def test_sum_hamiltonian_calls_np_dot_with_correct_args(self,mock_dot):
hmc = HMCNeuron('fake_arg','fake_arg','fake_arg')
momentum = np.array([.01,.02])
obj_error = 100
hmc.sum_hamiltonian(momentum,obj_error)
assert(np.array_equal(mock_dot.call_args[0][0], momentum))
assert(np.array_equal(mock_dot.call_args[0][1], momentum * .5))
def test_hmc_neuron_full_example(self):
"""
A complete HMCNeuron example:
Create a classification problem by initiating 0,1 class labels by
sampling 100 data points from a binomial distribution.
First, create two predictor inputs; Artificially simulate that one
predicts well and the other doesn't. Then, zero center and normalize
the inputs and initialize the neuron. Finally, integrate the converged
weights and make predictions.
"""
actual_outputs = np.random.binomial(1, .3, 100)
df = pd.DataFrame({'output':actual_outputs})
df['input_1'] = np.ones(df.shape[0])
df['input_2'] = np.ones(df.shape[0])
feat_cols = ['input_1','input_2']
# Make 'input_1' a good predictor
df.ix[df['output']!=1, 'input_1'] = 0
# Make 'input_2' a noisy, random, bad predictor
df.ix[df.sample(30).index, 'input_2'] = 0
df = zero_center(df, 'input_1')
df = zero_center(df, 'input_2')
df = normalize(df, 'input_1')
df = normalize(df, 'input_2')
hmc = HMCNeuron(df[feat_cols],df['output'],iterations=100)
hmc.average_converged_weights()
df['predictions'] = df[feat_cols].apply(
lambda row: hmc.predict_row(row), axis=1
)
print(hmc.weights)
print('doesnt',df[df['predictions']!= df['output']])
print('does',df[df['predictions']== df['output']])
def test_init_weights_shape_is_correct(self):
hmc = HMCNeuron('fake_arg','fake_arg','fake_arg')
weight_count = 3
weights = hmc.init_weights(weight_count)
actual = weights.shape[0]
expected = weight_count
assert(actual == expected)
def test_init_momentum_shape_is_correct(self):
hmc = HMCNeuron('fake_arg','fake_arg','fake_arg')
columns = 3
momentum = hmc.init_momentum(columns)
actual = momentum.shape[0]
expected = columns
assert(actual == expected)
def test_leapfrog_method_example(self):
hmc = HMCNeuron('fake_arg','fake_arg','fake_arg')
hmc.grad_obj = mock.Mock()
hmc.grad_obj.return_value = np.array([.02,.02]) # Fake gradient calc
weights = np.array([.1,.2]) # Fake initialized weights
momentum = np.array([.01,.02]) # Fake initial momentum
gradient = np.array([.01,.01]) # Another fake gradient
momentum = hmc.leapfrog_method(momentum,weights,gradient)
def test_sum_hamiltonian_example(self):
"""
The method sum_hamiltonian gives us the sum dot product of the
momentum and the momentum transposed plus the objective
function error.
"""
hmc = HMCNeuron('fake_arg','fake_arg','fake_arg')
momentum = np.array([.01,.02]) # Fake initial momentum
obj_error = 100
momentum = hmc.sum_hamiltonian(momentum,obj_error)
def test_leapfrog_method_must_call_grad_obj_count(self,mock_randint):
mock_randint.return_value = 100
hmc = HMCNeuron('fake_arg','fake_arg','fake_arg')
hmc.grad_obj = mock.Mock()
hmc.grad_obj.return_value = np.array([.02,.02])
weights = np.array([.1,.2])
momentum = np.array([.01,.02])
gradient = np.array([.01,.01])
momentum = hmc.leapfrog_method(momentum,weights,gradient)
actual = hmc.grad_obj.call_count
expected = 100
assert(actual == expected)
def test_leapfrog_method_args_must_be_np_arrays(self):
""" When the args aren't numpy arrays, it should fail."""
with nose.tools.assert_raises(TypeError) as te:
hmc = HMCNeuron('fake_arg','fake_arg','fake_arg')
hmc.grad_obj = mock.Mock()
hmc.grad_obj.return_value = [.02,.02]
weights = np.array([.1,.2])
momentum = np.array([.01,.02])
gradient = np.array([.01,.01])
momentum = hmc.leapfrog_method(momentum,weights,gradient)
expected = 'can\'t multiply sequence by non-int of type \'float\''
actual = str(te.exception)
assert(actual == expected)
def test_leapfrog_method_randomly_generates_tau(self,mock_randint):
"""
Make sure the number of gradient calculations is randomly
chosen.
"""
hmc = HMCNeuron('fake_arg','fake_arg','fake_arg')
hmc.grad_obj = mock.Mock()
hmc.grad_obj.return_value = np.array([.02,.02])
weights = np.array([.1,.2])
momentum = np.array([.01,.02])
gradient = np.array([.01,.01])
momentum = hmc.leapfrog_method(momentum,weights,gradient)
assert(mock_randint.called_once)
def test_leapfrog_method_return_value_must_be_np_array(self):
"""
The return value of leapfrog_method should be a numpy array;
If this were simply an array, a type error would be raised once
I multiply momentum by .5 (which is exactly what happens next in
the training algorithm).
"""
hmc = HMCNeuron('fake_arg','fake_arg','fake_arg')
hmc.grad_obj = mock.Mock()
hmc.grad_obj.return_value = np.array([.02,.02])
weights = np.array([.1,.2])
momentum = np.array([.01,.02])
gradient = np.array([1,1])
momentum = hmc.leapfrog_method(momentum,weights,gradient)
momentum * .5
def test_leapfrog_method_return_array_correct_dimensions(self):
"""
The return value of leapfrog_method should be 1 row with
an element for each input column.
"""
hmc = HMCNeuron('fake_arg','fake_arg','fake_arg')
hmc.grad_obj = mock.Mock()
hmc.grad_obj.return_value = np.array([.02,.02,.4])
weights = np.array([.1,.2,.2])
momentum = np.array([.01,.02,.1])
gradient = np.array([.01,.01,.1])
momentum = hmc.leapfrog_method(momentum,weights,gradient)
actual = len(momentum)
expected = 3
assert(actual==expected)
def test_hmc_neuron_init_stores_arguments(self):
"""
Test that inputs, outputs and iterations get set to the
instance.
"""
neuron = HMCNeuron('inputs','outputs','iterations')
assert(neuron.inputs == 'inputs')
assert(neuron.outputs == 'outputs')
assert(neuron.iterations == 'iterations')
def test_init_weights_shape_raises_when_too_many_args(self):
hmc = HMCNeuron('fake_arg','fake_arg','fake_arg')
hmc.init_weights(3,2)
def test_init_weights_shape_raises_when_no_args_raises(self):
hmc = HMCNeuron('fake_arg','fake_arg','fake_arg')
hmc.init_weights()
def test_hmc_neuron_init_raises_when_no_args(self):
neuron = HMCNeuron()
def test_store_weights_raises_when_too_many_args(self):
hmc = HMCNeuron('fake_arg','fake_arg','fake_arg')
hmc.store_weights(1,2,3)
def test_store_weights_raises_when_no_args(self):
hmc = HMCNeuron('fake_arg','fake_arg','fake_arg')
hmc.store_weights()
def test_init_momentum_shape_raises_when_no_args_raises(self):
hmc = HMCNeuron('fake_arg','fake_arg','fake_arg')
hmc.init_momentum()
def test_sum_hamiltonian_calls_np_dot_once(self,mock_dot):
hmc = HMCNeuron('fake_arg','fake_arg','fake_arg')
momentum = np.array([.01,.02])
obj_error = 100
hmc.sum_hamiltonian(momentum,obj_error)
assert(mock_dot.called_once)
def test_init_momentum_shape_raises_when_too_many_args(self):
hmc = HMCNeuron('fake_arg','fake_arg','fake_arg')
hmc.init_momentum(3,2)
def test_leapfrog_method_too_many_args(self):
hmc = HMCNeuron('fake_arg','fake_arg','fake_arg')
momentum = hmc.leapfrog_method([],[],[],[])
def test_leapfrog_method_not_enough_args(self):
hmc = HMCNeuron('fake_arg','fake_arg','fake_arg')
momentum = hmc.leapfrog_method()
def test_sum_hamiltonian_calls_np_dot_with_momentum(self):
hmc = HMCNeuron('fake_arg','fake_arg','fake_arg')
momentum = np.array([.01,.02])
obj_error = 100.0
actual = hmc.sum_hamiltonian(momentum,obj_error)
assert(isinstance(actual, float))
def test_store_weights_stores_weights(self):
hmc = HMCNeuron('fake_arg','fake_arg','fake_arg')
weights = [1,2,3]
hmc.store_weights(weights)
assert(hmc.weight_samples == [[1,2,3]])
def test_hmc_neuron_init_takes_three_args(self):
neuron = HMCNeuron('inputs','outputs','iterations')
