def set_trainer(config):
# load a checkpoint
if config.checkpoint is not None:
# load data
train_loader = load_data(config, 'train', False)
model, optimizer, word_map, start_epoch = load_checkpoint(config.checkpoint, device)
print('\nLoaded checkpoint from epoch %d.\n' % (start_epoch - 1))
# or initialize model
else:
start_epoch = 0
# load data
train_loader, embeddings, emb_size, word_map, n_classes, vocab_size = load_data(config, 'train', True)
model = models.setup(
config = config,
n_classes = n_classes,
vocab_size = vocab_size,
embeddings = embeddings,
emb_size = emb_size
)
optimizer = optim.Adam(
params = filter(lambda p: p.requires_grad, model.parameters()),
lr = config.lr
)
# loss functions
loss_function = nn.CrossEntropyLoss()
# move to device
model = model.to(device)
loss_function = loss_function.to(device)
trainer = Trainer(
num_epochs = config.num_epochs,
start_epoch = start_epoch,
train_loader = train_loader,
model = model,
model_name = config.model_name,
loss_function = loss_function,
optimizer = optimizer,
lr_decay = config.lr_decay,
word_map = word_map,
grad_clip = config.grad_clip,
print_freq = config.print_freq,
checkpoint_path = config.checkpoint_path,
checkpoint_basename = config.checkpoint_basename
)
return trainer
def main() -> None:
parser = argparse.ArgumentParser(description="Flower")
parser.add_argument(
"--server_address",
type=str,
default=DEFAULT_SERVER_ADDRESS,
help=f"gRPC server address (default: {DEFAULT_SERVER_ADDRESS})",
)
parser.add_argument(
"--cid", type=str, required=True, help="Client CID (no default)"
)
parser.add_argument(
"--log_host", type=str, help="Logserver address (no default)",
)
args = parser.parse_args()
# Configure logger
fl.common.logger.configure(f"client_{args.cid}", host=args.log_host)
# Load model and data
model = models.load_model(model_name=glb.MODEL, framework="PT")
model.to(DEVICE)
trainset, testset = datasets.load_data(dataset_name=glb.DATASET, framework="PT")
# Start client
client = PyTorchClient(args.cid, model, trainset, testset)
try:
fl.client.start_client(args.server_address, client)
except:
print("Either something went wrong or server finished execution!!")
def main():
model1 = 'vae'
model2 = 'dcgan'
dataset1 = ''
epoch = 200
batchsize = 50
output1 = 'output'
zdims = 256
# Make output direcotiry if not exists
if not os.path.isdir(output1):
os.mkdir(output1)
datasets = load_data(dataset1)
# Construct model
if model1 not in models:
raise Exception('Unknown model:', model1)
model = models[model1](input_shape=datasets.shape[1:],
z_dims=zdims,
output=output1)
# Training loop
datasets = datasets.images * 2.0 - 1.0
samples = np.random.normal(size=(100, zdims)).astype(np.float32)
model.main_loop(datasets,
samples,
epochs=epoch,
batchsize=batchsize,
reporter=['loss', 'g_loss', 'd_loss', 'g_acc', 'd_acc'])
def train(self,
tabular_path: str,
join_result_path: str,
model_path: str,
model_weights_path=None,
histogram_path=None) -> None:
"""
Train a classification model for spatial join cost estimator, then save the trained model to file
"""
# Extract train and test data, but only use train data
X_train, y_train = datasets.load_data(tabular_path,
RankingModel.TARGET,
RankingModel.DROP_COLUMNS)
X_train, X_val, y_train, y_val = train_test_split(X_train,
y_train,
test_size=0.2,
random_state=1)
query_train = [X_train.shape[0]]
query_val = [X_val.shape[0]]
gbm = lgb.LGBMRanker()
model = gbm.fit(X_train,
y_train,
group=query_train,
eval_set=[(X_val, y_val)],
eval_group=[query_val],
eval_at=[1, 2],
early_stopping_rounds=50)
# Fit and save the model
# model = self.rnk_model.fit(X_train, y_train)
pickle.dump(model, open(model_path, 'wb'))
def load_data(data_type):
if data_type == 'mnist':
return dsets.mnist.load_data()
elif data_type == 'svhn':
return dsets.svhn.load_data()
else:
return dsets.load_data(data_type)
def train(self,
tabular_path: str,
join_result_path: str,
model_path: str,
model_weights_path=None,
histogram_path=None) -> None:
"""
Train a classification model for spatial join cost estimator, then save the trained model to file
"""
# Extract train and test data, but only use train data
X_train, y_train, join_df = datasets.load_data(
tabular_path, ClassificationModel.TARGET,
ClassificationModel.DROP_COLUMNS,
ClassificationModel.SELECTED_COLUMNS)
# Fit and save the model
model = self.clf_model.fit(X_train, y_train)
pickle.dump(model, open(model_path, 'wb'))
# Feature importances
importances = model.feature_importances_
output_f = open('data/temp/feature_importances.csv', 'w')
output_f.writelines('feature_name,importance_score\n')
for fname, fscore in zip(ClassificationModel.SELECTED_COLUMNS,
importances):
print('{},{}'.format(fname, fscore))
output_f.writelines('{},{}\n'.format(fname, fscore))
output_f.close()
def plot_1d_experiment(params, output_path):
gp, gp_list, s_list, loss_list = pickle_data(
get_folder_name(params, output_path))
data, X_range = datasets.load_data(params["data_name"])
print X_range.shape
h, w = X_range.shape
for i in range(len(gp_list)):
gpi = gp_list[i]
gpi.Ytrain = -gpi.Ytrain
gpi.Ytrain_original = -gpi.Ytrain_original
gpi.mean = -gpi.mean
X_new, Y_new = s_list[i]
Y_new = -Y_new
plt.title("Iteration: " + str((i + 1) * params["save_every"]))
plt.xlabel('x')
plt.ylabel('y')
mean, cov = gpi.predict(X_range)
var = cov.diagonal()[:, np.newaxis]
plt.scatter(gpi.Xtrain,
gpi.Ytrain_original,
color='green',
marker='x',
s=50) # training data
plt.plot(X_range, mean, color='blue') # GP mean
plt.plot(X_range, mean + var, color='red') # GP mean
plt.plot(X_range, mean - var, color='red') # GP mean
plt.plot(X_range, 5 + 3 * data, color='black') # GP mean - std
plt.scatter(X_new, Y_new, color='purple', marker='*', s=100) # data
plt.xlim(-3 * np.pi, 3 * np.pi)
plt.show()
print X_new, Y_new
def run_experiment(params, output_path):
folder_name = get_folder_name(params, output_path)
if not os.path.exists(folder_name):
os.makedirs(folder_name)
print "Started EXPERIMENTS for", folder_name
k = 0
for p in params["first_points"]:
curr_time = current_milli_time() % 100000000000
out_file = open(folder_name + str(curr_time) + "_stdout", 'w')
save_stdout = sys.stdout
sys.stdout = out_file
data, X_range, X, Y, Z, is_us, goal_func = datasets.load_data(params["data_name"], True)
if params["acqui_name"] == "lookahead":
w_s, w_e = params["widths_range"]
a_s, a_e = params["amps_range"]
n_s, n_e = params["noise_range"]
widths = np.logspace(np.log(w_s), np.log(w_e), num=10, base=np.e)
amps = np.logspace(np.log(a_s), np.log(a_e), num=5, base=np.e)
noises = np.logspace(np.log(n_s), np.log(n_e), num=5, base=np.e)
# initialize gaussian process
gp = gaussian_process.GP_Regressor(params["dimensions"], params["noise"], params["cov_func"],
params["cov_grad"], [params["width"], params["amp"]])
# print 'Max location', np.unravel_index(np.argmax(data), data.shape)
#add first point
X_new = X_range[p, :][np.newaxis, :]
Y_new = goal_func(X_new)
gp.update(X_new, Y_new)
if params["acqui_name"] == "lookahead":
goal_func = partial(goal_func, sign=-1)
gp, gp_list, s_list, aux_list = GP_lookahead.lookahead_optimization(
gp,
params["iterations"],
goal_func,
X_range,
[widths, amps, noises],
lookahead_steps=params["steps"],
save_every=params["save_every"])
else:
acqui_func = acquisition_functions.get_function(params["acqui_name"], kappa=params["kappa"], gamma=params["gamma"])
# execute bayesian optimization
gp, gp_list, s_list, aux_list = bayesian_optimization.bayesian_optimization(gp, params["iterations"], goal_func,
acqui_func, X_range, params["bounds"],
save_every=params["save_every"])
pickle_data(folder_name, data=[gp, gp_list, s_list, aux_list], curr_time=curr_time)
sys.stdout = save_stdout
out_file.close()
k += 1
print "Experiment", k, "of", len(params["first_points"]), "done"
def train(self,
tabular_path: str,
join_result_path: str,
model_path: str,
model_weights_path=None,
histogram_path=None) -> None:
"""
Train a regression model for spatial join cost estimator, then save the trained model to file
"""
# Extract train and test data, but only use train data
# X_train, y_train, X_test, y_test = datasets.load_tabular_features_hadoop(RegressionModel.DISTRIBUTION, RegressionModel.MATCHED, RegressionModel.SCALE, RegressionModel.MINUS_ONE)
# X_train, y_train, X_test, y_test = datasets.load_tabular_features(join_result_path, tabular_path, RegressionModel.NORMALIZE, RegressionModel.MINUS_ONE, RegressionModel.TARGET)
X_train, y_train, join_df = datasets.load_data(
tabular_path, RegressionModel.TARGET, RegressionModel.DROP_COLUMNS,
RegressionModel.SELECTED_COLUMNS)
# X_train, X_val, y_train, y_val = train_test_split(X_train, y_train, test_size=0.2, random_state=1)
# query_val = [X_val.shape[0]]
#
# Fit and save the model
model = self.reg_model.fit(X_train, y_train)
# model = LGBMRanker()
# model.fit(X_train, y_train, eval_set=[(X_val, y_val)], eval_group=[query_val], eval_at=[1, 2],
# early_stopping_rounds=50)
pickle.dump(model, open(model_path, 'wb'))
def test(self,
tabular_path: str,
join_result_path: str,
model_path: str,
model_weights_path=None,
histogram_path=None) -> (float, float, float, float):
"""
Evaluate the accuracy metrics of a trained model for spatial join cost estimator
:return mean_squared_error, mean_absolute_percentage_error, mean_squared_logarithmic_error, mean_absolute_error
"""
# Extract train and test data, but only use test data
# X_train, y_train, X_test, y_test = datasets.load_tabular_features_hadoop(RegressionModel.DISTRIBUTION, RegressionModel.MATCHED, RegressionModel.SCALE, RegressionModel.MINUS_ONE)
# X_train, y_train, X_test, y_test = datasets.load_tabular_features(join_result_path, tabular_path, RegressionModel.NORMALIZE, RegressionModel.MINUS_ONE, RegressionModel.TARGET)
X_test, y_test = datasets.load_data(tabular_path, RankingModel.TARGET,
RankingModel.DROP_COLUMNS)
# Load the model and use it for prediction
loaded_model = pickle.load(open(model_path, 'rb'))
y_pred = loaded_model.predict(X_test)
# TODO: delete this dumping action. This is just for debugging
test_df = pd.DataFrame()
test_df['y_test'] = y_test
test_df['y_pred'] = y_pred
test_df.to_csv('data/temp/test_df.csv')
# Compute accuracy metrics
# ndcg = metrics.ndcg_score(y_test, y_pred)
# acc = metrics.accuracy_score(y_test, y_pred)
# print('Accuracy:', metrics.accuracy_score(y_test, y_pred))
ndcg = 0
return ndcg, ndcg, ndcg, ndcg
def main(args):
# 加载数据
x, y = load_data(args.dataset)
n_clusters = len(np.unique(y))
# 设置参数
if args.dataset == 'mnist' or args.dataset == 'fmnist':
args.update_interval = 140
args.pretrain_epochs = 301
ae_weights_init = tf.variance_scaling_initializer(scale=1. / 3., mode='fan_in', distribution='uniform')
# add feature dimension size to the beginning of hidden_dims
feature_dim = x.shape[1]
args.encoder_dims = [feature_dim] + args.encoder_dims
print(args.encoder_dims)
if args.pretrain == True:
# 预训练
print('Begin Pretraining')
t0 = time()
pretrainer = Pretrainer(args, ae_weights_init)
saver = pretrainer(x, y)
# print(saver)
print('Pretraining time: %ds' % round(time() - t0))
# 清理计算图
tf.reset_default_graph()
# Model训练
print('Begin Model training')
t1 = time()
trainer = Trainer(args, ae_weights_init, n_clusters)
trainer(x, y)
print('Model training time: %ds' % round(time() - t1))
def eval(config):
pl.seed_everything(config["seed"])
config["comment"] = f"{config['comment']}-seed-{config['seed']}"
config["n_splits"] = 99999 if config["times"]<0 and config["mode"]=="leave one" else config["n_splits"]
config["times"] = config["times"] if config["times"]>=0 else config["n_splits"]
dataset = load_data(**config)
predicts = []
labels = []
edges_index = []
split_idxs = []
for i, data in zip(range(config["times"]), dataset):
if config["times"]!=config["n_splits"] and config["times"]!=i+1:
continue
model, log_dir = build_model(config, data)
predict, label, edge_index = model.test_step(data, from_neighbor=config["from_neighbor"])
predicts.append(predict)
labels.append(label)
edges_index.append(edge_index)
split_idxs.append(torch.ones(len(predict), dtype=torch.int)*i)
anaylse_result(predict, label, edge_index,
dataset, split_idxs[-1], model, log_dir,
save=config.get("save", False),
tag=f"split_{i}")
model.info(f"split {i} end")
model.info(f"{model}")
model.info(f"{config}")
predicts = torch.cat(predicts)
labels = torch.cat(labels)
edges_index = torch.cat(edges_index, dim=-1)
split_idxs = torch.cat(split_idxs)
def main():
# Parsing arguments
parser = argparse.ArgumentParser(description='Training GANs or VAEs')
parser.add_argument('--model', type=str, required=True)
parser.add_argument('--dataset', type=str, required=True)
parser.add_argument('--epoch', type=int, default=200)
parser.add_argument('--batchsize', type=int, default=50)
parser.add_argument('--datasize', type=int, default=-1)
parser.add_argument('--output', default='output')
parser.add_argument('--zdims', type=int, default=256)
parser.add_argument('--gpu', type=int, default=0)
parser.add_argument('--resume', type=str, default=None)
parser.add_argument('--testmode', action='store_true')
args = parser.parse_args()
# Select GPU
os.environ['CUDA_VISIBLE_DEVICES'] = str(args.gpu)
# Make output direcotiry if not exists
if not os.path.isdir(args.output):
os.mkdir(args.output)
# Load datasets
if args.dataset == 'mnist':
datasets = mnist.load_data()
elif args.dataset == 'svhn':
datasets = svhn.load_data()
elif args.dataset == 'hair':
datasets = hairdata.load_data()
elif args.dataset == 'hair_4tags':
datasets = hairdata_4tags.load_data()
else:
datasets = load_data(args.dataset)
print('aa')
# Construct model
if args.model not in models:
raise Exception('Unknown model:', args.model)
model = models[args.model](input_shape=datasets.images.shape[1:],
num_attrs=len(datasets.attr_names),
z_dims=args.zdims,
output=args.output)
if args.resume is not None:
model.load_model(args.resume)
# Training loop
datasets.images = datasets.images * 2.0 - 1.0
samples = np.random.normal(size=(10, args.zdims)).astype(np.float32)
model.main_loop(datasets,
samples,
datasets.attr_names,
epochs=args.epoch,
batchsize=args.batchsize,
reporter=[
'loss', 'g_loss', 'd_loss', 'g_acc', 'd_acc', 'c_loss',
'ae_loss'
])
def test(self,
tabular_path: str,
join_result_path: str,
model_path: str,
model_weights_path=None,
histogram_path=None) -> (float, float, float, float):
"""
Evaluate the accuracy metrics of a trained model for spatial join cost estimator
:return mean_squared_error, mean_absolute_percentage_error, mean_squared_logarithmic_error, mean_absolute_error
"""
# Extract train and test data, but only use test data
# X_train, y_train, X_test, y_test = datasets.load_tabular_features_hadoop(RegressionModel.DISTRIBUTION, RegressionModel.MATCHED, RegressionModel.SCALE, RegressionModel.MINUS_ONE)
# X_train, y_train, X_test, y_test = datasets.load_tabular_features(join_result_path, tabular_path, RegressionModel.NORMALIZE, RegressionModel.MINUS_ONE, RegressionModel.TARGET)
X_test, y_test, join_df = datasets.load_data(
tabular_path, ClassificationModel.TARGET,
ClassificationModel.DROP_COLUMNS,
ClassificationModel.SELECTED_COLUMNS)
# Load the model and use it for prediction
loaded_model = pickle.load(open(model_path, 'rb'))
y_pred = loaded_model.predict(X_test)
# TODO: delete this dumping action. This is just for debugging
test_df = pd.DataFrame()
test_df['dataset1'] = join_df['dataset1']
test_df['dataset2'] = join_df['dataset2']
test_df['y_test'] = y_test
test_df['y_pred'] = y_pred
test_df.to_csv('data/temp/test_df.csv', index=None)
# Compute accuracy metrics
acc = metrics.accuracy_score(y_test, y_pred)
print('Accuracy:', metrics.accuracy_score(y_test, y_pred))
# Plot non-normalized confusion matrix
titles_options = [
("figures/confusion_matrix_without_normalization.png", None),
("figures/confusion_matrix_with_normalization.png", 'true')
]
class_names = ['BNLJ', 'PBSM', 'DJ', 'RepJ']
for title, normalize in titles_options:
plt.rcParams.update({'font.size': 14})
disp = plot_confusion_matrix(loaded_model,
X_test,
y_test,
display_labels=class_names,
cmap=plt.cm.Blues,
normalize=normalize)
disp.ax_.set_title("")
print(title)
print(disp.confusion_matrix)
plt.xlabel('Predicted algorithm', fontsize=16)
plt.ylabel('Actual best algorithm', fontsize=16)
plt.savefig(title)
return acc, acc, acc, acc
def main() -> None:
parser = argparse.ArgumentParser(description="Flower")
parser.add_argument(
"--server_address",
type=str,
default=DEFAULT_SERVER_ADDRESS,
help=f"gRPC server address (default: {DEFAULT_SERVER_ADDRESS})",
)
parser.add_argument("--cid",
type=str,
required=True,
help="Client CID (no default)")
parser.add_argument(
"--model",
type=str,
default="simple-cnn",
help="Model to use for training (default: simple-cnn)",
)
parser.add_argument(
"--dataset",
type=str,
default="cifar-10",
help="Dataset to use fro training (default: cifar-10)",
)
parser.add_argument(
"--device",
type=str,
default="CPU",
help="Device to run the model on (default: CPU)",
)
parser.add_argument(
"--log_host",
type=str,
help="Logserver address (no default)",
)
args = parser.parse_args()
# Configure logger
fl.common.logger.configure(f"client_{args.cid}", host=args.log_host)
# check for runnable device
global DEVICE
DEVICE = torch.device("cuda:0" if args.device == "GPU" else "cpu")
# Load model and data
model = models.load_model(model_name=args.model, framework="PT")
model.to(DEVICE)
trainset, testset = datasets.load_data(dataset_name=args.dataset,
framework="PT")
# Start client
client = PyTorchClient(args.cid, model, trainset, testset)
try:
fl.client.start_client(args.server_address, client)
except:
print("Either something went wrong or server finished execution!!")
def plot_experiment(params, output_path):
folder_name = get_folder_name(params, output_path)
all_experiments = set([f[:11] for f in os.listdir(folder_name) if utils.check_int(f[:11])])
all_experiments = sorted(all_experiments)
for experiment in all_experiments:
gp, gp_list, s_list = pickle_data(folder_name, file_prefix=experiment)
data, X_range, X, Y, Z, is_us, goal_func = datasets.load_data(params["data_name"], True)
print X_range.shape
h, w = X_range.shape
for i in range(len(gp_list)):
gpi = gp_list[i]
X_new, Y_new = s_list[i]
print X_new.shape
mean = np.zeros((X_range.shape[0], 1))
var = np.zeros((X_range.shape[0], 1))
for j in range(X_range.shape[0]):
mean[j], var[j] = gpi.predict(X_range[j, :][np.newaxis, :])
print gpi.Xtrain.shape
print gpi.Ytrain.shape
print Z.shape, is_us.shape, mean.flatten().shape, var.flatten().shape
Z[is_us] = mean.flatten()
p.contourf(X, Y, Z) # GP mean
p.colorbar()
p.scatter(gpi.Xtrain[:, 0], gpi.Xtrain[:, 1], color='green', marker='x', s=50) # training data
p.scatter(X_new[:, 0], X_new[:, 1], color='purple', marker='*', s=100) # test data
p.show()
Z[is_us] = var.flatten() ** 0.5
p.contourf(X, Y, Z) # GP mean
p.colorbar()
p.scatter(gpi.Xtrain[:, 0], gpi.Xtrain[:, 1], color='green', marker='x', s=50) # training data
p.scatter(X_new[:, 0], X_new[:, 1], color='purple', marker='*', s=100) # test data
p.show()
Z[is_us] = data.flatten()
p.contourf(X, Y, Z) # GP mean
p.colorbar()
p.scatter(gpi.Xtrain[:, 0], gpi.Xtrain[:, 1], color='green', marker='x', s=50) # training data
p.scatter(X_new[:, 0], X_new[:, 1], color='purple', marker='*', s=100) # test data
p.show()
Z[is_us] = mean.flatten()
fig = p.figure()
ax = fig.gca(projection='3d')
# surf = ax.plot_surface(X,Y,Z, rstride=1, cstride=1, cmap=cm.coolwarm,
# linewidth=0, antialiased=False)
ax.contourf(X, Y, Z)
# fig.colorbar(surf, shrink=0.5, aspect=5)
ax.set_zlim(-15, 90)
# p.colorbar()
ax.scatter(gpi.Xtrain[:, 0], gpi.Xtrain[:, 1], gpi.Ytrain_original, color='green', marker='x',
s=50) # training data
# p.scatter(X_new[:,0],X_new[:,1],color='purple',marker='*', s=100) # test data
p.show()
print X_new, Y_new
def main():
print('Prepare dataset')
# Dataset
data_train, data_valid, data_test = datasets.load_data(
args.dataset, args.data_path, args.representation, args.normalization)
print('Create model')
if args.representation == 'adj':
print('\t* Discrete Edges')
net = models.MpnnGGNN(in_size=2,
e=[1],
hidden_state_size=args.hidden_size,
message_size=args.hidden_size,
n_layers=args.nlayers,
discrete_edge=True,
out_type='regression',
target_size=data_train.getTargetSize())
elif args.representation == 'feat':
print('\t* Feature Edges')
net = models.MpnnGGNN(in_size=2,
e=2,
hidden_state_size=args.hidden_size,
message_size=args.hidden_size,
n_layers=args.nlayers,
discrete_edge=False,
out_type='regression',
target_size=data_train.getTargetSize())
else:
raise NameError('Representation ' + args.representation +
' not implemented!')
print('Check CUDA')
if args.cuda and args.ngpu > 1:
print('\t* Data Parallel **NOT TESTED**')
net = torch.nn.DataParallel(net, device_ids=list(range(args.ngpu)))
if args.cuda:
print('\t* CUDA')
net = net.cuda()
if args.load is not None:
print('Loading model')
checkpoint = load_checkpoint(args.load)
net.load_state_dict(checkpoint['state_dict'])
start_epoch = checkpoint['epoch']
best_acc = checkpoint['best_acc']
else:
raise NameError('Load path must be set!')
# Train
plot_dataset(data_train, net, args.ngpu > 0)
# Validation
plot_dataset(data_valid, net, args.ngpu > 0)
# Test
plot_dataset(data_test, net, args.ngpu > 0)
def experiment(options):
dataset_name = options['dataset_name']
urep_class = options['urep_class']
urep_ratio = options['urep_ratio']
use_validation_step = options['use_validation_step']
train_size = options['train_size']
test_size = options['test_size']
n_splits = options['n_splits']
X, y = datasets.load_data(dataset_name)
y = y.astype(int)
n_classes = y.max() + 1
if urep_class is not None:
test_data_imbalance(y, urep_class)
eval = [
0,
np.zeros(n_classes),
np.zeros(n_classes), 0,
np.zeros(n_classes),
np.zeros(n_classes)
]
if use_validation_step:
sss = StratifiedShuffleSplit(n_splits=1,
test_size=test_size,
train_size=train_size)
train_index, test_index = next(sss.split(X, y))
val_index = np.delete(np.arange(X.shape[0]),
np.concatenate((train_index, test_index)))
X_train, y_train = shuffle(X[train_index], y[train_index])
X_val, y_val = shuffle(X[val_index], y[val_index])
X_test, y_test = shuffle(X[test_index], y[test_index])
if urep_class is not None and urep_ratio is not None:
X_train, y_train = force_data_imbalance(X_train, y_train,
urep_class, urep_ratio)
test_data_imbalance(y_train, urep_class)
eval = train(X_train, y_train, X_val, y_val, X_test, y_test, options)
else:
skf = StratifiedKFold(n_splits=n_splits)
for train_index, test_index in skf.split(X, y):
X_train, y_train = shuffle(X[train_index], y[train_index])
X_test, y_test = shuffle(X[test_index], y[test_index])
if urep_class is not None and urep_ratio is not None:
X_train, y_train = force_data_imbalance(
X_train, y_train, urep_class, urep_ratio)
test_data_imbalance(y_train, urep_class)
split_eval = train(X_train, y_train, None, None, X_test, y_test,
options)
for i in range(6):
eval[i] += split_eval[i] / n_splits
for perf in eval:
print(perf)
return eval
def _get_data_and_model(args):
# prepare dataset
if args.method in ['FcDEC', 'FcIDEC', 'FcDEC-DA', 'FcIDEC-DA']:
x, y = load_data(args.dataset, args.subset_key)
elif args.method in ['ConvDEC', 'ConvIDEC', 'ConvDEC-DA', 'ConvIDEC-DA']:
x, y = load_data_conv(args.dataset, args.subset_key)
else:
raise ValueError(
"Invalid value for method, which can only be in ['FcDEC', 'FcIDEC', 'ConvDEC', 'ConvIDEC', "
"'FcDEC-DA', 'FcIDEC-DA', 'ConvDEC-DA', 'ConvIDEC-DA']")
# prepare optimizer
if args.optimizer in ['sgd', 'SGD']:
optimizer = SGD(args.lr, 0.9)
else:
optimizer = Adam()
# prepare the model
if y is None:
n_clusters = args.n_clusters
else:
n_clusters = len(np.unique(y))
if 'FcDEC' in args.method:
model = FcDEC(dims=[x.shape[-1], 500, 500, 2000, 10],
n_clusters=n_clusters)
model.compile(optimizer=optimizer, loss='kld')
elif 'FcIDEC' in args.method:
model = FcIDEC(dims=[x.shape[-1], 500, 500, 2000, 10],
n_clusters=n_clusters)
model.compile(optimizer=optimizer,
loss=['kld', 'mse'],
loss_weights=[0.1, 1.0])
elif 'ConvDEC' in args.method:
model = ConvDEC(input_shape=x.shape[1:],
filters=[32, 64, 128, 10],
n_clusters=n_clusters)
model.compile(optimizer=optimizer, loss='kld')
elif 'ConvIDEC' in args.method:
model = ConvIDEC(input_shape=x.shape[1:],
filters=[32, 64, 128, 10],
n_clusters=n_clusters)
model.compile(optimizer=optimizer,
loss=['kld', 'mse'],
loss_weights=[0.1, 1.0])
else:
raise ValueError(
"Invalid value for method, which can only be in ['FcDEC', 'FcIDEC', 'ConvDEC', 'ConvIDEC', "
"'FcDEC-DA', 'FcIDEC-DA', 'ConvDEC-DA', 'ConvIDEC-DA']")
# if -DA method, we'll force aug_pretrain and aug_cluster is True
if '-DA' in args.method:
args.aug_pretrain = True
args.aug_cluster = True
return (x, y), model
def divide_data(options):
data, _, _, _ = load_data((options.dim, options.dim), options.data_path)
random.shuffle(data)
length = len(data)
bound = int(length * options.data_prop)
for i in range(bound):
scipy.misc.imsave('../data/train/train_{0}.jpg'.format(i), data[i])
for i in range(bound, length):
scipy.misc.imsave('../data/test/test_{0}.jpg'.format(i - bound),
data[i])
def main():
parser = argparse.ArgumentParser()
parser.add_argument("dataset")
parser.add_argument(
"--t_zero",
default=None,
type=int,
help="Last task identifier before the first evaluation task identifier"
)
parser.add_argument("--history", default=0, type=int, help="History size")
parser.add_argument("--backend",
default="dgl",
type=str,
choices=["dgl", "geometric"])
parser.add_argument(
"--basedir",
help=
"Basedir for preprocessed dataset, else create subdirectory in input")
args = parser.parse_args()
graph_or_edge_index, features, labels, years = load_data(
args.dataset, backend=args.backend)
basedir = args.basedir if args.basedir else args.dataset
outdir = os.path.join(
basedir,
lifelong_nodeclf_identifier(args.dataset, args.t_zero, args.history,
args.backend))
# Cast to torch tensors
features = torch.as_tensor(features, dtype=torch.float)
labels = torch.as_tensor(labels, dtype=torch.long)
years = torch.as_tensor(years, dtype=torch.long)
if args.backend == "geometric":
dataset = make_lifelong_nodeclf_dataset(outdir,
years,
features,
labels,
edge_index=graph_or_edge_index,
t_zero=args.t_zero,
cumulate=args.history)
elif args.backend == 'dgl':
dataset = make_lifelong_nodeclf_dataset(outdir,
years,
features,
labels,
dgl_graph=graph_or_edge_index,
t_zero=args.t_zero,
cumulate=args.history)
else:
raise ValueError("Unknown backend")
print(dataset)
def main() -> None:
parser = argparse.ArgumentParser(description="Flower")
parser.add_argument(
"--server_address",
type=str,
default=DEFAULT_SERVER_ADDRESS,
help=f"gRPC server address (default: {DEFAULT_SERVER_ADDRESS})",
)
parser.add_argument("--cid",
type=str,
required=True,
help="Client CID (no default)")
parser.add_argument(
"--model",
type=str,
default="simple-cnn",
help="Model to use for training (default: simple-cnn)",
)
parser.add_argument(
"--dataset",
type=str,
default="cifar-10",
help="Dataset to use fro training (default: cifar-10)",
)
parser.add_argument(
"--device",
type=str,
default="CPU",
help="Device to run the model on (default: CPU)",
)
parser.add_argument(
"--log_host",
type=str,
help="Logserver address (no default)",
)
args = parser.parse_args()
# Configure logger
fl.common.logger.configure(f"client_{args.cid}", host=args.log_host)
# Load model and data
model = models.load_model(model_name=args.model, framework="TF")
xy_train, xy_test = datasets.load_data(dataset_name=args.dataset,
framework="TF")
# Start client
keras_client = TfKerasClient(args.cid, model, xy_train, xy_test)
client = fl.client.keras_client.KerasClientWrapper(keras_client)
try:
fl.client.start_client(args.server_address, client)
except:
print("Either something went wrong or server finished execution!!")
def run_exp(dbs, da_s1, da_s2, expdir, ae_weights_dir, trials=5, verbose=0,
pretrain_epochs=50, finetune_epochs=50, use_multiprocessing=True):
# Log files
if not os.path.exists(expdir):
os.makedirs(expdir)
logfile = open(expdir + '/results.csv', 'a')
logwriter = csv.DictWriter(logfile, fieldnames=['trials', 'acc', 'nmi', 'time'])
logwriter.writeheader()
# Begin training on different datasets
for db in dbs:
logwriter.writerow(dict(trials=db, acc='', nmi='', time=''))
# load dataset
x, y = load_data(db)
# setting parameters
n_clusters = len(np.unique(y))
dims = [x.shape[-1], 500, 500, 2000, 10]
# Training
results = np.zeros(shape=[trials, 3], dtype=float) # init metrics before finetuning
for i in range(trials): # base
t0 = time()
save_dir = os.path.join(expdir, db, 'trial%d' % i)
if not os.path.exists(save_dir):
os.makedirs(save_dir)
# prepare model
model = ASPC(dims, n_clusters)
model.compile(optimizer=Adam(0.0001), loss='mse')
# pretraining
ae_weights = 'ae_weights.h5'
if ae_weights_dir is None:
model.pretrain(x, y, optimizer=SGD(1.0, 0.9), epochs=pretrain_epochs,
save_dir=save_dir, da_s1=da_s1, verbose=verbose, use_multiprocessing=use_multiprocessing)
ae_weights = os.path.join(save_dir, ae_weights)
else:
ae_weights = os.path.join(ae_weights_dir, db, 'trial%d' % i, ae_weights)
# finetuning
results[i, :2] = model.fit(x, y, epochs=finetune_epochs if db!='fmnist' else 10,
da_s2=da_s2, save_dir=save_dir, ae_weights=ae_weights,
use_multiprocessing=use_multiprocessing)
results[i, 2] = time() - t0
for t, line in enumerate(results):
logwriter.writerow(dict(trials=t, acc=line[0], nmi=line[1], time=line[2]))
mean = np.mean(results, 0)
logwriter.writerow(dict(trials='avg', acc=mean[0], nmi=mean[1], time=mean[2]))
logfile.flush()
logfile.close()
def main(_):
# Parsing arguments
parser = argparse.ArgumentParser(description='Training GANs or VAEs')
parser.add_argument('--model', type=str, required=True)
parser.add_argument('--dataset', type=str, required=True)
parser.add_argument('--datasize', type=int, default=-1)
parser.add_argument('--epoch', type=int, default=200)
parser.add_argument('--batchsize', type=int, default=50)
parser.add_argument('--output', default='output')
parser.add_argument('--zdims', type=int, default=256)
parser.add_argument('--gpu', type=int, default=0)
parser.add_argument('--resume', type=str, default=None)
parser.add_argument('--testmode', action='store_true')
args = parser.parse_args()
# select gpu
os.environ['CUDA_VISIBLE_DEVICES'] = str(args.gpu)
# Make output direcotiry if not exists
if not os.path.isdir(args.output):
os.mkdir(args.output)
# Load datasets
if args.dataset == 'mnist':
datasets = mnist.load_data()
elif args.dataset == 'svhn':
datasets = svhn.load_data()
else:
datasets = load_data(args.dataset, args.datasize)
# Construct model
if args.model not in models:
raise Exception('Unknown model:', args.model)
model = models[args.model](
batchsize=args.batchsize,
input_shape=datasets.shape[1:],
attr_names=None or datasets.attr_names,
z_dims=args.zdims,
output=args.output,
resume=args.resume
)
if args.testmode:
model.test_mode = True
tf.set_random_seed(12345)
# Training loop
datasets.images = datasets.images.astype('float32') * 2.0 - 1.0
model.main_loop(datasets,
epochs=args.epoch)
def main():
print('Prepare dataset')
# Dataset
data_train, data_valid, data_test = datasets.load_data(
args.dataset, args.data_path, args.representation, args.normalization)
# Data Loader
train_loader = torch.utils.data.DataLoader(
data_train,
collate_fn=datasets.collate_fn_multiple_size,
batch_size=args.batch_size,
num_workers=args.prefetch,
pin_memory=True)
valid_loader = torch.utils.data.DataLoader(
data_valid,
batch_size=1,
collate_fn=datasets.collate_fn_multiple_size,
num_workers=args.prefetch,
pin_memory=True)
test_loader = torch.utils.data.DataLoader(
data_test,
batch_size=1,
collate_fn=datasets.collate_fn_multiple_size,
num_workers=args.prefetch,
pin_memory=True)
print('Create model')
if args.distance == 'SoftHd':
net = GraphEditDistance.SoftHd()
else:
net = GraphEditDistance.Hd()
print('Loss & optimizer')
evaluation = knn
print('Check CUDA')
if args.cuda and args.ngpu > 1:
print('\t* Data Parallel **NOT TESTED**')
net = torch.nn.DataParallel(net, device_ids=list(range(args.ngpu)))
if args.cuda:
print('\t* CUDA')
net = net.cuda()
print('Validation')
acc_valid = test(valid_loader, train_loader, net, args.ngpu > 0,
evaluation)
# Evaluate best model in Test
print('Test:')
acc_test = test(test_loader, train_loader, net, args.ngpu > 0, evaluation)
def main(_):
"""config.py の CLASSES、引数で処理を実施."""
datasets = load_data(one_hot=True)
if _.train:
train(datasets, epochs=120)
else:
print(datasets.test.labels[:10])
print(predict(datasets.test.images[:10]))
print(predict([datasets.test.images[0]]))
print(predict([datasets.test.images[0]], dtype='int'))
print(predict([datasets.test.images[0]], dtype='argmax'))
def evaluate_experiment(params, output_path):
if params["data_name"] == "1d":
data, X_range = datasets.load_data(params["data_name"])
else:
data, X_range, X, Y, Z, is_us = datasets.load_data(
params["data_name"], True)
gp, gp_list, s_list, loss_list = pickle_data(
get_folder_name(params, output_path))
if params["data_name"] == "ozone":
goal_func = partial(utils.usa_goal_func,
data=data,
X_range=X_range,
sign=-1)
elif params["data_name"] == "1d":
goal_func = partial(utils.d1_goal_func, sign=-1)
else:
raise NameError("Data set name unknown.")
y_opt = data[np.argmax(data)]
x_opt = X_range[np.argmax(data), :]
for i in range(len(gp_list)):
gpi = gp_list[i]
gpi.update_hyperparams_ml(params["bounds"])
mean = np.zeros((X_range.shape[0], 1))
var = np.zeros((X_range.shape[0], 1))
for j in range(X_range.shape[0]):
mean[j], var[j] = gpi.predict(X_range[j, :][np.newaxis, :])
x_best = X_range[np.argmax(mean), :]
x_first = gpi.Xtrain[0, :]
print "Iteration:", (i + 1) * params["save_every"]
print "Gap measure:", utils.gap_measure(goal_func, x_first, x_best,
y_opt)
print "Closeness measure:", utils.closeness_measure(x_best, x_opt)
def main():
# Parse hyperparams
hparams = rebar.default_hparams
hparams.parse(FLAGS.hparams)
print(hparams.values())
train_xs, valid_xs, test_xs = datasets.load_data(hparams)
mean_xs = np.mean(train_xs, axis=0) # Compute mean centering on training
training_steps = 2000000
model = getattr(rebar, hparams.model)
sbn = model(hparams, mean_xs=mean_xs)
scores = train(sbn, train_xs, valid_xs, test_xs,
training_steps=training_steps, debug=False)
def do_fid_inception(args):
args_json = json.load(open(os.path.join(args.dir, 'hps.txt')))
vars(args).update(args_json)
dset = load_data(args.dataset, True)
ckpt_dir = tf.train.get_checkpoint_state(args.dir).model_checkpoint_path
val_images = dset.test.images
G = Graph(args, val_images.shape[1], val_images.shape[3])
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
sess = tf.Session(config=config)
sess.run(tf.global_variables_initializer())
saver = tfv1.train.Saver(keep_checkpoint_every_n_hours=2, max_to_keep=1)
print('RESTORING WEIGHTS FROM', ckpt_dir)
saver.restore(sess, ckpt_dir)
score = compute_fid_inception(dset, sess, G)
print('FID SCORE = {}'.format(score))
def main(args):
model = models.get_model(args.model, args.dataset)
dataset = model.dataset
train, test = datasets.load_data(dataset)
train, validation = datasets.split_data(train, fractions=[0.8, 0.2])
trainer.train(model, train, validation, args.epochs, args.batch_size,
args.learning_rate,
dataset_update=args.dataset_update, increments=args.splits,
use_ewc=args.ewc, ewc_lambda=args.ewc_lambda,
ewc_samples=args.ewc_samples,
use_fim=args.fim, fim_threshold=args.fim_threshold,
fim_samples=args.fim_samples,
use_incdet=args.incdet,
incdet_threshold=args.incdet_threshold)
def process_dataset_iris(pkl_dataset_dir):
dataset_load = load_data(pkl_dataset_dir)
data = dataset_load[0]
data = trans_complex_to_real(data)
target = dataset_load[1]
return Bunch(data=data, target=target)
__author__ = 'lizuyao'
import pickle
import datasets
import sys
import matplotlib
matplotlib.use('Agg')
import matplotlib.pyplot as plt
from unbalanced_dataset import UnderSampler
from mpl_toolkits.mplot3d import Axes3D
X_reduced=pickle.load(open(sys.argv[1], "rb"))
fileName = sys.argv[2]
X, Y = datasets.load_data(fileName)
# Generate the new dataset using under-sampling method
verbose = False
# 'Random under-sampling'
# ratio of majority elements to sample with respect to the number of minority cases.
US = UnderSampler(ratio=1.,verbose=verbose)
X_reduced, Y = US.fit_transform(X_reduced, Y)
# To getter a better understanding of interaction of the dimensions
# plot the first three tsne dimensions
fig = plt.figure(1, figsize=(8, 6))
ax = Axes3D(fig, elev=-150, azim=110)
ax.scatter(X_reduced[:, 0], X_reduced[:, 1], X_reduced[:, 2], c=Y, cmap=plt.cm.Paired)
ax.set_title("First three tsne directions")
ax.set_xlabel("1st eigenvector")
ax.w_xaxis.set_ticklabels([])
ax.set_ylabel("2nd eigenvector")
ax.w_yaxis.set_ticklabels([])
datasets = []
for dataset_id in 'ABC':
datasets.append([url_pattern % (dataset_id, train_or_test)
for train_or_test in ['train', 'test']])
models = ['LDA', 'LogitReg', 'LinearReg', 'QDA']
n_models = len(models)
n_datasets = len(datasets)
scores = np.ndarray((n_datasets * 2, n_models))
for dataset, j in zip(datasets, xrange(n_datasets)):
train, test = dataset
# load the data
dataset_name = os.path.basename(train.replace(".train", ""))
print ">" * 80, "Begin (%s)" % dataset_name
train = load_data(train, data_dir=os.getcwd())
test = load_data(test, data_dir=os.getcwd())
X_train = train[..., :-1]
Y_train = train[..., -1]
X_test = test[..., :-1]
Y_test = test[..., -1]
# fit models
for model_id, i in zip(models, xrange(n_models)):
print "\nRunning %s ..." % model_id
model = eval(model_id)(verbose=0).fit(X_train, Y_train)
print "... done."
model.print_params()
mistrain = (model.predict(X_train) != Y_train
).sum() / (1. * len(Y_train))
self.means_ = self.labels_.dot(X) # uncorrected
counts = self.labels_.sum(axis=1) # number of points in each class
for k in xrange(self.n_classes):
# update proportions
self.priors_[k] = counts[k] / N
# update means
self.means_[k] /= counts[k]
# update covariance matrices
X_k = X - self.means_[k]
D = np.diag(self.labels_[k])
self.covariance_matrices_[k] = X_k.T.dot(D.dot(
X_k)) / counts[k]
return self
if __name__ == '__main__':
import matplotlib.pyplot as plt
X = load_data(
"http://www.di.ens.fr/~fbach/courses/fall2013/EMGaussian.data")
# fit
em = EM(4).fit(X, max_iter=50)
# plot results
plt.scatter(*X.T, c=em.labels_.argmax(axis=0))
plt.show()
