Y_train = image_data_sets.train.labels * 2 - 1
train = DataSet(L_train, Y_train)
test = DataSet(L_test, Y_test)
data_sets = base.Datasets(train=train, validation=None, test=test)
input_dim = data_sets.train.x.shape[1]
# Train with hinge
rbf_model = SmoothHinge(
temp=0,
use_bias=use_bias,
input_dim=input_dim,
weight_decay=weight_decay,
num_classes=num_classes,
batch_size=batch_size,
data_sets=data_sets,
initial_learning_rate=initial_learning_rate,
keep_probs=keep_probs,
decay_epochs=decay_epochs,
mini_batch=False,
train_dir='output',
log_dir='log',
# model_name='dogfish_rbf_hinge_t-0')
model_name='catpanda_rbf_hinge_t-0')
rbf_model.train()
hinge_W = rbf_model.sess.run(rbf_model.params)[0]
# Then load weights into smoothed version
tf.reset_default_graph()
rbf_model = SmoothHinge(
temp=0.001,
num_pos_copies = int(total_copies / 2)
num_neg_copies = total_copies - num_pos_copies
weight_decay = 0.09
tf.reset_default_graph()
model = SmoothHinge(
input_dim=input_dim,
temp=temp,
weight_decay=weight_decay,
use_bias=True,
num_classes=num_classes,
batch_size=batch_size,
data_sets=data_sets,
initial_learning_rate=initial_learning_rate,
decay_epochs=None,
mini_batch=False,
train_dir=output_root,
log_dir='log',
model_name=
'smooth_hinge_%s_sphere-%s_slab-%s_start-copy_lflip-%s_step-%s_t-%s_eps-%s_wd-%s_rs-%s'
% (dataset_name, project_sphere, project_slab, label_flip,
step_size, temp, epsilon, weight_decay, random_seed))
X_modified, Y_modified = data.copy_random_points(
X_train,
Y_train,
target_class=1,
num_copies=num_pos_copies,
random_seed=random_seed,
num_temps = len(temps)
num_params = 784
# Get weights from hinge
tf.reset_default_graph()
temp = 0
model = SmoothHinge(use_bias=use_bias,
temp=temp,
input_dim=input_dim,
weight_decay=weight_decay,
num_classes=num_classes,
batch_size=batch_size,
data_sets=data_sets,
initial_learning_rate=initial_learning_rate,
keep_probs=keep_probs,
decay_epochs=decay_epochs,
mini_batch=False,
train_dir='output',
log_dir='log',
model_name='smooth_hinge_17_t-%s' % temp)
model.train()
model.load_checkpoint(iter_to_load=0)
hinge_W = model.sess.run(model.params)[0]
model_margins = model.sess.run(model.margin,
feed_dict=model.all_test_feed_dict)
# Look at np.argsort(model_margins)[:10] to pick a test example
def init_model(self):
"""
Initialize a tf model based on model_name and datasets
"""
# TODO: make it easier to use non-default hyperparams?
# we can always infer # classes of from the training data
num_classes = len(set(self.data_sets.train.labels))
model_name = self.task + '_' + self.model_name
print('Num classes', num_classes)
if self.model_name == 'binary_logistic':
#num_classes = 2
assert num_classes == 2
weight_decay = 0.0001
batch_size = 100
initial_learning_rate = 0.001
keep_probs = None
decay_epochs = [1000, 10000]
max_lbfgs_iter = 1000
self.model = BinaryLogisticRegressionWithLBFGS(
input_dim=self.input_dim,
weight_decay=weight_decay,
max_lbfgs_iter=max_lbfgs_iter,
num_classes=num_classes,
batch_size=batch_size,
data_sets=self.data_sets,
initial_learning_rate=initial_learning_rate,
keep_probs=keep_probs,
decay_epochs=decay_epochs,
mini_batch=False,
train_dir='output',
log_dir='log',
model_name=model_name
)
elif self.model_name == 'multi_logistic':
#num_classes = 10
weight_decay = 0.01
batch_size = 1400
initial_learning_rate = 0.001
keep_probs = None
max_lbfgs_iter = 1000
decay_epochs = [1000, 10000]
self.model = LogisticRegressionWithLBFGS(
input_dim=self.input_dim,
weight_decay=weight_decay,
max_lbfgs_iter=max_lbfgs_iter,
num_classes=num_classes,
batch_size=batch_size,
data_sets=self.data_sets,
initial_learning_rate=initial_learning_rate,
keep_probs=keep_probs,
decay_epochs=decay_epochs,
mini_batch=False,
train_dir='output',
log_dir='log',
model_name=model_name)
elif self.model_name == 'cnn':
assert num_classes == 10
weight_decay = 0.001
batch_size = 500
initial_learning_rate = 0.0001
decay_epochs = [10000, 20000]
hidden1_units = 8
hidden2_units = 8
hidden3_units = 8
conv_patch_size = 3
keep_probs = [1.0, 1.0]
self.model = All_CNN_C(
input_side=self.input_side,
input_channels=self.input_channels,
conv_patch_size=conv_patch_size,
hidden1_units=hidden1_units,
hidden2_units=hidden2_units,
hidden3_units=hidden3_units,
weight_decay=weight_decay,
num_classes=num_classes,
batch_size=batch_size,
data_sets=self.data_sets,
initial_learning_rate=initial_learning_rate,
damping=1e-2,
decay_epochs=decay_epochs,
mini_batch=True,
train_dir='output',
log_dir='log',
model_name=model_name
)
elif self.task == 'income':
num_classes = 2
input_dim = self.data_sets.train.x.shape[1]
weight_decay = 0.0001
# weight_decay = 1000 / len(lr_data_sets.train.labels)
batch_size = 10
initial_learning_rate = 0.001
keep_probs = None
decay_epochs = [1000, 10000]
max_lbfgs_iter = 1000
self.model = BinaryLogisticRegressionWithLBFGS(
input_dim=input_dim,
weight_decay=weight_decay,
max_lbfgs_iter=max_lbfgs_iter,
num_classes=num_classes,
batch_size=batch_size,
data_sets=self.data_sets,
initial_learning_rate=initial_learning_rate,
keep_probs=keep_probs,
decay_epochs=decay_epochs,
mini_batch=False,
train_dir='output',
log_dir='log',
model_name='income_logreg'
)
elif self.model_name == 'hinge_svm':
#num_classes = 2
weight_decay = 0.01
use_bias = False
batch_size = 100
initial_learning_rate = 0.001
keep_probs = None
decay_epochs = [1000, 10000]
temps = [0, 0.001, 0.1]
num_temps = len(temps)
num_params = 784
temp = 0
self.model = SmoothHinge(
use_bias=use_bias,
temp=temp,
input_dim=self.input_dim,
weight_decay=weight_decay,
num_classes=num_classes,
batch_size=batch_size,
data_sets=self.data_sets,
initial_learning_rate=initial_learning_rate,
keep_probs=keep_probs,
decay_epochs=decay_epochs,
mini_batch=False,
train_dir='output',
log_dir='log',
model_name='smooth_hinge_17_t-%s' % temp)