output(),
models=models,
outputs=dataset.get_num_classes(),
l1_reg=FLAGS.l1_regularization,
l2_reg=FLAGS.l2_regularization,
loss_fn=MeanSquared(),
clip_gradient=FLAGS.gradient_clip)
# Define metrics
metrics = [
me.AbsError(),
me.MeanPredError(),
me.MedianPredError(),
me.Mean(),
me.Std(),
me.AbsErrorInterval(cuts=[60, 120, 200, 400])
]
if FLAGS.training:
# Start training
joint.train(dataset,
batch_size=FLAGS.batch_size,
track_models=FLAGS.checkpoints,
track_summaries=FLAGS.summaries,
steps=FLAGS.steps,
gpu_frac=FLAGS.gpu_frac,
metrics=metrics)
else:
# Evaluate test
def runTest(self):
self.base_test(me.Mean(), 3.90)