def batch_size_linear_search():
min = 8
max = 600
step_size = 8
optimizer = lambda x: torch.optim.SGD(x, lr=0.1)
experiment_name = "batch_size_linear_search"
t = Timer()
batch_size_times = {}
for i, batch_size in enumerate(range(min, max, step_size)):
t.start()
main(experiment_name, optimizer, epochs=i + 2, batch_size=batch_size)
elapsed_time = t.stop()
batch_size_times[batch_size] = elapsed_time
pickle.dump(batch_size_times, open("batch_size_times.pickle", "wb"))
# Plot
batch_sizes = []
times = []
for k in sorted(batch_size_times):
batch_sizes.append(k)
times.append(batch_size_times[k])
plt.plot(np.array(batch_sizes), np.array(times))
plt.xlabel("Batch Size")
plt.ylabel("Epoch Time")
plt.title("Batch Size vs Epoch Time")
plt.show()
def main(experiment_name,
optimizer,
output_directory_root="experiments/resnet18_logistic_cifar10",
epochs=60,
batch_size=512,
num_workers=1):
output_directory = os.path.join(output_directory_root, experiment_name)
if not os.path.isdir(output_directory):
os.makedirs(output_directory, exist_ok=True)
# Setup regular log file + tensorboard
logfile_path = os.path.join(output_directory, "logfile.txt")
setup_logger_tqdm(logfile_path)
tensorboard_log_directory = os.path.join("runs",
"resnet18_logistic_cifar10",
experiment_name)
tensorboard_summary_writer = SummaryWriter(
log_dir=tensorboard_log_directory)
# Choose Training Device
use_cuda = torch.cuda.is_available()
logger.info(f"CUDA Available? {use_cuda}")
device = "cuda" if use_cuda else "cpu"
# Datasets and Loaders
train_set_loader, test_set_loader = get_data_loaders(
batch_size, num_workers)
# Create Model & Optimizer
model = torchvision.models.resnet18(pretrained=True)
for param in model.parameters():
param.requires_grad = False
num_classes = 10
model.fc = nn.Linear(model.fc.in_features, 10)
model.to(device)
optimizer = optimizer(model.parameters())
logger.info("=========== Commencing Training ===========")
logger.info(f"Epoch Count: {epochs}")
logger.info(f"Batch Size: {batch_size}")
# Load Checkpoint
checkpoint_file_path = os.path.join(output_directory, "checkpoint.pth")
start_epoch = 0
if os.path.exists(checkpoint_file_path):
logger.info("Checkpoint Found - Loading!")
checkpoint = torch.load(checkpoint_file_path)
logger.info(f"Last completed epoch: {checkpoint['epoch']}")
logger.info(f"Average Train Loss: {checkpoint['train_loss']}")
logger.info(f"Top-1 Train Accuracy: {checkpoint['train_accuracy']}")
logger.info(f"Top-1 Test Accuracy: {checkpoint['test_accuracy']}")
start_epoch = checkpoint["epoch"] + 1
logger.info(f"Resuming at epoch {start_epoch}")
model.load_state_dict(checkpoint["model_state_dict"])
optimizer.load_state_dict(checkpoint["optimizer_state_dict"])
else:
logger.info("No checkpoint found, starting from scratch.")
# Training Loop
t = Timer()
for epoch in range(start_epoch, epochs):
t.start()
logger.info("-" * 10)
logger.info(f"Epoch {epoch}")
logger.info("-" * 10)
train_loss, train_accuracy = train_model(device, model,
train_set_loader, optimizer)
tensorboard_summary_writer.add_scalar("train_loss", train_loss, epoch)
tensorboard_summary_writer.add_scalar("train_accuracy", train_accuracy,
epoch)
test_accuracy = test_model(device, model, test_set_loader, optimizer)
tensorboard_summary_writer.add_scalar("test_accuracy", test_accuracy,
epoch)
# Save Checkpoint
logger.info("Saving checkpoint.")
torch.save(
{
'epoch': epoch,
'model_state_dict': model.state_dict(),
'optimizer_state_dict': optimizer.state_dict(),
'train_loss': train_loss,
'train_accuracy': train_accuracy,
'test_accuracy': test_accuracy
}, checkpoint_file_path)
elapsed_time = t.stop()
logger.info(f"End of epoch {epoch}, took {elapsed_time:0.4f} seconds.")
logger.info(f"Average Train Loss: {train_loss}")
logger.info(f"Top-1 Train Accuracy: {train_accuracy}")
logger.info(f"Top-1 Test Accuracy: {test_accuracy}")
logger.info("")
def main(device, mp_args, dataloader_func, model, optimizer_callback,
output_directory, tensorboard_log_directory, epochs):
global_rank = mp_args.nr * mp_args.gpus + device
dist.init_process_group(backend='nccl',
init_method='env://',
world_size=mp_args.world_size,
rank=global_rank)
output_directory = os.path.join(output_directory, f"rank_{global_rank}")
if not os.path.isdir(output_directory):
os.makedirs(output_directory, exist_ok=True)
# Setup regular log file
logfile_path = os.path.join(output_directory, "logfile.txt")
setup_logger_tqdm(logfile_path)
# Setup TensorBoard logging
tensorboard_log_directory = os.path.join(tensorboard_log_directory,
f"rank_{global_rank}")
tensorboard_summary_writer = SummaryWriter(
log_dir=tensorboard_log_directory)
# Dataloaders
train_set_loader, test_set_loader = dataloader_func(
mp_args.world_size, global_rank)
# Model & Optimizer
model.to(device)
optimizer = optimizer_callback(model)
model = nn.parallel.DistributedDataParallel(model, device_ids=[device])
logger.info(f"Epoch Count: {epochs}")
# Load Checkpoint
checkpoint_file_path = os.path.join(output_directory, "checkpoint.pth")
start_epoch = 0
if os.path.exists(checkpoint_file_path):
logger.info("Checkpoint Found - Loading!")
checkpoint = torch.load(checkpoint_file_path)
logger.info(f"Last completed epoch: {checkpoint['epoch']}")
logger.info(f"Average Train Loss: {checkpoint['train_loss']}")
logger.info(f"Top-1 Train Accuracy: {checkpoint['train_accuracy']}")
logger.info(f"Top-1 Test Accuracy: {checkpoint['test_accuracy']}")
start_epoch = checkpoint["epoch"] + 1
logger.info(f"Resuming at epoch {start_epoch}")
model.load_state_dict(checkpoint["model_state_dict"])
optimizer.load_state_dict(checkpoint["optimizer_state_dict"])
else:
logger.info("No checkpoint found, starting from scratch.")
# Training Loop
t = Timer()
#progress = tqdm(total=epochs, initial=start_epoch, desc="Epochs")
for epoch in range(start_epoch, epochs):
t.start()
logger.info(f"Commence EPOCH {epoch}")
# Train
train_loss, train_accuracy = train_model(device, model,
train_set_loader, optimizer)
tensorboard_summary_writer.add_scalar("train_loss", train_loss, epoch)
tensorboard_summary_writer.add_scalar("train_accuracy", train_accuracy,
epoch)
# Test
test_accuracy = test_model(device, model, test_set_loader)
tensorboard_summary_writer.add_scalar("test_accuracy", test_accuracy,
epoch)
# Save Checkpoint
logger.info("Saving checkpoint.")
torch.save(
{
'epoch': epoch,
'model_state_dict': model.state_dict(),
'optimizer_state_dict': optimizer.state_dict(),
'train_loss': train_loss,
'train_accuracy': train_accuracy,
'test_accuracy': test_accuracy
}, checkpoint_file_path)
elapsed_time = t.stop()
logger.info(f"End of epoch {epoch}, took {elapsed_time:0.4f} seconds.")
logger.info(f"Average Train Loss: {train_loss}")
logger.info(f"Top-1 Train Accuracy: {train_accuracy}")
logger.info(f"Top-1 Test Accuracy: {test_accuracy}")
def main(dataloader_func,
model,
optimizer_callback,
output_directory,
tensorboard_log_directory,
lr_scheduler=None,
epochs=150):
if not os.path.isdir(output_directory):
os.makedirs(output_directory, exist_ok=True)
# Setup regular log file
logfile_path = os.path.join(output_directory, "logfile.txt")
setup_logger_tqdm(logfile_path)
# Setup TensorBoard logging
tensorboard_summary_writer = SummaryWriter(
log_dir=tensorboard_log_directory)
# Choose Training Device
use_cuda = torch.cuda.is_available()
logger.info(f"CUDA Available? {use_cuda}")
device = "cuda" if use_cuda else "cpu"
# Dataloaders
train_set_loader, test_set_loader = dataloader_func()
# Model & Optimizer
model.to(device)
optimizer = optimizer_callback(model)
if lr_scheduler:
lr_scheduler = lr_scheduler(optimizer)
logger.info(f"Epoch Count: {epochs}")
# Load Checkpoint
checkpoint_file_path = os.path.join(output_directory, "checkpoint.pth")
start_epoch = 0
if os.path.exists(checkpoint_file_path):
logger.info("Checkpoint Found - Loading!")
checkpoint = torch.load(checkpoint_file_path)
logger.info(f"Last completed epoch: {checkpoint['epoch']}")
logger.info(f"Average Train Loss: {checkpoint['train_loss']}")
logger.info(f"Top-1 Train Accuracy: {checkpoint['train_accuracy']}")
logger.info(f"Top-1 Test Accuracy: {checkpoint['test_accuracy']}")
start_epoch = checkpoint["epoch"] + 1
logger.info(f"Resuming at epoch {start_epoch}")
model.load_state_dict(checkpoint["model_state_dict"])
optimizer.load_state_dict(checkpoint["optimizer_state_dict"])
if lr_scheduler:
lr_scheduler.load_state_dict(checkpoint["lr_scheduler_state_dict"])
else:
logger.info("No checkpoint found, starting from scratch.")
# Training Loop
t = Timer()
for epoch in range(start_epoch, epochs):
t.start()
logger.info(f"Commence EPOCH {epoch}")
# Train
train_loss, train_accuracy = train_model(device, model,
train_set_loader, optimizer)
tensorboard_summary_writer.add_scalar("train_loss", train_loss, epoch)
tensorboard_summary_writer.add_scalar("train_accuracy", train_accuracy,
epoch)
# Test
test_accuracy = test_model(device, model, test_set_loader, optimizer)
tensorboard_summary_writer.add_scalar("test_accuracy", test_accuracy,
epoch)
scheduler_dict = None
if lr_scheduler:
lr_scheduler.step()
scheduler_dict = lr_scheduler.state_dict()
# Save Checkpoint
logger.info("Saving checkpoint.")
torch.save(
{
'epoch': epoch,
'model_state_dict': model.state_dict(),
'optimizer_state_dict': optimizer.state_dict(),
'lr_scheduler_state_dict': scheduler_dict,
'train_loss': train_loss,
'train_accuracy': train_accuracy,
'test_accuracy': test_accuracy
}, checkpoint_file_path)
elapsed_time = t.stop()
logger.info(f"End of epoch {epoch}, took {elapsed_time:0.4f} seconds.")
logger.info(f"Average Train Loss: {train_loss}")
logger.info(f"Top-1 Train Accuracy: {train_accuracy}")
logger.info(f"Top-1 Test Accuracy: {test_accuracy}")
def main(carrier_path,
marking_network,
target_network,
target_checkpoint,
batch_size=256,
num_workers=1,
align=True,
test_set_loader=None):
# Setup Device
use_cuda = torch.cuda.is_available()
logger.info(f"CUDA Available? {use_cuda}")
device = torch.device("cuda" if use_cuda else "cpu")
# Load Carrier
carrier = torch.load(carrier_path).numpy()
t = Timer()
t.start()
# Align spaces (Or Not)
W = target_checkpoint["model_state_dict"]["fc.weight"].cpu().numpy()
if align:
logger.info(
"Aligning marking and target network feature space with least squares"
)
marking_network.to(device)
marking_network.eval()
target_network.to(device)
target_network.eval()
# Setup Dataloader
if not test_set_loader:
test_set_loader = get_data_loader(batch_size, num_workers)
logger.info(
"Extracting image features from marking and target networks.")
features_marking, _ = extract_features(test_set_loader,
marking_network,
device,
verbose=False)
features_target, _ = extract_features(test_set_loader,
target_network,
device,
verbose=False)
features_marking = features_marking.numpy()
features_target = features_target.numpy()
X, residuals, rank, s = np.linalg.lstsq(features_marking,
features_target)
logger.info(
"Norm of residual: %.4e" %
np.linalg.norm(np.dot(features_marking, X) - features_target)**2)
W = np.matmul(W, X.T)
# Computing scores
W /= np.linalg.norm(W, axis=1, keepdims=True)
scores = np.sum(W * carrier, axis=1)
#print(f"SCORES: {scores}")
logger.info("Mean p-value is at %d times sigma" %
int(scores.mean() * np.sqrt(W.shape[0] * carrier.shape[1])))
logger.info("Epoch of the model: %d" % target_checkpoint["epoch"])
p_vals = [cosine_pvalue(c, d=carrier.shape[1]) for c in list(scores)]
#print(f"Cosine P values: {p_vals}")
#print(f"np.sum(np.log(p_vals)): {np.sum(np.log(p_vals))}")
#logger.info(p_vals)
combined_pval = combine_pvalues(p_vals)[1]
logger.info(f"log10(p)={np.log10(combined_pval)}")
elapsed_time = t.stop()
logger.info("Total took %.2f" % (elapsed_time))
return (scores, p_vals, combined_pval)
def main(device,
mp_args,
experiment_name,
optimizer,
output_directory_root="experiments/resnet18_distributed",
lr_scheduler=None,
epochs=150,
batch_size=512,
num_workers=1):
global_rank = mp_args.nr * mp_args.gpus + device
dist.init_process_group(backend='nccl',
init_method='env://',
world_size=mp_args.world_size,
rank=global_rank)
output_directory = os.path.join(output_directory_root, experiment_name,
f"rank_{global_rank}")
if not os.path.isdir(output_directory):
os.makedirs(output_directory, exist_ok=True)
# Setup regular log file + tensorboard
logfile_path = os.path.join(output_directory, "logfile.txt")
setup_logger_tqdm(logfile_path)
tensorboard_log_directory = os.path.join("runs", "resnet18_distributed",
experiment_name,
f"rank_{global_rank}")
tensorboard_summary_writer = SummaryWriter(
log_dir=tensorboard_log_directory)
# Datasets and Loaders
train_set_loader, test_set_loader = get_data_loaders(
mp_args.world_size, global_rank, batch_size, num_workers)
# Create Model & Optimizer (uses Partial Functions)
model = torchvision.models.resnet18(pretrained=False, num_classes=10)
model.to(device)
optimizer = optimizer(model.parameters())
model = nn.parallel.DistributedDataParallel(model, device_ids=[device])
if lr_scheduler:
lr_scheduler = lr_scheduler(optimizer)
logger.info("=========== Commencing Training ===========")
logger.info(f"Epoch Count: {epochs}")
logger.info(f"Batch Size: {batch_size}")
# Load Checkpoint
checkpoint_file_path = os.path.join(output_directory, "checkpoint.pth")
start_epoch = 0
if os.path.exists(checkpoint_file_path):
logger.info("Checkpoint Found - Loading!")
checkpoint = torch.load(checkpoint_file_path)
logger.info(f"Last completed epoch: {checkpoint['epoch']}")
logger.info(f"Average Train Loss: {checkpoint['train_loss']}")
logger.info(f"Top-1 Train Accuracy: {checkpoint['train_accuracy']}")
logger.info(f"Top-1 Test Accuracy: {checkpoint['test_accuracy']}")
start_epoch = checkpoint["epoch"] + 1
logger.info(f"Resuming at epoch {start_epoch}")
model.load_state_dict(checkpoint["model_state_dict"])
optimizer.load_state_dict(checkpoint["optimizer_state_dict"])
if lr_scheduler:
lr_scheduler.load_state_dict(checkpoint["lr_scheduler_state_dict"])
else:
logger.info("No checkpoint found, starting from scratch.")
# Training Loop
t = Timer()
for epoch in range(start_epoch, epochs):
t.start()
logger.info("-" * 10)
logger.info(f"Epoch {epoch}")
logger.info("-" * 10)
train_loss, train_accuracy = train_model(device, model,
train_set_loader, optimizer)
tensorboard_summary_writer.add_scalar("train_loss", train_loss, epoch)
tensorboard_summary_writer.add_scalar("train_accuracy", train_accuracy,
epoch)
test_accuracy = test_model(device, model, test_set_loader, optimizer)
tensorboard_summary_writer.add_scalar("test_accuracy", test_accuracy,
epoch)
scheduler_dict = None
if lr_scheduler:
lr_scheduler.step()
scheduler_dict = lr_scheduler.state_dict()
# Save Checkpoint
logger.info("Saving checkpoint.")
torch.save(
{
'epoch': epoch,
'model_state_dict': model.state_dict(),
'optimizer_state_dict': optimizer.state_dict(),
'lr_scheduler_state_dict': scheduler_dict,
'train_loss': train_loss,
'train_accuracy': train_accuracy,
'test_accuracy': test_accuracy
}, checkpoint_file_path)
elapsed_time = t.stop()
logger.info(f"End of epoch {epoch}, took {elapsed_time:0.4f} seconds.")
logger.info(f"Average Train Loss: {train_loss}")
logger.info(f"Top-1 Train Accuracy: {train_accuracy}")
logger.info(f"Top-1 Test Accuracy: {test_accuracy}")
logger.info("")
