def test_aum_finalize(tmp_path, aum_data):
inputs, outputs = aum_data
save_dir = tmp_path.as_posix()
aum_calculator = AUMCalculator(save_dir=save_dir, compressed=False)
for data in inputs:
aum_calculator.update(data['logits'], data['targets'],
data['sample_ids'])
aum_calculator.finalize()
final_vals = pd.read_csv(os.path.join(save_dir, 'aum_values.csv'))
detailed_vals = pd.read_csv(os.path.join(save_dir, 'full_aum_records.csv'))
# Lets first verify detailed vals
records = []
for output in outputs:
records.extend(output.values())
expected_detailed_vals = pd.DataFrame([
asdict(record) for record in records
]).sort_values(by=['sample_id', 'num_measurements']).reset_index(drop=True)
assert detailed_vals.equals(expected_detailed_vals)
# Now lets verfiy the final vals
final_dict = {record.sample_id: record.aum for record in records}
expected_final_vals = []
for key, val in final_dict.items():
expected_final_vals.append({'sample_id': key, 'aum': val})
expected_final_vals = pd.DataFrame(expected_final_vals).sort_values(
by='aum', ascending=False).reset_index(drop=True)
assert final_vals.equals(expected_final_vals)
def train(self,
num_epochs=300,
batch_size=256,
test_at_end=True,
lr=0.1,
wd=1e-4,
momentum=0.9,
lr_drops=[0.5, 0.75],
aum_wtr=False,
rand_weight=False,
**kwargs):
"""
Training script
:param int num_epochs: (default 300)
:param int batch_size: (default 256)
:param float lr: Learning rate
:param float wd: Weight decay
:param float momentum: Momentum
:param list lr_drops: When to drop the learning rate (by a factor of 10) as a percentage of total training time.
:param str aum_wtr: (optional) The path of the model/results directory to load AUM_WTR weights from.
:param bool rand_weight (optional, default false): uses rectified normal random weighting if True.
"""
# Model
model = self.model
if torch.cuda.is_available():
model = model.cuda()
if torch.cuda.device_count() > 1:
model = torch.nn.DataParallel(model).cuda()
# Optimizer
optimizer = torch.optim.SGD(model.parameters(),
lr=lr,
weight_decay=wd,
momentum=momentum,
nesterov=True)
milestones = [
int(lr_drop * num_epochs) for lr_drop in (lr_drops or [])
]
scheduler = torch.optim.lr_scheduler.MultiStepLR(optimizer,
milestones=milestones,
gamma=0.1)
logging.info(f"\nOPTIMIZER:\n{optimizer}")
logging.info(f"SCHEDULER:\n{scheduler.milestones}")
# Initialize AUM caluclator object
aum_calculator = AUMCalculator(save_dir=self.savedir, compressed=False)
train_data = OrderedDict()
train_data["train_indices"] = self.train_set.indices
train_data["valid_indices"] = (self.valid_set.indices
if self.valid_set is not None else
torch.tensor([], dtype=torch.long))
train_data["true_targets"] = self.train_set.targets
train_data["assigned_targets"] = self.train_set.assigned_targets
# Storage to log results
results = []
# Train model
best_error = 1
for epoch in range(num_epochs):
train_results = self.train_epoch(model=model,
optimizer=optimizer,
epoch=epoch,
num_epochs=num_epochs,
batch_size=batch_size,
aum_calculator=aum_calculator,
aum_wtr=aum_wtr,
rand_weight=rand_weight,
**kwargs)
if self.valid_set is not None:
valid_results = self.test(model=model,
split="valid",
batch_size=batch_size,
epoch=epoch,
**kwargs)
else:
valid_results = self.test(model,
split="test",
batch_size=batch_size,
epoch=epoch,
**kwargs)
scheduler.step()
# Determine if model is the best
if self.valid_set is not None:
self.save()
elif best_error > valid_results.error:
best_error = valid_results.error
logging.info('New best error: %.4f' % valid_results.error)
self.save()
# Log results
logging.info(f"\nTraining {repr(train_results)}")
logging.info(f"\nValidation {repr(valid_results)}")
logging.info('')
results.append(
OrderedDict([("epoch", f"{epoch + 1:03d}"),
*[(f"train_{field}", val)
for field, val in train_results.items()],
*[(f"valid_{field}", val)
for field, val in valid_results.items()]]))
pd.DataFrame(results).set_index("epoch").to_csv(
os.path.join(self.savedir, "train_log.csv"))
# Save metadata around train set (like which labels were flipped)
torch.save(train_data, os.path.join(self.savedir,
"train_data.pth"))
# Once we're finished training calculate aum
aum_calculator.finalize()
# Maybe test (last epoch)
if test_at_end and self.valid_set is not None:
test_results = self.test(model=model, **kwargs)
logging.info(f"\nTest (no early stopping) {repr(test_results)}")
shutil.copyfile(
os.path.join(self.savedir, "results_test.csv"),
os.path.join(self.savedir, "results_test_noearlystop.csv"))
results.append(
OrderedDict([(f"test_{field}", val)
for field, val in test_results.items()]))
pd.DataFrame(results).set_index("epoch").to_csv(
os.path.join(self.savedir, "train_log.csv"))
# Load best model
self.save(suffix=".last")
self.load()
# Maybe test (best epoch)
if test_at_end and self.valid_set is not None:
test_results = self.test(model=model, **kwargs)
logging.info(f"\nEarly Stopped Model Test {repr(test_results)}")
results.append(
OrderedDict([(f"test_best_{field}", val)
for field, val in test_results.items()]))
pd.DataFrame(results).set_index("epoch").to_csv(
os.path.join(self.savedir, "train_log.csv"))
return self
def main(args):
pprint(vars(args))
# Setup experiment folder structure
# Create output folder if it doesn't exist
Path(args.output_dir).mkdir(parents=True, exist_ok=True)
# save out args
with open(os.path.join(args.output_dir, 'args.txt'), 'w+') as f:
pprint(vars(args), f)
# Setup summary writer
summary_writer = SummaryWriter(
log_dir=os.path.join(args.output_dir, 'tb_logs'))
# Set seeds
set_seed(42)
# Load dataset
# Data transforms
mean = [0.5071, 0.4867, 0.4408]
stdv = [0.2675, 0.2565, 0.2761]
train_transforms = transforms.Compose([
transforms.RandomCrop(32, padding=4),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize(mean=mean, std=stdv),
])
test_transforms = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize(mean=mean, std=stdv),
])
# Datasets
train_set = datasets.CIFAR100(args.data_dir,
train=True,
transform=train_transforms,
download=True)
val_set = datasets.CIFAR100(args.data_dir,
train=True,
transform=test_transforms)
test_set = datasets.CIFAR100(args.data_dir,
train=False,
transform=test_transforms)
indices = torch.randperm(len(train_set))
train_indices = indices[:len(indices) - args.valid_size]
valid_indices = indices[len(indices) - args.valid_size:]
train_set = torch.utils.data.Subset(train_set, train_indices)
val_set = torch.utils.data.Subset(val_set, valid_indices)
train_set = DatasetWithIndex(train_set)
val_set = DatasetWithIndex(val_set)
test_set = DatasetWithIndex(test_set)
val_loader = DataLoader(val_set,
batch_size=args.val_batch_size,
shuffle=False,
pin_memory=(torch.cuda.is_available()))
test_loader = DataLoader(test_set,
batch_size=args.val_batch_size,
shuffle=False,
pin_memory=(torch.cuda.is_available()))
# Load Model
device = 'cuda' if torch.cuda.is_available() else 'cpu'
model = resnet34(num_classes=100)
model = model.to(device)
num_params = sum(x.numel() for x in model.parameters() if x.requires_grad)
print(model)
f'Number of parameters: {num_params}'
# Create optimizer & lr scheduler
parameters = [p for p in model.parameters() if p.requires_grad]
optimizer = torch.optim.SGD(parameters,
lr=args.learning_rate,
momentum=args.momentum,
nesterov=True)
milestones = [0.5 * args.num_epochs, 0.75 * args.num_epochs]
scheduler = torch.optim.lr_scheduler.MultiStepLR(optimizer,
milestones=milestones,
gamma=0.1)
# Keep track of AUM
aum_calculator = AUMCalculator(args.output_dir,
compressed=(not args.detailed_aum))
# Keep track of things
global_step = 0
best_error = math.inf
print('Beginning training')
for epoch in range(args.num_epochs):
train_loader = DataLoader(train_set,
batch_size=args.train_batch_size,
shuffle=True,
pin_memory=(torch.cuda.is_available()),
num_workers=0)
train_metrics = {
'loss': AverageMeter(),
'error': AverageMeter(),
'batch_time': AverageMeter()
}
num_batches = len(train_loader)
for batch_step, batch in enumerate(train_loader):
train_step(args, summary_writer, train_metrics, aum_calculator,
args.log_interval, batch_step, num_batches, batch,
epoch, args.num_epochs, global_step, model, optimizer,
device)
global_step += 1
scheduler.step()
val_metrics = {
'loss': AverageMeter(),
'error': AverageMeter(),
'batch_time': AverageMeter()
}
num_batches = len(val_loader)
for batch_step, batch in enumerate(val_loader):
eval_step(args, 'VAL', val_metrics, args.log_interval, batch_step,
num_batches, batch, epoch, args.num_epochs, model,
device)
# log eval metrics to tensorboard
summary_writer.add_scalar('val/error', val_metrics['error'].avg,
global_step)
summary_writer.add_scalar('val/loss', val_metrics['loss'].avg,
global_step)
summary_writer.add_scalar('val/batch_time',
val_metrics['batch_time'].avg, global_step)
# Save best model
if val_metrics['error'].avg < best_error:
best_error = val_metrics['error'].avg
torch.save(model.state_dict(),
os.path.join(args.output_dir, 'best.pt'))
# Finalize aum calculator
aum_calculator.finalize()
# Eval best model on on test set
model.load_state_dict(torch.load(os.path.join(args.output_dir, 'best.pt')))
test_metrics = {
'loss': AverageMeter(),
'error': AverageMeter(),
'batch_time': AverageMeter()
}
num_batches = len(test_loader)
for batch_step, batch in enumerate(test_loader):
eval_step(args, 'TEST', test_metrics, args.log_interval, batch_step,
num_batches, batch, -1, -1, model, device)
# log eval metrics to tensorboard
summary_writer.add_scalar('test/error', test_metrics['error'].avg,
global_step)
summary_writer.add_scalar('test/loss', test_metrics['loss'].avg,
global_step)
summary_writer.add_scalar('test/batch_time',
test_metrics['batch_time'].avg, global_step)
# log test metrics to console
results = '\t'.join([
'FINAL TEST RESULTS',
f'Loss: {test_metrics["loss"].avg:.3f}',
f'Error: {test_metrics["error"].avg:.3f}',
])
print(results)