def test_mlr_train(self): mlr = MLR(3, 3) N = 1000 inputs = zeros([3, N]) inputs[randint(3, size=N), range(N)] = 1. self.assertLess(mlr._check_gradient(inputs, inputs, 1e-4), 1e-6) mlr.train(inputs, inputs) # prediction should be perfect (almost always) self.assertLess(sum(mlr.sample(inputs) - inputs), 2)
def test_gradient(self):
mlr = MLR(10, 5)
N = 1000
inputs = randn(mlr.dim_in, N)
outputs = zeros([mlr.dim_out, N])
outputs[randint(mlr.dim_out, size=N), range(N)] = 1.
err = mlr._check_gradient(inputs, outputs, 1e-5)
self.assertLess(err, 1e-8)
# without regularization
for param in ['weights', 'biases']:
err = mlr._check_gradient(inputs,
outputs,
1e-5,
parameters={
'train_weights': param == 'weights',
'train_biases': param == 'biases',
})
self.assertLess(err, 1e-8)
# with regularization
for norm in ['L1', 'L2']:
for param in ['weights', 'biases']:
err = mlr._check_gradient(inputs,
outputs,
1e-7,
parameters={
'train_weights':
param == 'weights',
'train_biases':
param == 'biases',
'regularize_weights': {
'strength': 0.7,
'norm': norm
},
'regularize_biases': {
'strength': 0.3,
'norm': norm
},
})
self.assertLess(err, 1e-6)
def test_mlr_pickle(self):
tmp_file = mkstemp()[1]
model0 = MLR(10, 3)
model0.weights = randn(*model0.weights.shape)
model0.biases = randn(*model0.biases.shape)
# store model
with open(tmp_file, 'w') as handle:
dump({'model': model0}, handle)
# load model
with open(tmp_file) as handle:
model1 = load(handle)['model']
# make sure parameters haven't changed
self.assertEqual(model0.dim_in, model1.dim_in)
self.assertEqual(model0.dim_out, model1.dim_out)
self.assertLess(max(abs(model0.biases - model1.biases)), 1e-20)
self.assertLess(max(abs(model0.weights - model1.weights)), 1e-20)
def test_mlr_pickle(self):
tmp_file = mkstemp()[1]
model0 = MLR(10, 3)
model0.weights = randn(*model0.weights.shape)
model0.biases = randn(*model0.biases.shape)
# store model
with open(tmp_file, 'w') as handle:
dump({'model': model0}, handle)
# load model
with open(tmp_file) as handle:
model1 = load(handle)['model']
# make sure parameters haven't changed
self.assertEqual(model0.dim_in, model1.dim_in)
self.assertEqual(model0.dim_out, model1.dim_out)
self.assertLess(max(abs(model0.biases - model1.biases)), 1e-20)
self.assertLess(max(abs(model0.weights - model1.weights)), 1e-20)
def test_gradient(self):
mlr = MLR(10, 5)
N = 1000
inputs = randn(mlr.dim_in, N)
outputs = zeros([mlr.dim_out, N])
outputs[randint(mlr.dim_out, size=N), range(N)] = 1.
err = mlr._check_gradient( inputs, outputs, 1e-5)
self.assertLess(err, 1e-8)
# without regularization
for param in ['weights', 'biases']:
err = mlr._check_gradient(
inputs,
outputs,
1e-5,
parameters={
'train_weights': param == 'weights',
'train_biases': param == 'biases',
})
self.assertLess(err, 1e-8)
# with regularization
for norm in ['L1', 'L2']:
for param in ['weights', 'biases']:
err = mlr._check_gradient(
inputs,
outputs,
1e-7,
parameters={
'train_weights': param == 'weights',
'train_biases': param == 'biases',
'regularize_weights': {'strength': 0.7, 'norm': norm},
'regularize_biases': {'strength': 0.3, 'norm': norm},
})
self.assertLess(err, 1e-6)
def test_mlr_train(self):
mlr = MLR(3, 3)
N = 1000
inputs = zeros([3, N])
inputs[randint(3, size=N), range(N)] = 1.
self.assertLess(mlr._check_gradient(inputs, inputs, 1e-4), 1e-6)
mlr.train(inputs, inputs)
# prediction should be perfect (almost always)
self.assertLess(sum(mlr.sample(inputs) - inputs), 2)
