def __init__(self, num_subfunctions=50, num_dims=10, objective='l2'):
# don't introduce a Theano dependency until we have to
from utils import _tonp
self.name = 'DeepAE'
layer_sizes = [ 28*28, 1000, 500, 250, 30]
#layer_sizes = [ 28*28, 20]
# Load data
X, y = load_mnist()
# break the data up into minibatches
self.subfunction_references = []
for mb in range(num_subfunctions):
self.subfunction_references.append([X[:, mb::num_subfunctions], y[mb::num_subfunctions]])
# evaluate on subset of training data
self.n_full = 10000
idx = random_choice(X.shape[1], self.n_full, replace=False)
##use all the training data for a smoother plot
#idx = np.array(range(X.shape[1]))
self.full_objective_references = [[X[:,idx].copy(), y[idx].copy()]]
from dropout.deepae import build_f_df # here so theano not required for import
self.theano_f_df, self.model = build_f_df(layer_sizes, use_bias=True, objective=objective)
crossent_params = False
if crossent_params:
history = np.load('/home/poole/Sum-of-Functions-Optimizer/sfo_output.npz')
out = dict(history=history['arr_0'])
params = out['history'].item()['x']['SFO']
self.theta_init = params
else:
self.theta_init = [param.get_value() for param in self.model.params]
def __init__(self, num_subfunctions=50, num_dims=10, objective='l2'):
# don't introduce a Theano dependency until we have to
from utils import _tonp
self.name = 'DeepAE'
layer_sizes = [28 * 28, 1000, 500, 250, 30]
#layer_sizes = [ 28*28, 20]
# Load data
X, y = load_mnist()
# break the data up into minibatches
self.subfunction_references = []
for mb in range(num_subfunctions):
self.subfunction_references.append(
[X[:, mb::num_subfunctions], y[mb::num_subfunctions]])
# evaluate on subset of training data
self.n_full = 10000
idx = random_choice(X.shape[1], self.n_full, replace=False)
##use all the training data for a smoother plot
#idx = np.array(range(X.shape[1]))
self.full_objective_references = [[X[:, idx].copy(), y[idx].copy()]]
from dropout.deepae import build_f_df # here so theano not required for import
self.theano_f_df, self.model = build_f_df(layer_sizes,
use_bias=True,
objective=objective)
crossent_params = False
if crossent_params:
history = np.load(
'/home/poole/Sum-of-Functions-Optimizer/sfo_output.npz')
out = dict(history=history['arr_0'])
params = out['history'].item()['x']['SFO']
self.theta_init = params
else:
self.theta_init = [
param.get_value() for param in self.model.params
]
def __init__(self, num_subfunctions=100, num_dims=10, rectifier='soft'):
self.name = 'MLP'
#layer_sizes = [ 28*28, 1200, 10 ]
#layer_sizes = [ 28*28, 1200, 10 ]
#layer_sizes = [ 28*28, 500, 120, num_dims ]
#layer_sizes = [ 28*28, 120, 12, num_dims ]
layer_sizes = [ 28*28, 1200, 1200, num_dims ]
# Load data
X, y = load_mnist()
# break the data up into minibatches
self.subfunction_references = []
for mb in range(num_subfunctions):
self.subfunction_references.append([X[:, mb::num_subfunctions], y[mb::num_subfunctions]])
# evaluate on subset of training data
idx = random_choice(X.shape[1], 5000, replace=False)
#use all the training data for a smoother plot
#idx = np.array(range(X.shape[1]))
self.full_objective_references = [[X[:,idx].copy(), y[idx].copy()]]
from dropout.mlp import build_f_df # here so theano not required for import
self.theano_f_df, self.model = build_f_df(layer_sizes, rectifier=rectifier,
use_bias=True)
self.theta_init = [param.get_value() for param in self.model.params]
def __init__(self, num_subfunctions=100, num_dims=10, rectifier='soft'):
self.name = 'MLP'
#layer_sizes = [ 28*28, 1200, 10 ]
#layer_sizes = [ 28*28, 1200, 10 ]
#layer_sizes = [ 28*28, 500, 120, num_dims ]
#layer_sizes = [ 28*28, 120, 12, num_dims ]
layer_sizes = [28 * 28, 1200, 1200, num_dims]
# Load data
X, y = load_mnist()
# break the data up into minibatches
self.subfunction_references = []
for mb in range(num_subfunctions):
self.subfunction_references.append(
[X[:, mb::num_subfunctions], y[mb::num_subfunctions]])
# evaluate on subset of training data
idx = random_choice(X.shape[1], 5000, replace=False)
#use all the training data for a smoother plot
#idx = np.array(range(X.shape[1]))
self.full_objective_references = [[X[:, idx].copy(), y[idx].copy()]]
from dropout.mlp import build_f_df # here so theano not required for import
self.theano_f_df, self.model = build_f_df(layer_sizes,
rectifier=rectifier,
use_bias=True)
self.theta_init = [param.get_value() for param in self.model.params]
