def get_model() -> common.LernomaticModel:
if GLOBAL_OPTS['model'] == 'resnet':
model = resnets.WideResnet(depth=GLOBAL_OPTS['resnet_depth'],
num_classes=10,
input_channels=3)
elif GLOBAL_OPTS['model'] == 'cifar':
model = cifar.CIFAR10Net()
else:
raise ValueError('Unknown model type [%s]' % str(GLOBAL_OPTS['model']))
return model
def test_history_extend(self) -> None:
test_checkpoint = 'checkpoint/test_history_extend.pkl'
test_history = 'checkpoint/test_history_extend_history.pkl'
model = cifar.CIFAR10Net()
# TODO : adjust this so we don't need 10 epochs of training
# Get trainer object
test_num_epochs = 10
trainer = cifar_trainer.CIFAR10Trainer(
model,
save_every = 0,
print_every = 50,
device_id = util.get_device_id(),
# loader options,
num_epochs = test_num_epochs,
learning_rate = 3e-4,
batch_size = 64,
num_workers = self.test_num_workers,
)
print('Training original model')
trainer.train()
trainer.save_checkpoint(test_checkpoint)
trainer.save_history(test_history)
# Load a new trainer, train for another 10 epochs (20 total)
extend_trainer = cifar_trainer.CIFAR10Trainer(
model,
save_every = 0,
print_every = 50,
device_id = util.get_device_id(),
# loader options,
num_epochs = 10,
learning_rate = 3e-4,
batch_size = 64,
num_workers = self.test_num_workers,
)
print('Loading checkpoint [%s] into extend trainer...' % str(test_checkpoint))
extend_trainer.load_checkpoint(test_checkpoint)
print('Loading history [%s] into extend trainer...' % str(test_history))
extend_trainer.load_history(test_history)
# Check history before extending
assert trainer.device_id == extend_trainer.device_id
print('extend_trainer device : %s' % str(extend_trainer.device))
assert test_num_epochs == extend_trainer.num_epochs
assert 10 == extend_trainer.cur_epoch
assert extend_trainer.loss_history is not None
assert (10 * len(extend_trainer.train_loader)) == len(extend_trainer.loss_history)
extend_trainer.set_num_epochs(20)
assert(20 * len(extend_trainer.train_loader) == len(extend_trainer.loss_history))
for i in range(10 * len(extend_trainer.train_loader)):
print('Checking loss iter [%d / %d]' % (i, 20 * len(extend_trainer.train_loader)), end='\r')
assert trainer.loss_history[i] == extend_trainer.loss_history[i]
extend_trainer.train()
def get_model(model_type:str,
num_classes:int=10,
input_channels:int=3,
depth:int=58) -> common.LernomaticModel:
if model_type == 'resnet':
model = resnets.WideResnet(
depth=depth,
num_classes=num_classes,
input_channels = input_channels
)
elif model_type == 'cifar':
model = cifar.CIFAR10Net()
else:
raise ValueError('Unknown model type [%s]' % str(GLOBAL_OPTS['model']))
return model
def get_trainer() -> cifar_trainer.CIFAR10Trainer:
# get a model to test on and its corresponding trainer
model = cifar.CIFAR10Net()
trainer = cifar_trainer.CIFAR10Trainer(
model,
# turn off checkpointing
save_every = 0,
print_every = GLOBAL_TEST_PARAMS['test_print_every'],
# data options
batch_size = GLOBAL_TEST_PARAMS['test_batch_size'],
# training options
learning_rate = GLOBAL_TEST_PARAMS['test_learning_rate'],
num_epochs = GLOBAL_TEST_PARAMS['train_num_epochs'],
device_id = util.get_device_id(),
)
return trainer
def get_trainer(learning_rate: float = None) -> cifar_trainer.CIFAR10Trainer:
if learning_rate is not None:
test_learning_rate = learning_rate
else:
test_learning_rate = GLOBAL_OPTS['learning_rate']
model = cifar.CIFAR10Net()
trainer = cifar_trainer.CIFAR10Trainer(
model,
# data options
batch_size=GLOBAL_OPTS['batch_size'],
num_workers=GLOBAL_OPTS['num_workers'],
num_epochs=GLOBAL_OPTS['num_epochs'],
# set initial learning rate
learning_rate=test_learning_rate,
# other options
save_every=0,
print_every=200,
device_id=util.get_device_id(),
verbose=GLOBAL_OPTS['verbose'])
return trainer
def test_train(self) -> None:
test_checkpoint = 'checkpoint/trainer_train_test.pkl'
model = cifar.CIFAR10Net()
# Get trainer object
trainer = cifar_trainer.CIFAR10Trainer(
model,
save_every = 0,
print_every = 50,
device_id = util.get_device_id(),
# loader options,
num_epochs = self.test_num_epochs,
learning_rate = 3e-4,
batch_size = 128,
num_workers = self.test_num_workers,
)
if self.verbose:
print('Created trainer object')
print(trainer)
# train for one epoch
trainer.train()
trainer.save_checkpoint(test_checkpoint)
fig, ax = get_figure_subplots()
vis_loss_history.plot_train_history_2subplots(
ax,
trainer.loss_history,
acc_history = trainer.acc_history,
cur_epoch = trainer.cur_epoch,
iter_per_epoch = trainer.iter_per_epoch,
loss_title = 'CIFAR-10 Trainer Test loss',
acc_title = 'CIFAR-10 Trainer Test accuracy'
)
fig.savefig('figures/trainer_train_test_history.png', bbox_inches='tight')
def get_model() -> common.LernomaticModel:
model = cifar.CIFAR10Net()
return model
def main() -> None:
# Get a model
model = cifar.CIFAR10Net()
# Get a trainer
trainer = cifar_trainer.CIFAR10Trainer(
model,
# training parameters
batch_size=GLOBAL_OPTS['batch_size'],
num_epochs=GLOBAL_OPTS['num_epochs'],
learning_rate=GLOBAL_OPTS['learning_rate'],
momentum=GLOBAL_OPTS['momentum'],
weight_decay=GLOBAL_OPTS['weight_decay'],
# device
device_id=GLOBAL_OPTS['device_id'],
# checkpoint
checkpoint_dir=GLOBAL_OPTS['checkpoint_dir'],
checkpoint_name=GLOBAL_OPTS['checkpoint_name'],
# display,
print_every=GLOBAL_OPTS['print_every'],
save_every=GLOBAL_OPTS['save_every'],
verbose=GLOBAL_OPTS['verbose'])
if GLOBAL_OPTS['tensorboard_dir'] is not None:
writer = tensorboard.SummaryWriter(
log_dir=GLOBAL_OPTS['tensorboard_dir'])
trainer.set_tb_writer(writer)
# Optionally do a search pass here and add a scheduler
if GLOBAL_OPTS['find_lr']:
lr_finder = expr_util.get_lr_finder(trainer)
lr_find_start_time = time.time()
lr_finder.find()
lr_find_min, lr_find_max = lr_finder.get_lr_range()
lr_find_end_time = time.time()
lr_find_total_time = lr_find_end_time - lr_find_start_time
print('Found learning rate range %.4f -> %.4f' %
(lr_find_min, lr_find_max))
print('Total find time [%s] ' %\
str(timedelta(seconds = lr_find_total_time))
)
# Now get a scheduler
stepsize = trainer.get_num_epochs() * len(trainer.train_loader) // 2
# get scheduler
lr_scheduler = expr_util.get_scheduler(
lr_find_min,
lr_find_max,
stepsize,
sched_type='TriangularScheduler')
trainer.set_lr_scheduler(lr_scheduler)
# train the model
train_start_time = time.time()
trainer.train()
train_end_time = time.time()
train_total_time = train_end_time - train_start_time
print('Total training time [%s] (%d epochs) %s' %\
(repr(trainer), trainer.cur_epoch,
str(timedelta(seconds = train_total_time)))
)
# Visualise the output
train_fig, train_ax = vis_loss_history.get_figure_subplots()
vis_loss_history.plot_train_history_2subplots(
train_ax,
trainer.get_loss_history(),
acc_history=trainer.get_acc_history(),
cur_epoch=trainer.cur_epoch,
iter_per_epoch=trainer.iter_per_epoch,
loss_title='CIFAR-10 Training Loss',
acc_title='CIFAR-10 Training Accuracy ')
train_fig.savefig(GLOBAL_OPTS['fig_name'], bbox_inches='tight')
def test_save_load_checkpoint(self) -> None:
test_checkpoint = 'checkpoint/trainer_test_save_load.pkl'
# get a model
model = cifar.CIFAR10Net()
# get a trainer
src_tr = cifar_trainer.CIFAR10Trainer(
model,
num_epochs = self.test_num_epochs,
save_every = 0,
device_id = util.get_device_id(),
batch_size = self.test_batch_size,
num_workers = self.test_num_workers
)
if self.verbose:
print('Created trainer object')
print(src_tr)
# train for one epoch
src_tr.train()
src_tr.save_checkpoint(test_checkpoint)
# Make a new trainer and load all parameters into that
# I guess we need to put some kind of loader and model here...
dst_tr = cifar_trainer.CIFAR10Trainer(
model,
device_id = util.get_device_id()
)
dst_tr.load_checkpoint(test_checkpoint)
# Test object parameters
assert src_tr.num_epochs == dst_tr.num_epochs
assert src_tr.learning_rate == dst_tr.learning_rate
assert src_tr.weight_decay == dst_tr.weight_decay
assert src_tr.print_every == dst_tr.print_every
assert src_tr.save_every == dst_tr.save_every
assert src_tr.device_id == dst_tr.device_id
print('\t Comparing model parameters ')
src_model_params = src_tr.get_model_params()
dst_model_params = dst_tr.get_model_params()
assert len(src_model_params.items()) == len(dst_model_params.items())
# p1, p2 are k,v tuple pairs of each model parameters
# k = str
# v = torch.Tensor
for n, (p1, p2) in enumerate(zip(src_model_params.items(), dst_model_params.items())):
assert p1[0] == p2[0]
print('Checking parameter %s [%d/%d] \t\t' % (str(p1[0]), n+1, len(src_model_params.items())), end='\r')
assert torch.equal(p1[1], p2[1]) == True
print('\n ...done')
# Test loss history
val_loss_history = 'test_save_load_history.pkl'
src_tr.save_history(val_loss_history)
dst_tr.load_history(val_loss_history)
print('\t Comparing loss history....')
assert src_tr.loss_iter == dst_tr.loss_iter
for n in range(src_tr.loss_iter):
print('Checking loss element [%d/%d]' % (n, src_tr.loss_iter), end='\r')
assert src_tr.loss_history[n] == dst_tr.loss_history[n]
# Try to train for another epoch
dst_tr.set_num_epochs(src_tr.num_epochs+1)
assert dst_tr.num_epochs == src_tr.num_epochs+1
dst_tr.train()
assert src_tr.num_epochs+1 == dst_tr.cur_epoch
print('\n ...done')
os.remove(test_checkpoint)
def test_save_load_device_map(self) -> None:
test_checkpoint = 'checkpoint/trainer_save_load_device_map.pkl'
test_history = 'checkpoint/trainer_save_load_device_map_history.pkl'
model = cifar.CIFAR10Net()
# Get trainer object
src_tr = cifar_trainer.CIFAR10Trainer(
model,
save_every = 0,
print_every = 50,
device_id = util.get_device_id(),
# loader options,
num_epochs = self.test_num_epochs,
batch_size = self.test_batch_size,
num_workers = self.test_num_workers,
)
if self.verbose:
print('Created trainer object')
print(src_tr)
# train for one epoch
src_tr.train()
src_tr.save_checkpoint(test_checkpoint)
assert src_tr.acc_history is not None
src_tr.save_history(test_history)
# Now try to load a checkpoint and ensure that there is an
# acc history attribute that is not None
dst_tr = cifar_trainer.CIFAR10Trainer(
model,
device_id = -1,
verbose = self.verbose
)
dst_tr.load_checkpoint(test_checkpoint)
# Test object parameters
assert src_tr.num_epochs == dst_tr.num_epochs
assert src_tr.learning_rate == dst_tr.learning_rate
assert src_tr.weight_decay == dst_tr.weight_decay
assert src_tr.print_every == dst_tr.print_every
assert src_tr.save_every == dst_tr.save_every
print('\t Comparing model parameters ')
src_model_params = src_tr.get_model_params()
dst_model_params = dst_tr.get_model_params()
assert len(src_model_params.items()) == len(dst_model_params.items())
print('\t Comparing model parameters ')
src_model_params = src_tr.get_model_params()
dst_model_params = dst_tr.get_model_params()
assert len(src_model_params.items()) == len(dst_model_params.items())
# p1, p2 are k,v tuple pairs of each model parameters
# k = str
# v = torch.Tensor
print('Checking that tensors from checkpoint have been tranferred to new device')
for n, (p1, p2) in enumerate(zip(src_model_params.items(), dst_model_params.items())):
assert p1[0] == p2[0]
print('Checking parameter %s [%d/%d] \t\t' % (str(p1[0]), n, len(src_model_params.items())), end='\r')
assert 'cpu' == p2[1].device.type
print('\n ...done')
# Test loss history
dst_tr.load_history(test_history)
assert dst_tr.acc_history is not None
print('\t Comparing loss history....')
assert src_tr.loss_iter == dst_tr.loss_iter
for n in range(src_tr.loss_iter):
print('Checking loss element [%d/%d]' % (n, src_tr.loss_iter), end='\r')
assert src_tr.loss_history[n] == dst_tr.loss_history[n]