def run_one_scenario(cost_models,
list_algos,
filename: str,
batch_size: int = 1,
stochastic: bool = True,
nb_epoch: int = 250,
step_size=None,
compression: CompressionModel = None,
use_averaging: bool = False,
fraction_sampled_workers: int = 1) -> None:
all_descent = {}
stochasticity = 'sto' if stochastic else "full"
if stochastic:
experiments_settings = "{0}-b{1}".format(stochasticity, batch_size)
else:
experiments_settings = stochasticity
for type_params in tqdm(list_algos):
multiple_sg_descent = multiple_run_descent(
type_params,
cost_models=cost_models,
compression_model=compression,
use_averaging=use_averaging,
stochastic=stochastic,
nb_epoch=nb_epoch,
step_formula=step_size,
batch_size=batch_size,
logs_file=filename,
fraction_sampled_workers=fraction_sampled_workers)
all_descent[type_params.name()] = multiple_sg_descent
res = ResultsOfSeveralDescents(all_descent, len(cost_models))
pickle_saver(res, "{0}/descent-{1}".format(filename,
experiments_settings))
def prepare_noniid_dataset(data, pivot_label: str, data_path: str, pickle_path: str, nb_cluster: int, double_check: bool =False):
# The TSNE representation is independent of the number of devices.
tsne_file = "{0}-tsne".format(data_path)
if not file_exist("{0}.pkl".format(tsne_file)):
# Running TNSE to obtain a 2D representation of data
logging.debug("The TSNE representation ({0}) doesn't exist."
.format(tsne_file))
embedded_data = tsne(data)
pickle_saver(embedded_data, tsne_file)
tsne_cluster_file = "{0}/tsne-cluster".format(pickle_path)
if not file_exist("{0}.pkl".format(tsne_cluster_file)):
# Finding clusters in the TNSE.
logging.debug("Finding non-iid clusters in the TNSE represesentation: {0}.pkl".format(tsne_file))
embedded_data = pickle_loader("{0}".format(tsne_file))
logging.debug("Saving found clusters : {0}.pkl".format(tsne_cluster_file))
predicted_cluster = find_cluster(embedded_data, nb_cluster)
pickle_saver(predicted_cluster, "{0}".format(tsne_cluster_file))
predicted_cluster = pickle_loader("{0}".format(tsne_cluster_file))
# With the found clusters, splitting data.
X, Y = clustering_data(data, predicted_cluster, pivot_label, nb_cluster)
if double_check:
logging.debug("Checking data cluserization, wait until completion before seeing the plots.")
# Checking that splitting data by cluster is valid.
check_data_clusterisation(X, Y, nb_cluster)
return X, Y
def run_one_scenario(cost_models,
list_algos,
logs_file: str,
experiments_settings: str,
batch_size: int = 1,
stochastic: bool = True,
nb_epoch: int = 250,
step_size=None,
compression: CompressionModel = None,
use_averaging: bool = False,
fraction_sampled_workers: int = 1,
modify_run=None) -> None:
pickle_file = "{0}/descent-{1}".format(logs_file, experiments_settings)
if modify_run is None:
if file_exist(pickle_file + ".pkl"):
remove_file(pickle_file + ".pkl")
algos = list_algos
else:
algos = [list_algos[i] for i in modify_run]
for type_params in tqdm(algos):
multiple_sg_descent = multiple_run_descent(
type_params,
cost_models=cost_models,
compression_model=compression,
use_averaging=use_averaging,
stochastic=stochastic,
nb_epoch=nb_epoch,
step_formula=step_size,
batch_size=batch_size,
logs_file=logs_file,
fraction_sampled_workers=fraction_sampled_workers)
if logs_file:
logs = open("{0}/logs.txt".format(logs_file), "a+")
logs.write(
"{0} size of the multiple SG descent: {1:.2e} bits\n".format(
type_params.name(), asizeof.asizeof(multiple_sg_descent)))
logs.close()
if file_exist(pickle_file + ".pkl"):
res = pickle_loader(pickle_file)
res.add_descent(multiple_sg_descent,
type_params.name(),
deep_learning_run=False)
else:
res = ResultsOfSeveralDescents(len(cost_models))
res.add_descent(multiple_sg_descent,
type_params.name(),
deep_learning_run=False)
pickle_saver(res, pickle_file)
del res
del multiple_sg_descent
images = images.to(device)
labels = labels.to(device)
# Forward pass
outputs = model(images)
loss = criterion(outputs, labels)
# Backward and optimize
optimizer.zero_grad()
loss.backward()
optimizer.step()
test_loss_val, test_acc_val = accuracy_and_loss(model, test_loader, criterion, device)
run.update_run(loss.item(), test_loss_val, test_acc_val)
pickle_saver(run, "run_cifar10")
if (i+1) % 100 == 0:
with open("log_cifar10.txt", 'a') as f:
print ("Epoch [{}/{}], Step [{}/{}] Loss: {:.4f}"
.format(epoch+1, num_epochs, i+1, total_step, loss.item()), file=f)
# Decay learning rate
if (epoch+1) % 20 == 0:
curr_lr /= 3
update_lr(optimizer, curr_lr)
# Test the model
model.eval()
with torch.no_grad():
correct = 0
def run_experiments_in_deeplearning(dataset: str):
with open("log.txt", 'a') as f:
print("==== NEW RUN ====", file=f)
fraction_sampled_workers = 1
batch_size = 128
nb_devices = 20
algos = "mcm-vs-existing"
iid = "non-iid"
data_path, pickle_path, algos_pickle_path, picture_path = create_path_and_folders(
nb_devices, dataset, iid, algos, fraction_sampled_workers)
default_up_compression = SQuantization(quantization_levels[dataset],
norm=norm_quantization[dataset])
default_down_compression = SQuantization(quantization_levels[dataset],
norm=norm_quantization[dataset])
exp_name = "{0}_m{1}_lr{2}_sup{3}_sdwn{4}_b{5}_wd{6}".format(
models[dataset].__name__, momentums[dataset],
optimal_steps_size[dataset], default_up_compression.level,
default_down_compression.level, batch_size, weight_decay[dataset])
all_descent = {}
for type_params in [VanillaSGD(), Diana(), Artemis(), MCM()]:
print(type_params)
torch.cuda.empty_cache()
params = type_params.define(
cost_models=None,
n_dimensions=None,
nb_epoch=200,
nb_devices=nb_devices,
batch_size=batch_size,
fraction_sampled_workers=1,
up_compression_model=default_up_compression,
down_compression_model=default_down_compression)
params = cast_to_DL(params, dataset, models[dataset],
optimal_steps_size[dataset], weight_decay[dataset])
params.log_file = "log.txt"
params.momentum = momentums[dataset]
params.print()
with open(params.log_file, 'a') as f:
print(type_params, file=f)
print("Optimal step size: ", params.optimal_step_size, file=f)
multiple_sg_descent = run_tuned_exp(params)
all_descent[type_params.name()] = multiple_sg_descent
res = ResultsOfSeveralDescents(all_descent, nb_devices)
# res.add_descent(multiple_sg_descent, type_params.name())
pickle_saver(res, "{0}/{1}".format(algos_pickle_path, exp_name))
res = pickle_loader("{0}/{1}".format(algos_pickle_path, exp_name))
# Plotting without averaging
plot_error_dist(res.get_loss(np.array(0), in_log=True),
res.names,
res.nb_devices_for_the_run,
batch_size=batch_size,
all_error=res.get_std(np.array(0), in_log=True),
x_legend="Number of passes on data",
ylegends="train_loss",
picture_name="{0}/{1}_train_losses".format(
picture_path, exp_name))
plot_error_dist(res.get_loss(np.array(0), in_log=True),
res.names,
res.nb_devices_for_the_run,
batch_size=batch_size,
x_points=res.X_number_of_bits,
ylegends="train_loss",
all_error=res.get_std(np.array(0), in_log=True),
x_legend="Communicated bits",
picture_name="{0}/{1}_train_losses_bits".format(
picture_path, exp_name))
plot_error_dist(res.get_test_accuracies(),
res.names,
res.nb_devices_for_the_run,
ylegends="accuracy",
all_error=res.get_test_accuracies_std(),
x_legend="Number of passes on data",
batch_size=batch_size,
picture_name="{0}/{1}_test_accuracies".format(
picture_path, exp_name))
plot_error_dist(res.get_test_losses(in_log=True),
res.names,
res.nb_devices_for_the_run,
batch_size=batch_size,
all_error=res.get_test_losses_std(in_log=True),
x_legend="Number of passes on data",
ylegends="test_loss",
picture_name="{0}/{1}_test_losses".format(
picture_path, exp_name))
def run_for_different_scenarios(cost_models,
list_algos,
values,
labels,
experiments_settings: str,
logs_file: str,
batch_size: int = 1,
stochastic: bool = True,
nb_epoch: int = 250,
step_formula=None,
compression: CompressionModel = None,
scenario: str = "step") -> None:
assert scenario in ["step", "compression", "alpha"], "There is three possible scenarios : to analyze by step size," \
" by compression operators, or by value of alpha."
nb_devices_for_the_run = len(cost_models)
all_kind_of_compression_res = []
all_descent_various_gamma = {}
descent_by_algo_and_step_size = {}
# Corresponds to descent with optimal gamma for each algorithm
optimal_descents = {}
for param_algo in tqdm(list_algos):
losses_by_algo, losses_avg_by_algo, norm_ef_by_algo, dist_model_by_algo = [], [], [], []
h_i_to_optimal_grad_by_algo, var_models_by_algo = [], []
descent_by_step_size = {}
for (value, label) in zip(values, labels):
if scenario == "step":
multiple_sg_descent = multiple_run_descent(
param_algo,
cost_models=cost_models,
use_averaging=True,
stochastic=stochastic,
batch_size=batch_size,
step_formula=value,
nb_epoch=nb_epoch,
compression_model=compression,
logs_file=logs_file)
if scenario in ["compression", "alpha"]:
multiple_sg_descent = multiple_run_descent(
param_algo,
cost_models=cost_models,
use_averaging=True,
stochastic=stochastic,
batch_size=batch_size,
step_formula=step_formula,
compression_model=value,
nb_epoch=nb_epoch,
logs_file=logs_file)
descent_by_step_size[label] = multiple_sg_descent
losses_by_label, losses_avg_by_label, norm_ef_by_label, dist_model_by_label = [], [], [], []
h_i_to_optimal_grad_by_label, var_models_by_label = [], []
# Picking the minimum values for each of the run.
for seq_losses, seq_losses_avg, seq_norm_ef, seq_dist_model, seq_h_i_optimal, seq_var_models in \
zip(multiple_sg_descent.train_losses, multiple_sg_descent.averaged_train_losses,
multiple_sg_descent.norm_error_feedback, multiple_sg_descent.dist_to_model,
multiple_sg_descent.h_i_to_optimal_grad, multiple_sg_descent.var_models):
losses_by_label.append(min(seq_losses))
losses_avg_by_label.append(min(seq_losses_avg))
norm_ef_by_label.append(seq_norm_ef[-1])
dist_model_by_label.append(seq_dist_model[-1])
var_models_by_label.append(seq_var_models[-1])
h_i_to_optimal_grad_by_label.append(seq_h_i_optimal[-1])
losses_by_algo.append(losses_by_label)
losses_avg_by_algo.append(losses_avg_by_label)
norm_ef_by_algo.append(norm_ef_by_label)
dist_model_by_algo.append(dist_model_by_label)
var_models_by_algo.append(var_models_by_label)
h_i_to_optimal_grad_by_algo.append(h_i_to_optimal_grad_by_label)
res_by_algo_and_step_size = ResultsOfSeveralDescents(
nb_devices_for_the_run)
res_by_algo_and_step_size.add_dict_of_descent(descent_by_step_size)
descent_by_algo_and_step_size[
param_algo.name()] = res_by_algo_and_step_size
# Find optimal descent for the algo:
min_loss_desc = 10e12
opt_desc = None
for desc in descent_by_step_size.values():
if min_loss_desc > min([
desc.train_losses[j][-1]
for j in range(len(desc.train_losses))
]):
min_loss_desc = min([
desc.train_losses[j][-1]
for j in range(len(desc.train_losses))
])
opt_desc = desc
# Adding the optimal descent to the dict of optimal descent
optimal_descents[param_algo.name()] = opt_desc
artificial_multiple_descent = AverageOfSeveralIdenticalRun()
artificial_multiple_descent.append_list(losses_by_algo,
losses_avg_by_algo,
norm_ef_by_algo,
dist_model_by_algo,
h_i_to_optimal_grad_by_algo,
var_models_by_algo)
all_descent_various_gamma[
param_algo.name()] = artificial_multiple_descent
all_kind_of_compression_res.append(all_descent_various_gamma)
res_various_gamma = ResultsOfSeveralDescents(nb_devices_for_the_run)
res_various_gamma.add_dict_of_descent(all_descent_various_gamma,
deep_learning_run=False)
pickle_saver(
res_various_gamma, "{0}/{1}-{2}".format(logs_file, scenario,
experiments_settings))
res_opt_gamma = ResultsOfSeveralDescents(nb_devices_for_the_run)
res_opt_gamma.add_dict_of_descent(optimal_descents,
deep_learning_run=False)
pickle_saver(
res_opt_gamma, "{0}/{1}-optimal-{2}".format(logs_file, scenario,
experiments_settings))
pickle_saver(
descent_by_algo_and_step_size,
"{0}/{1}-descent_by_algo-{2}".format(logs_file, scenario,
experiments_settings))
def run_experiments_in_deeplearning(dataset: str,
plot_only: bool = False) -> None:
"""Runs and plots experiments for a given dataset using an appropriate neural network.
:param dataset: Name of the dataset
:param plot_only: True if the goal is not to rerun all experiments but only to regenerate figures.
"""
fraction_sampled_workers = 1
batch_size = batch_sizes[dataset]
nb_devices = 20
algos = sys.argv[2]
iid = sys.argv[3]
stochastic = True
create_folder_if_not_existing(algos)
log_file = algos + "/log_" + dataset + "_" + iid + ".txt"
with open(log_file, 'a') as f:
print("==== NEW RUN ====", file=f)
with open(log_file, 'a') as f:
print(
"stochastic -> {0}, iid -> {1}, batch_size -> {2}, norm -> {3}, s -> {4}, momentum -> {5}, model -> {6}"
.format(stochastic, iid, batch_size, norm_quantization[dataset],
quantization_levels[dataset], momentums[dataset],
models[dataset].__name__),
file=f)
data_path, pickle_path, algos_pickle_path, picture_path = create_path_and_folders(
nb_devices, dataset, iid, algos, fraction_sampled_workers)
default_up_compression = SQuantization(quantization_levels[dataset],
norm=norm_quantization[dataset])
default_down_compression = SQuantization(quantization_levels[dataset],
norm=norm_quantization[dataset])
loaders = create_loaders(dataset, iid, nb_devices, batch_size, stochastic)
_, train_loader_workers_full, _ = loaders
dim = next(iter(train_loader_workers_full[0]))[0].shape[1]
if optimal_steps_size[dataset] is None:
L = compute_L(train_loader_workers_full)
optimal_steps_size[dataset] = 1 / L
print("Step size:", optimal_steps_size[dataset])
exp_name = name_of_the_experiments(dataset, stochastic)
pickle_file = "{0}/{1}".format(algos_pickle_path, exp_name)
list_algos = choose_algo(algos, stochastic, fraction_sampled_workers)
if not plot_only:
if file_exist(pickle_file + ".pkl"):
remove_file(pickle_file + ".pkl")
for type_params in list_algos:
print(type_params)
torch.cuda.empty_cache()
params = type_params.define(
cost_models=None,
n_dimensions=dim,
nb_epoch=300,
nb_devices=nb_devices,
stochastic=stochastic,
batch_size=batch_size,
fraction_sampled_workers=fraction_sampled_workers,
up_compression_model=default_up_compression,
down_compression_model=default_down_compression)
params = cast_to_DL(params, dataset, models[dataset],
optimal_steps_size[dataset],
weight_decay[dataset], iid)
params.log_file = log_file
params.momentum = momentums[dataset]
params.criterion = criterion[dataset]
params.print()
with open(params.log_file, 'a') as f:
print(type_params, file=f)
print("Optimal step size: ", params.optimal_step_size, file=f)
multiple_descent = AverageOfSeveralIdenticalRun()
seed_everything(seed=42)
start = time.time()
for i in range(NB_RUN):
print('Run {:3d}/{:3d}:'.format(i + 1, NB_RUN))
fixed_params = copy.deepcopy(params)
try:
training = Train(loaders, fixed_params)
multiple_descent.append_from_DL(training.run_training())
except ValueError as err:
print(err)
continue
with open(log_file, 'a') as f:
print("Time of the run: {:.2f}s".format(time.time() - start),
file=f)
with open(params.log_file, 'a') as f:
print("{0} size of the multiple SG descent: {1:.2e} bits\n".
format(type_params.name(),
asizeof.asizeof(multiple_descent)),
file=f)
if file_exist(pickle_file + ".pkl"):
res = pickle_loader(pickle_file)
res.add_descent(multiple_descent,
type_params.name(),
deep_learning_run=True)
else:
res = ResultsOfSeveralDescents(nb_devices)
res.add_descent(multiple_descent,
type_params.name(),
deep_learning_run=True)
pickle_saver(res, pickle_file)
# obj_min_cvx = pickle_loader("{0}/obj_min".format(pickle_path))
obj_min = 0 #pickle_loader("{0}/obj_min".format(pickle_path))
res = pickle_loader(pickle_file)
# obj_min = min(res.get_loss(np.array(0), in_log=False)[0])
# print("Obj min in convex:", obj_min_cvx)
print("Obj min in dl:", obj_min)
# Plotting
plot_error_dist(res.get_loss(np.array(obj_min)),
res.names,
all_error=res.get_std(np.array(obj_min)),
x_legend="Number of passes on data",
ylegends="train_loss",
picture_name="{0}/{1}_train_losses".format(
picture_path, exp_name))
plot_error_dist(res.get_loss(np.array(obj_min)),
res.names,
x_points=res.X_number_of_bits,
ylegends="train_loss",
all_error=res.get_std(np.array(obj_min)),
x_legend="Communicated bits",
picture_name="{0}/{1}_train_losses_bits".format(
picture_path, exp_name))
plot_error_dist(res.get_test_accuracies(),
res.names,
ylegends="accuracy",
all_error=res.get_test_accuracies_std(),
x_legend="Number of passes on data",
picture_name="{0}/{1}_test_accuracies".format(
picture_path, exp_name))
plot_error_dist(res.get_test_losses(in_log=True),
res.names,
all_error=res.get_test_losses_std(in_log=True),
x_legend="Number of passes on data",
ylegends="test_loss",
picture_name="{0}/{1}_test_losses".format(
picture_path, exp_name))