def run_alg(x_pub, y_pub, x_priv, y_priv, params, full_model_id):
##################################
# representation learning
#################################
x = x_pub
y = y_pub
# separate validation set if needed
val_x = None
#val_y = None
if validation_split:
logging.info("Splitting into training and validation sets")
from sklearn.model_selection import train_test_split
train_x, val_x, train_y, val_y = train_test_split(x, y, test_size=validation_split, random_state=0)
x, y = train_x, train_y
logging.info(" * training set shape: %d x %d" % x.shape)
logging.info(" * validation set shape: %d x %d" % val_x.shape)
data_dim = x.shape[1]
logging.info(" * data shape after preprocessing: %d x %d" % x.shape)
repr_dim = int(round(params.repr_dim))
logging.info("Learning the representation on public data...")
logging.info(" * learning a representation of size %d", repr_dim)
start_time = time.time()
# init the algorithm
#alg = make_alg(data_dim, repr_dim, num_classes)
#alg = make_alg(data_dim, repr_dim)
from models.vae_pytorch import VAE
alg = VAE().init(
input_dim = data_dim,
latent_dim = repr_dim,
#enc_dims = [],
enc_dims = [int(10 ** params.hidden_layer_size_mul_log10)*repr_dim] * int(params.n_hidden_layers),
dec_dims = 'same',
enc_activations = 'relu',
dec_activations = 'relu',
prediction_mean_activation = 'sigmoid',
prediction_var = 'gs',
prediction_log_var_min = math.log(0.01**2),
normalize_input_type = 'quantiles',
normalize_input_quantile = 0.05,
normalize_input_axis = 'global',
normalize_input_target = (0, 1),
normalize_input_clip = True,
optimizer = 'Adam',
optimizer_params = {'lr': 10.0 ** params.learning_rate_log10},
n_epochs = n_epochs,
early_stopping = True,
reduce_lr_on_plateau = False,
batch_size = batch_size)
# create output dir if does not exist
#ensure_dir_exists('res')
# define the progress saving function
ensure_dir_exists('param_opt/progress')
progress_filename = 'param_opt/progress/encdec-mse-%s.txt' % (full_model_id)
progress_file = open(progress_filename, 'w', encoding='utf-8')
#aux_progress_filename = 'param_opt/progress/aux-ce-%s.txt' % (full_model_id)
#aux_progress_file = open(aux_progress_filename, 'w', encoding='utf-8')
if val_x is not None:
val_progress_filename = 'param_opt/progress/encdec-validation-mse-%s.txt' % (full_model_id)
val_progress_file = open(val_progress_filename, 'w', encoding='utf-8')
#aux_val_progress_filename = 'param_opt/progress/aux-validation-ce-%s.txt' % (full_model_id)
#aux_val_progress_file = open(aux_val_progress_filename, 'w', encoding='utf-8')
def save_progress():
x_pred = alg.decode(alg.encode(x))
rel_mse = relative_mean_squared_error(x, x_pred)
progress_file.write("%g\n" % rel_mse)
#aux_pred = alg.predict_secondary(x)
#aux_rel_ce = relative_cross_entropy(y, aux_pred)
#aux_progress_file.write("%g\n" % aux_rel_ce)
if val_x is not None:
val_x_pred = alg.decode(alg.encode(val_x))
rel_mse = relative_mean_squared_error(val_x, val_x_pred)
val_progress_file.write("%g\n" % rel_mse)
#val_aux_pred = alg.predict_secondary(val_x)
#aux_rel_ce = relative_cross_entropy(val_y, val_aux_pred)
#aux_val_progress_file.write("%g\n" % aux_rel_ce)
# fit to the training data
ensure_dir_exists("param_opt/log/")
alg.learn(x, validation_data=val_x,
log_file_prefix=("param_opt/log/%s" % (full_model_id)),
per_epoch_callback_funs=[save_progress],
deadline=None, max_duration=repr_max_duration)
# test reconstruction error
x_pred = alg.decode(alg.encode(x))
rel_mse = relative_mean_squared_error(x, x_pred)
val_x_pred = alg.decode(alg.encode(val_x))
val_rel_mse = relative_mean_squared_error(val_x, val_x_pred)
logging.info(" * final error: rel_mse = %g, val_rel_mse = %g",
rel_mse, val_rel_mse)
elapsed = time.time() - start_time
logging.info(" * running time = %s", pretty_duration(elapsed))
##################################
# representation mapping
#################################
x = x_priv
y = y_priv
# get the representation
logging.info("Making the representation of private data...")
x_repr = alg.encode(x)
# test to predict the data itself
x_pred = alg.decode(x_repr)
rel_mse = relative_mean_squared_error(x, x_pred)
logging.info(" * reconstruct the data: rel_mse = %g", rel_mse)
##################################
# prediction
#################################
x = x_repr
# private or non-private logistic regression
private = True
# test prediction with cross validation
logging.info("Prediction with %d-fold cross validation...", pred_cv_folds)
from sklearn.model_selection import StratifiedKFold
cv = StratifiedKFold(n_splits=pred_cv_folds, shuffle=True, random_state=0)
avg_test_acc = 0
for fold, (train, test) in enumerate(cv.split(x, y)):
logging.info("Fold %d...", fold)
x_train, x_test, y_train, y_test = x[train], x[test], y[train], y[test]
# init rng
#np.random.seed(seed0)
logging.info("Bounding the data to 1-sphere...")
if scale_fun == "norm_max":
logging.info(" * scale by max norm")
scale_factor = np.amax(np.linalg.norm(x_train, axis=1))
elif scale_fun == "dims_max":
logging.info(" * scale each dimension by max absolute value")
scale_factor = np.amax(np.abs(x_train), axis=0)
elif scale_fun == "norm_avg":
logging.info(" * scale by average norm")
scale_factor = np.mean(np.linalg.norm(x_train, axis=1))
elif scale_fun == "dims_std":
logging.info(" * scale each dimension by standard deviation")
scale_factor = np.std(x_train, axis=0)
elif scale_fun == "none":
scale_factor = 1.0
else:
assert False
x_train /= scale_factor * scale_const
x_test /= scale_factor * scale_const
if clip == "norm":
logging.info(" * clip norms to max 1")
x_train /= np.maximum(np.linalg.norm(x_train, axis=1, keepdims=True) * (1 + bounding_slack), 1)
x_test /= np.maximum(np.linalg.norm(x_test, axis=1, keepdims=True) * (1 + bounding_slack),1)
elif clip == "dims":
assert False, "not implemented"
elif clip == "none":
logging.info(" * no clipping -> no bounding")
assert private == False #or np.isinf(epsilon)
else:
assert False
# fit
logging.info("Fitting a model...")
if private:
logging.info(" * DP logistic regression: epsilon=%g, alpha=%g", epsilon, regularizer_strength)
from models.logistic_regression import DPLogisticRegression
model = DPLogisticRegression().init(repr_dim, classes=np.unique(y),
alpha=regularizer_strength, epsilon=epsilon)
else:
logging.info(" * logistic regression: alpha=%g", regularizer_strength)
from sklearn.linear_model import LogisticRegression
model = LogisticRegression(C=1/regularizer_strength)
model.fit(x_train, y_train)
#print(model.predict(x_test))
# compute mean accuracy on test set
logging.info("Testing the model...")
#acc = model.score(x_test, y_test)
from sklearn.metrics import accuracy_score
train_acc = accuracy_score(y_train, model.predict(x_train))
test_acc = accuracy_score(y_test, model.predict(x_test))
logging.info(" * train accuracy = %.6f", train_acc)
logging.info(" * test accuracy = %.6f", test_acc)
avg_test_acc += test_acc
avg_test_acc /= pred_cv_folds
logging.info("Average test accuracy = %.6f", avg_test_acc)
return avg_test_acc
def task(args):
import pandas
repr_dim, (alg_id, _, make_alg), seed = args
logging.info("dataset = %s, algorithm = %s", data_set, alg_id)
# read the data sets
logging.info("Reading data...")
data = pandas.read_hdf("data/%s.h5" % (data_set), data_type)
logging.info(" * gene expression shape: %d x %d" % data.shape)
#aux_target = pandas.read_hdf("data/TCGA_cancertype.h5", 'cancer_types')
#logging.info(" * auxiliary target size: %d" % aux_target.shape)
#common_samples = data.index.intersection(aux_target.index)
#data = data.loc[common_samples]
#aux_target = aux_target.loc[common_samples]
#logging.info(" * number of common samples: %d" % common_samples.size)
from common import categorical_to_binary
x = data.as_matrix()
#y = categorical_to_binary(aux_target.values)
#num_classes = y.shape[1]
#x = x[:,0:2000]
# normalize the input to _total_ unit variance and per-feature zero mean
if normalize_data:
x -= np.mean(x)
x /= np.std(x)
x -= np.mean(x, axis=0)
# FIXME!
#x = (x - np.amin(x,axis=0)) / (np.amax(x,axis=0) - np.amin(x,axis=0))
#x = (x - np.amin(x)) / (np.amax(x) - np.amin(x))
# init rng
np.random.seed(seed)
import torch
torch.manual_seed(seed)
torch.cuda.manual_seed(seed)
#if args.cuda ?????:
# torch.cuda.manual_seed(seed)
# separate validation set if needed
val_x = None
#val_y = None
if validation_split:
logging.info("Splitting into training and validation sets")
m = x.shape[0]
perm = np.random.permutation(m)
x = x[perm, :]
#y = y[perm,:]
split_point = int(validation_split * m)
(val_x, x) = (x[:split_point, :], x[split_point:, :])
#(val_y, y) = (y[:split_point,:], y[split_point:,:])
logging.info(" * training set shape: %d x %d" % x.shape)
logging.info(" * validation set shape: %d x %d" % val_x.shape)
data_dim = x.shape[1]
logging.info(" * data shape after preprocessing: %d x %d" % x.shape)
logging.info("Running the algorithm...")
logging.info(" * learning a representation of size %d", repr_dim)
start_time = time.time()
# init the algorithm
#alg = make_alg(data_dim, repr_dim, num_classes)
alg = make_alg(data_dim, repr_dim)
# create output dir if does not exist
ensure_dir_exists('res')
full_model_id = "%s-%d-%s-s%d%s" % (data_set, repr_dim, alg_id, seed,
id_suffix)
# define the progress saving function
progress_filename = 'res/progress-encdec-mse-%s.txt' % (full_model_id)
progress_file = open(progress_filename, 'w', encoding='utf-8')
#aux_progress_filename = 'res/progress-aux-ce-%s.txt' % (full_model_id)
#aux_progress_file = open(aux_progress_filename, 'w', encoding='utf-8')
if val_x is not None:
val_progress_filename = 'res/progress-encdec-validation-mse-%s.txt' % (
full_model_id)
val_progress_file = open(val_progress_filename, 'w', encoding='utf-8')
#aux_val_progress_filename = 'res/progress-aux-validation-ce-%s.txt' % (full_model_id)
#aux_val_progress_file = open(aux_val_progress_filename, 'w', encoding='utf-8')
def save_progress():
x_pred = alg.decode(alg.encode(x))
rel_mse = relative_mean_squared_error(x, x_pred)
progress_file.write("%g\n" % rel_mse)
#aux_pred = alg.predict_secondary(x)
#aux_rel_ce = relative_cross_entropy(y, aux_pred)
#aux_progress_file.write("%g\n" % aux_rel_ce)
if val_x is not None:
val_x_pred = alg.decode(alg.encode(val_x))
rel_mse = relative_mean_squared_error(val_x, val_x_pred)
val_progress_file.write("%g\n" % rel_mse)
#val_aux_pred = alg.predict_secondary(val_x)
#aux_rel_ce = relative_cross_entropy(val_y, val_aux_pred)
#aux_val_progress_file.write("%g\n" % aux_rel_ce)
# fit to the training data
alg.learn(x,
validation_data=val_x,
log_file_prefix=("log/%s" % (full_model_id)),
per_epoch_callback_funs=[save_progress],
deadline=deadline,
max_duration=max_duration)
# test reconstruction error
x_pred = alg.decode(alg.encode(x))
rel_mse = relative_mean_squared_error(x, x_pred)
val_x_pred = alg.decode(alg.encode(val_x))
val_rel_mse = relative_mean_squared_error(val_x, val_x_pred)
logging.info(" * final error: rel_mse = %g, val_rel_mse = %g", rel_mse,
val_rel_mse)
elapsed = time.time() - start_time
logging.info(" * running time = %s", pretty_duration(elapsed))
# save model
logging.info("Saving the learned model...")
ensure_dir_exists('repr_models')
alg.save("repr_models/%s" % (full_model_id))
def learn_repr(x, y, params, full_model_id):
# separate validation set if needed
val_x = None
#val_y = None
if params.repr_learn_validation_split:
logging.info("Splitting into training and validation sets")
from sklearn.model_selection import train_test_split
train_x, val_x, train_y, val_y = train_test_split(
x, y, test_size=params.repr_learn_validation_split, random_state=0)
x, y = train_x, train_y
logging.info(" * training set shape: %d x %d" % x.shape)
logging.info(" * validation set shape: %d x %d" % val_x.shape)
data_dim = x.shape[1]
logging.info(" * data shape after preprocessing: %d x %d" % x.shape)
repr_dim = int(round(params.repr_dim))
logging.info("Learning the representation on public data...")
logging.info(" * learning a representation of size %d", repr_dim)
start_time = time.time()
(_, _, _, make_alg, _) = select_repr_alg(params.repr_alg)
# init the algorithm
#alg = make_alg(data_dim, repr_dim, num_classes)
#alg = make_alg(data_dim, repr_dim)
alg = make_alg(data_dim, repr_dim, params)
# create output dir if does not exist
#ensure_dir_exists('res')
# define the progress saving function
ensure_dir_exists('param_opt/progress')
progress_filename = 'param_opt/progress/encdec-mse-%s.txt' % (
full_model_id)
progress_file = open(progress_filename, 'w', encoding='utf-8')
#aux_progress_filename = 'param_opt/progress/aux-ce-%s.txt' % (full_model_id)
#aux_progress_file = open(aux_progress_filename, 'w', encoding='utf-8')
if val_x is not None:
val_progress_filename = 'param_opt/progress/encdec-validation-mse-%s.txt' % (
full_model_id)
val_progress_file = open(val_progress_filename, 'w', encoding='utf-8')
#aux_val_progress_filename = 'param_opt/progress/aux-validation-ce-%s.txt' % (full_model_id)
#aux_val_progress_file = open(aux_val_progress_filename, 'w', encoding='utf-8')
def save_progress():
x_pred = alg.decode(alg.encode(x))
rel_mse = relative_mean_squared_error(x, x_pred)
progress_file.write("%g\n" % rel_mse)
#aux_pred = alg.predict_secondary(x)
#aux_rel_ce = relative_cross_entropy(y, aux_pred)
#aux_progress_file.write("%g\n" % aux_rel_ce)
if val_x is not None:
val_x_pred = alg.decode(alg.encode(val_x))
rel_mse = relative_mean_squared_error(val_x, val_x_pred)
val_progress_file.write("%g\n" % rel_mse)
#val_aux_pred = alg.predict_secondary(val_x)
#aux_rel_ce = relative_cross_entropy(val_y, val_aux_pred)
#aux_val_progress_file.write("%g\n" % aux_rel_ce)
# fit to the training data
ensure_dir_exists("param_opt/log/")
alg.learn(x,
validation_data=val_x,
log_file_prefix=("param_opt/log/%s" % (full_model_id)),
per_epoch_callback_funs=[save_progress],
deadline=None,
max_duration=params.repr_learn_max_duration)
# test reconstruction error
x_pred = alg.decode(alg.encode(x))
rel_mse = relative_mean_squared_error(x, x_pred)
if val_x is not None:
val_x_pred = alg.decode(alg.encode(val_x))
val_rel_mse = relative_mean_squared_error(val_x, val_x_pred)
else:
val_rel_mse = np.nan
logging.info(" * final error: rel_mse = %g, val_rel_mse = %g", rel_mse,
val_rel_mse)
elapsed = time.time() - start_time
logging.info(" * running time = %s", pretty_duration(elapsed))
return alg
def task(args):
import pandas
param_id, priv_cancertypes, seed = args
logging.info("priv classes = %s, params_id = %s, seed = %d",
priv_cancertypes, param_id, seed)
#repr_dim, (alg_id, _, make_alg), seed = args
#logging.info("algorithm = %s, seed = %d", alg_id, seed)
# read the data sets
alg_id = param_id
logging.info("Loading parameters...")
params = np.load("run_parameters/params.npy")
params = params[param_id]
logging.info("Reading data...")
gene_expr = pandas.read_hdf("data/%s.h5" % (data_set), data_type)
logging.info(" * gene expression shape: %d x %d" % gene_expr.shape)
logging.info("Filtering out genes with low expressions...")
low_expr = (np.median(gene_expr, axis=0) < 0.0)
gene_expr = gene_expr.iloc[:, ~low_expr]
logging.info(" * %d of %d remaining (%d removed)" %
(sum(~low_expr), low_expr.size, sum(low_expr)))
logging.info("Loading cancer types...")
cancer_type = pandas.read_hdf("data/%s.h5" % (target_set), target_type)
assert np.array_equal(gene_expr.index, cancer_type.index)
# split
logging.info("Splitting...")
priv = cancer_type.isin(priv_cancertypes)
logging.info(" * %d private samples, %d public samples (of %d total)" %
(sum(priv), sum(~priv), priv.size))
from common import categorical_to_binary
x_pub = gene_expr[~priv].as_matrix()
y_pub = cancer_type[~priv].cat.codes.as_matrix()
x_priv = gene_expr[priv].as_matrix()
y_priv = cancer_type[priv].cat.codes.as_matrix()
#y = categorical_to_binary(aux_target.values)
#num_classes = y.shape[1]
data_name = '-'.join(priv_cancertypes).replace(' ', '_')
# A hack to have a different seed if the algorithm is run multiple times
# with the same parameters. Destroys reproducibility...
import time
seed0 = int(time.time() * 100) % (2**32)
# init rng
np.random.seed(seed0)
import torch
torch.manual_seed(seed0)
if torch.cuda.is_available() and torch.cuda.device_count() > 0:
torch.cuda.manual_seed(seed0)
##################################
# representation learning
#################################
x = x_pub
y = y_pub
# separate validation set if needed
val_x = None
#val_y = None
if validation_split:
logging.info("Splitting into training and validation sets")
from sklearn.model_selection import train_test_split
train_x, val_x, train_y, val_y = train_test_split(
x, y, test_size=validation_split, random_state=0)
x, y = train_x, train_y
#m = x.shape[0]
#perm = np.random.permutation(m)
#x = x[perm,:]
#y = y[perm,:]
#split_point = int(validation_split * m)
#(val_x, x) = (x[:split_point,:], x[split_point:,:])
#(val_y, y) = (y[:split_point,:], y[split_point:,:])
logging.info(" * training set shape: %d x %d" % x.shape)
logging.info(" * validation set shape: %d x %d" % val_x.shape)
data_dim = x.shape[1]
logging.info(" * data shape after preprocessing: %d x %d" % x.shape)
logging.info("Learning the representaiton on public data...")
logging.info(" * learning a representation of size %d", repr_dim)
start_time = time.time()
# init the algorithm
#alg = make_alg(data_dim, repr_dim, num_classes)
#alg = make_alg(data_dim, repr_dim)
from models.vae_pytorch import VAE
alg = VAE().init(
input_dim=data_dim,
latent_dim=repr_dim,
#enc_dims = [],
enc_dims=[int(10**params.hidden_layer_size_mul_log10) * repr_dim] *
int(params.n_hidden_layers),
dec_dims='same',
enc_activations='relu',
dec_activations='relu',
prediction_mean_activation='sigmoid',
prediction_var='gs',
prediction_log_var_min=math.log(0.01**2),
normalize_input_type='quantiles',
normalize_input_quantile=0.05,
normalize_input_axis='global',
normalize_input_target=(0, 1),
normalize_input_clip=True,
optimizer='Adam',
optimizer_params={'lr': 10.0**params.learning_rate_log10},
n_epochs=n_epochs,
early_stopping=True,
reduce_lr_on_plateau=False,
batch_size=batch_size)
# create output dir if does not exist
ensure_dir_exists('res')
full_model_id = "%s-%d-%s-s%d%s" % (data_name, repr_dim, alg_id, seed,
id_suffix)
# define the progress saving function
progress_filename = 'res/progress-encdec-mse-%s.txt' % (full_model_id)
progress_file = open(progress_filename, 'w', encoding='utf-8')
#aux_progress_filename = 'res/progress-aux-ce-%s.txt' % (full_model_id)
#aux_progress_file = open(aux_progress_filename, 'w', encoding='utf-8')
if val_x is not None:
val_progress_filename = 'res/progress-encdec-validation-mse-%s.txt' % (
full_model_id)
val_progress_file = open(val_progress_filename, 'w', encoding='utf-8')
#aux_val_progress_filename = 'res/progress-aux-validation-ce-%s.txt' % (full_model_id)
#aux_val_progress_file = open(aux_val_progress_filename, 'w', encoding='utf-8')
def save_progress():
x_pred = alg.decode(alg.encode(x))
rel_mse = relative_mean_squared_error(x, x_pred)
progress_file.write("%g\n" % rel_mse)
#aux_pred = alg.predict_secondary(x)
#aux_rel_ce = relative_cross_entropy(y, aux_pred)
#aux_progress_file.write("%g\n" % aux_rel_ce)
if val_x is not None:
val_x_pred = alg.decode(alg.encode(val_x))
rel_mse = relative_mean_squared_error(val_x, val_x_pred)
val_progress_file.write("%g\n" % rel_mse)
#val_aux_pred = alg.predict_secondary(val_x)
#aux_rel_ce = relative_cross_entropy(val_y, val_aux_pred)
#aux_val_progress_file.write("%g\n" % aux_rel_ce)
# fit to the training data
alg.learn(x,
validation_data=val_x,
log_file_prefix=("log/%s" % (full_model_id)),
per_epoch_callback_funs=[save_progress],
deadline=deadline,
max_duration=max_duration)
# test reconstruction error
x_pred = alg.decode(alg.encode(x))
rel_mse = relative_mean_squared_error(x, x_pred)
val_x_pred = alg.decode(alg.encode(val_x))
val_rel_mse = relative_mean_squared_error(val_x, val_x_pred)
logging.info(" * final error: rel_mse = %g, val_rel_mse = %g", rel_mse,
val_rel_mse)
elapsed = time.time() - start_time
logging.info(" * running time = %s", pretty_duration(elapsed))
# save model
#logging.info("Saving the learned model...")
#ensure_dir_exists('repr_models')
#alg.save("repr_models/%s" % (full_model_id))
##################################
# representation mapping
#################################
x = x_priv
y = y_priv
# get the representation
logging.info("Making the representation of private data...")
x_repr = alg.encode(x)
# test to predict the data itself
x_pred = alg.decode(x_repr)
rel_mse = relative_mean_squared_error(x, x_pred)
logging.info(" * reconstruct the data: rel_mse = %g", rel_mse)
ensure_dir_exists("res")
with open("res/private-encdec-rel_mse-%d-%s-%s-s%d%s.txt" %
(repr_dim, data_name, alg_id, seed, id_suffix),
'w',
encoding='utf-8') as f:
f.write("%.6f\n" % rel_mse)
# save the representation
#logging.info("Saving the representation...")
#ensure_dir_exists("data_repr")
#np.savetxt("data_repr/repr-%s-%d-%s-s%d%s.csv" %
# (data_name, repr_dim, alg_id, seed, id_suffix),
# x_repr, delimiter=',')
##################################
# prediction
#################################
x = x_repr
# split train and test sets
logging.info("Splitting to train and test sets...")
from sklearn.model_selection import train_test_split
x_train, x_test, y_train, y_test = train_test_split(
x, y, test_size=pred_test_size, random_state=0)
logging.info(" * train samples: %d" % x_train.shape[0])
logging.info(" * test samples: %d" % x_test.shape[0])
# init rng
np.random.seed(seed0)
#print(np.amax(np.linalg.norm(x_train, axis=1)))
#print(np.mean(np.linalg.norm(x_train, axis=1)))
logging.info("Bounding the data to 1-sphere...")
if scale_fun == "norm_max":
logging.info(" * scale by max norm")
scale_factor = np.amax(np.linalg.norm(x_train, axis=1))
elif scale_fun == "dims_max":
logging.info(" * scale each dimension by max absolute value")
scale_factor = np.amax(np.abs(x_train), axis=0)
elif scale_fun == "norm_avg":
logging.info(" * scale by average norm")
scale_factor = np.mean(np.linalg.norm(x_train, axis=1))
elif scale_fun == "dims_std":
logging.info(" * scale each dimension by standard deviation")
scale_factor = np.std(x_train, axis=0)
elif scale_fun == "none":
scale_factor = 1.0
else:
assert False
x_train /= scale_factor * scale_const
x_test /= scale_factor * scale_const
#print(np.amax(np.linalg.norm(x_train, axis=1, keepdims=True)))
if clip == "norm":
logging.info(" * clip norms to max 1")
x_train /= np.maximum(
np.linalg.norm(x_train, axis=1, keepdims=True) *
(1 + bounding_slack), 1)
x_test /= np.maximum(
np.linalg.norm(x_test, axis=1, keepdims=True) *
(1 + bounding_slack), 1)
elif clip == "dims":
assert False, "not implemented"
elif clip == "none":
logging.info(" * no clipping -> no bounding")
assert private == False #or np.isinf(epsilon)
else:
assert False
#for private in [False, True]:
for private in [True]:
# fit
logging.info("Fitting a model...")
if private:
logging.info(" * DP logistic regression: epsilon=%g, alpha=%g",
epsilon, regularizer_strength)
from models.logistic_regression import DPLogisticRegression
model = DPLogisticRegression().init(repr_dim,
classes=np.unique(y),
alpha=regularizer_strength,
epsilon=epsilon)
else:
logging.info(" * logistic regression: alpha=%g",
regularizer_strength)
from sklearn.linear_model import LogisticRegression
model = LogisticRegression(C=1 / regularizer_strength)
model.fit(x_train, y_train)
#print(model.predict(x_test))
# compute mean accuracy on test set
logging.info("Testing the model...")
#acc = model.score(x_test, y_test)
from sklearn.metrics import accuracy_score
train_acc = accuracy_score(y_train, model.predict(x_train))
test_acc = accuracy_score(y_test, model.predict(x_test))
logging.info(" * train accuracy = %.6f", train_acc)
logging.info(" * test accuracy = %.6f", test_acc)
logging.info("Writing results to disk...")
ensure_dir_exists("res")
filename = (
"res/cancertype-pred-accuracy-%d-%s-%s-s%d-%s-%d-%s%s.txt" %
(repr_dim, data_name, alg_id, seed, scale_fun, scale_const, clip,
("-e%g" % (epsilon) if private else "-nonpriv")))
logging.info(" * filename: %s", filename)
with open(filename, 'w', encoding='utf-8') as f:
f.write("%.6f\n" % test_acc)
filename = "param_opt/opt_result%s-%s.txt" % (id_suffix, full_model_id)
with open(filename, 'w', encoding='utf-8') as f:
f.write("%.6f\n" % test_acc)