def get_param_names3(num_inputs, layer_sizes, num_outputs, m_trainable_arr,
b_trainable_arr):
"""
see calc_num_weights3 for relevant of suffix
probably a way to code this function with less conditionals, but not work
spending any more time on
"""
weight_shapes = tools.get_weight_shapes3(num_inputs, layer_sizes,
num_outputs, m_trainable_arr,
b_trainable_arr)
param_names = []
i = 0
j = 0
weight_offset = 1
bias_offset = 1
if np.all(~np.array(m_trainable_arr)) and np.all(
~np.array(b_trainable_arr)):
print "no trainable parameters"
return param_names
for k in range(len(m_trainable_arr)):
if m_trainable_arr[k] and b_trainable_arr[k]:
weight_row = weight_shapes[j][0]
weight_col = weight_shapes[j][1]
bias = weight_shapes[j + 1][0]
j += 2
elif m_trainable_arr[k] and not b_trainable_arr[k]:
weight_row = weight_shapes[j][0]
weight_col = weight_shapes[j][1]
bias = 0
j += 1
elif not m_trainable_arr[k] and b_trainable_arr[k]:
weight_row = 0
weight_col = 0
bias = weight_shapes[j][0]
j += 1
else:
weight_row = 0
weight_col = 0
bias = 0
for l in range(weight_row):
for m in range(weight_col):
param_names.append(
('p%i' % (i + 1),
'w_{%i,%i,%i}' % (k + weight_offset, l + 1, m + 1)))
i += 1
for l in range(bias):
param_names.append(
('p%i' % (i + 1), 'b_{%i,%i}' % (k + bias_offset, l + 1)))
i += 1
return param_names
def main(run_string):
###### load training data
data = 'bh_50'
data_suffix = '_tr_1.csv'
data_dir = '../../data/uci/'
data_prefix = data_dir + data
x_tr, y_tr = input_tools.get_x_y_tr_data(data_prefix, data_suffix)
x_tr = np.genfromtxt('../../data/linear_input_data.txt', delimiter=',')
y_tr = x_tr
batch_size = x_tr.shape[0]
###### get weight information
weights_dir = '../../data/' #for forward test
a1_size = 0
num_inputs = tools.get_num_inputs(x_tr)
num_outputs = tools.get_num_outputs(y_tr)
layer_sizes = [1, num_inputs] * 2
m_trainable_arr = [True, True] * 2 + [False]
b_trainable_arr = [True, True] * 2 + [False]
num_weights = tools.calc_num_weights3(num_inputs, layer_sizes, num_outputs,
m_trainable_arr, b_trainable_arr)
###### check shapes of training data
x_tr, y_tr = tools.reshape_x_y_twod(x_tr, y_tr)
###### setup prior
hyper_type = "deterministic" # "stochastic" or "deterministic"
var_type = "deterministic" # "stochastic" or "deterministic"
weight_shapes = tools.get_weight_shapes3(num_inputs, layer_sizes,
num_outputs, m_trainable_arr,
b_trainable_arr)
dependence_lengths = tools.get_degen_dependence_lengths(weight_shapes,
independent=True)
if hyper_type == "deterministic" and var_type == "deterministic":
prior_types = [4]
prior_hyperparams = [[0., 1.]]
param_prior_types = [0]
prior = inverse_priors.inverse_prior(prior_types, prior_hyperparams,
dependence_lengths,
param_prior_types, num_weights)
n_stoc = 0
n_stoc_var = 0
elif hyper_type == "stochastic" and var_type == "deterministic":
granularity = 'single'
hyper_dependence_lengths = tools.get_hyper_dependence_lengths(
weight_shapes, granularity)
hyperprior_types = [9]
prior_types = [4]
hyperprior_params = [[1. / 2., 1. / (2. * 100)]]
prior_hyperparams = [0.]
param_hyperprior_types = [0]
param_prior_types = [0]
n_stoc = len(hyper_dependence_lengths)
prior = isp.inverse_stoc_hyper_prior(
hyperprior_types, prior_types, hyperprior_params,
prior_hyperparams, hyper_dependence_lengths, dependence_lengths,
param_hyperprior_types, param_prior_types, n_stoc, num_weights)
n_stoc_var = 0
elif hyper_type == "stochastic" and var_type == "stochastic":
granularity = 'single'
hyper_dependence_lengths = tools.get_hyper_dependence_lengths(
weight_shapes, granularity)
var_dependence_lengths = [1]
n_stoc_var = len(var_dependence_lengths)
hyperprior_types = [9]
var_prior_types = [10]
prior_types = [4]
hyperprior_params = [[1. / 2., 1. / (2. * 100)]]
var_prior_params = [[1. / 2., 1. / (2. * 100)]]
prior_hyperparams = [0.]
param_hyperprior_types = [0]
var_param_prior_types = [0]
param_prior_types = [0]
n_stoc = len(hyper_dependence_lengths)
prior = isvp.inverse_stoc_var_hyper_prior(
hyperprior_types, var_prior_types, prior_types, hyperprior_params,
var_prior_params, prior_hyperparams, hyper_dependence_lengths,
var_dependence_lengths, dependence_lengths, param_hyperprior_types,
var_param_prior_types, param_prior_types, n_stoc, n_stoc_var,
num_weights)
###### test prior output from nn setup
if "nn_prior_test" in run_string:
prior_tests.nn_prior_test(prior, n_stoc + n_stoc_var + num_weights)
#set up np model
np_nn = npms.mlp_ResNet_2
npm = np_model(np_nn, x_tr, y_tr, batch_size, layer_sizes, m_trainable_arr,
b_trainable_arr, n_stoc_var)
ll_type = 'gauss' # 'gauss', 'av_gauss', 'categorical_crossentropy', 'av_categorical_crossentropy'
npm.setup_LL(ll_type)
###### test llhood output
if "forward_test_linear" in run_string:
forward_tests.forward_test_linear([npm], num_weights + n_stoc_var,
weights_dir)
###### setup polychord
nDerived = 0
settings = PyPolyChord.settings.PolyChordSettings(
n_stoc + n_stoc_var + num_weights, nDerived)
settings.base_dir = './np_chains/'
settings.file_root = data + "_slp_sh_sv_sm"
settings.nlive = 1000
###### run polychord
if "polychord1" in run_string:
PyPolyChord.run_polychord(npm, n_stoc, n_stoc_var, num_weights,
nDerived, settings, prior,
polychord_tools.dumper)
def get_param_names_svh3(num_inputs,
layer_sizes,
num_outputs,
m_trainable_arr,
b_trainable_arr,
granularity=None,
n_stoc_var=0):
"""
NOT TESTED
see calc_num_weights3 for relevant of suffix
"""
weight_shapes = tools.get_weight_shapes3(num_inputs, layer_sizes,
num_outputs, m_trainable_arr,
b_trainable_arr)
param_names = []
i = 0
j = 0
weight_offset = 1
bias_offset = 1
if np.all(~np.array(m_trainable_arr)) and np.all(
~np.array(b_trainable_arr)):
print "no trainable parameters"
return param_names
if granularity == 'single':
param_names.append(('p1', 'h_{1}'))
i += 1
elif granularity == 'layer':
for k in range(len(m_trainable_arr)):
if m_trainable_arr[k] or b_trainable_arr[k]:
param_names.append(('p%i' % (i + 1), 'h_{%i}' % (k + 1)))
i += 1
elif granularity == 'input_size':
t = 0
w = 0
for k in range(len(m_trainable_arr)):
if m_trainable_arr[k] and b_trainable_arr[k]:
t += weight_shapes[w][0]
t += 1
w += 2
elif m_trainable_arr[k] and not b_trainable_arr[k]:
t += weight_shapes[w][0]
w += 1
elif not m_trainable_arr[k] and b_trainable_arr[k]:
t += 1
w += 1
else:
pass
for n in range(t):
param_names.append(
('p%i' % (i + 1), 'h_{%i, %i}' % (k + 1, n + 1)))
i += 1
t = 0
else: #no stochastic prior hyperparams
pass
if n_stoc_var:
for k in range(n_stoc_var):
param_names.append(('p%i' % (i + 1), 'v_{%i}' % (k + 1)))
i += 1
else: #no stochastic lhood variance parameters
pass
for k in range(len(m_trainable_arr)):
if m_trainable_arr[k] and b_trainable_arr[k]:
weight_row = weight_shapes[j][0]
weight_col = weight_shapes[j][1]
bias = weight_shapes[j + 1][0]
j += 2
elif m_trainable_arr[k] and not b_trainable_arr[k]:
weight_row = weight_shapes[j][0]
weight_col = weight_shapes[j][1]
bias = 0
j += 1
elif not m_trainable_arr[k] and b_trainable_arr[k]:
weight_row = 0
weight_col = 0
bias = weight_shapes[j][0]
j += 1
else:
weight_row = 0
weight_col = 0
bias = 0
for l in range(weight_row):
for m in range(weight_col):
param_names.append(
('p%i' % (i + 1),
'w_{%i,%i,%i}' % (k + weight_offset, l + 1, m + 1)))
i += 1
for l in range(bias):
param_names.append(
('p%i' % (i + 1), 'b_{%i,%i}' % (k + bias_offset, l + 1)))
i += 1
return param_names