def trainval(exp_dict, savedir_base, reset=False):
# bookkeeping
# ---------------
# get experiment directory
exp_id = hu.hash_dict(exp_dict)
savedir = os.path.join(savedir_base, exp_id)
if reset:
# delete and backup experiment
hc.delete_experiment(savedir, backup_flag=True)
# create folder and save the experiment dictionary
os.makedirs(savedir, exist_ok=True)
hu.save_json(os.path.join(savedir, 'exp_dict.json'), exp_dict)
pprint.pprint(exp_dict)
print('Experiment saved in %s' % savedir)
# set seed
# ---------------
seed = 42 + exp_dict['runs']
np.random.seed(seed)
torch.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
# Dataset
# -----------
# train loader
train_set = datasets.get_dataset(dataset_name=exp_dict["dataset"],
train_flag=True,
datadir=savedir_base,
exp_dict=exp_dict)
train_loader = torch.utils.data.DataLoader(
train_set,
drop_last=True,
shuffle=True,
batch_size=exp_dict["batch_size"])
# val set
val_set = datasets.get_dataset(dataset_name=exp_dict["dataset"],
train_flag=False,
datadir=savedir_base,
exp_dict=exp_dict)
# Model
# -----------
model = models.get_model(exp_dict["model"], train_set=train_set).cuda()
# Choose loss and metric function
loss_function = metrics.get_metric_function(exp_dict["loss_func"])
# Compute fstar
# -------------
if exp_dict['opt'].get('fstar_flag'):
ut.compute_fstar(train_set, loss_function, savedir_base, exp_dict)
# Load Optimizer
n_batches_per_epoch = len(train_set) / float(exp_dict["batch_size"])
opt = optimizers.get_optimizer(opt_dict=exp_dict["opt"],
params=model.parameters(),
n_batches_per_epoch=n_batches_per_epoch)
# Checkpoint
# -----------
model_path = os.path.join(savedir, 'model.pth')
score_list_path = os.path.join(savedir, 'score_list.pkl')
opt_path = os.path.join(savedir, 'opt_state_dict.pth')
if os.path.exists(score_list_path):
# resume experiment
score_list = hu.load_pkl(score_list_path)
model.load_state_dict(torch.load(model_path))
opt.load_state_dict(torch.load(opt_path))
s_epoch = score_list[-1]['epoch'] + 1
else:
# restart experiment
score_list = []
s_epoch = 0
# Train & Val
# ------------
print('Starting experiment at epoch %d/%d' %
(s_epoch, exp_dict['max_epoch']))
for e in range(s_epoch, exp_dict['max_epoch']):
# Set seed
seed = e + exp_dict['runs']
np.random.seed(seed)
torch.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
score_dict = {}
# Compute train loss over train set
score_dict["train_loss"] = metrics.compute_metric_on_dataset(
model, train_set, metric_name=exp_dict["loss_func"])
# Compute val acc over val set
score_dict["val_acc"] = metrics.compute_metric_on_dataset(
model, val_set, metric_name=exp_dict["acc_func"])
# Train over train loader
model.train()
print("%d - Training model with %s..." % (e, exp_dict["loss_func"]))
# train and validate
s_time = time.time()
for batch in tqdm.tqdm(train_loader):
images, labels = batch["images"].cuda(), batch["labels"].cuda()
opt.zero_grad()
# closure
def closure():
return loss_function(model, images, labels, backwards=True)
opt.step(closure)
e_time = time.time()
# Record metrics
score_dict["epoch"] = e
score_dict["step_size"] = opt.state["step_size"]
score_dict["step_size_avg"] = opt.state["step_size_avg"]
score_dict["n_forwards"] = opt.state["n_forwards"]
score_dict["n_backwards"] = opt.state["n_backwards"]
score_dict["grad_norm"] = opt.state["grad_norm"]
score_dict["batch_size"] = train_loader.batch_size
score_dict["train_epoch_time"] = e_time - s_time
score_list += [score_dict]
# Report and save
print(pd.DataFrame(score_list).tail())
hu.save_pkl(score_list_path, score_list)
hu.torch_save(model_path, model.state_dict())
hu.torch_save(opt_path, opt.state_dict())
print("Saved: %s" % savedir)
print('Experiment completed')
def trainval_svrg(exp_dict, savedir, datadir, metrics_flag=True):
'''
SVRG-specific training and validation loop.
'''
pprint.pprint(exp_dict)
# Load Train Dataset
train_set = datasets.get_dataset(dataset_name=exp_dict["dataset"],
train_flag=True,
datadir=datadir,
exp_dict=exp_dict)
train_loader = DataLoader(train_set,
drop_last=False,
shuffle=True,
batch_size=exp_dict["batch_size"])
# Load Val Dataset
val_set = datasets.get_dataset(dataset_name=exp_dict["dataset"],
train_flag=False,
datadir=datadir,
exp_dict=exp_dict)
# Load model
model = models.get_model(exp_dict["model"], train_set=train_set).cuda()
# Choose loss and metric function
loss_function = metrics.get_metric_function(exp_dict["loss_func"])
# lookup the learning rate
lr = get_svrg_step_size(exp_dict)
# Load Optimizer
opt = get_svrg_optimizer(model,
loss_function,
train_loader=train_loader,
lr=lr)
# Resume from last saved state_dict
if (not os.path.exists(savedir + "/run_dict.pkl")
or not os.path.exists(savedir + "/score_list.pkl")):
ut.save_pkl(savedir + "/run_dict.pkl", {"running": 1})
score_list = []
s_epoch = 0
else:
score_list = ut.load_pkl(savedir + "/score_list.pkl")
model.load_state_dict(torch.load(savedir + "/model_state_dict.pth"))
opt.load_state_dict(torch.load(savedir + "/opt_state_dict.pth"))
s_epoch = score_list[-1]["epoch"] + 1
for epoch in range(s_epoch, exp_dict["max_epoch"]):
score_dict = {"epoch": epoch}
if metrics_flag:
# 1. Compute train loss over train set
score_dict["train_loss"] = metrics.compute_metric_on_dataset(
model, train_set, metric_name=exp_dict["loss_func"])
# 2. Compute val acc over val set
score_dict["val_acc"] = metrics.compute_metric_on_dataset(
model, val_set, metric_name=exp_dict["acc_func"])
# 3. Train over train loader
model.train()
print("%d - Training model with %s..." %
(epoch, exp_dict["loss_func"]))
s_time = time.time()
for images, labels in tqdm.tqdm(train_loader):
images, labels = images.cuda(), labels.cuda()
opt.zero_grad()
closure = lambda svrg_model: loss_function(
svrg_model, images, labels, backwards=True)
opt.step(closure)
e_time = time.time()
# Record step size and batch size
score_dict["step_size"] = opt.state["step_size"]
score_dict["batch_size"] = train_loader.batch_size
score_dict["train_epoch_time"] = e_time - s_time
# Add score_dict to score_list
score_list += [score_dict]
# Report and save
print(pd.DataFrame(score_list).tail())
ut.save_pkl(savedir + "/score_list.pkl", score_list)
ut.torch_save(savedir + "/model_state_dict.pth", model.state_dict())
ut.torch_save(savedir + "/opt_state_dict.pth", opt.state_dict())
print("Saved: %s" % savedir)
return score_list
def newminimum(exp_id,
savedir_base,
datadir,
name,
exp_dict,
metrics_flag=True):
# bookkeeping
# ---------------
# get experiment directory
old_modeldir = os.path.join(savedir_base, exp_id)
savedir = os.path.join(savedir_base, exp_id, name)
old_exp_dict = hu.load_json(os.path.join(old_modeldir, 'exp_dict.json'))
# TODO: compare exp dict for possible errors:
# optimizer have to be the same
# same network, dataset
# create folder and save the experiment dictionary
os.makedirs(savedir, exist_ok=True)
hu.save_json(os.path.join(savedir, 'exp_dict.json'), exp_dict)
pprint.pprint(exp_dict)
print('Experiment saved in %s' % savedir)
# set seed
# ---------------
seed = 42 + exp_dict['runs']
np.random.seed(seed)
torch.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
# Dataset
# -----------
# Load Train Dataset
train_set = datasets.get_dataset(dataset_name=exp_dict["dataset"],
train_flag=True,
datadir=datadir,
exp_dict=exp_dict)
train_loader = torch.utils.data.DataLoader(
train_set,
drop_last=True,
shuffle=True,
batch_size=exp_dict["batch_size"])
# Load Val Dataset
val_set = datasets.get_dataset(dataset_name=exp_dict["dataset"],
train_flag=False,
datadir=datadir,
exp_dict=exp_dict)
# Model
# -----------
model = models.get_model(exp_dict["model"], train_set=train_set)
# Choose loss and metric function
loss_function = metrics.get_metric_function(exp_dict["loss_func"])
# Load Optimizer
n_batches_per_epoch = len(train_set) / float(exp_dict["batch_size"])
opt = optimizers.get_optimizer(opt=exp_dict["opt"],
params=model.parameters(),
n_batches_per_epoch=n_batches_per_epoch)
# Checkpoint
# -----------
model_path = os.path.join(savedir, 'model.pth')
score_list_path = os.path.join(savedir, 'score_list.pkl')
opt_path = os.path.join(savedir, 'opt_state_dict.pth')
old_model_path = os.path.join(old_modeldir, 'model.pth')
old_score_list_path = os.path.join(old_modeldir, 'score_list.pkl')
old_opt_path = os.path.join(old_modeldir, 'opt_state_dict.pth')
score_list = hu.load_pkl(old_score_list_path)
model.load_state_dict(torch.load(old_model_path))
opt.load_state_dict(torch.load(old_opt_path))
s_epoch = score_list[-1]['epoch'] + 1
# save current model state for comparison
minimum = []
for param in model.parameters():
minimum.append(param.clone())
# Train & Val
# ------------
print('Starting experiment at epoch %d/%d' %
(s_epoch, exp_dict['max_epoch']))
for epoch in range(s_epoch, exp_dict['max_epoch']):
# Set seed
np.random.seed(exp_dict['runs'] + epoch)
torch.manual_seed(exp_dict['runs'] + epoch)
# torch.cuda.manual_seed_all(exp_dict['runs']+epoch) not needed since no cuda available
score_dict = {"epoch": epoch}
if metrics_flag:
# 1. Compute train loss over train set
score_dict["train_loss"] = metrics.compute_metric_on_dataset(
model, train_set, metric_name='softmax_loss')
# metric_name=exp_dict["loss_func"])
# TODO: which loss should be used? (normal or with reguralizer?)
# 2. Compute val acc over val set
score_dict["val_acc"] = metrics.compute_metric_on_dataset(
model, val_set, metric_name=exp_dict["acc_func"])
# 3. Train over train loader
model.train()
print("%d - Training model with %s..." %
(epoch, exp_dict["loss_func"]))
s_time = time.time()
for images, labels in tqdm.tqdm(train_loader):
# images, labels = images.cuda(), labels.cuda() no cuda available
opt.zero_grad()
loss = loss_function(model, images, labels, minimum,
0.1) # just works for custom loss function
loss.backward()
opt.step()
e_time = time.time()
# Record metrics
score_dict["step_size"] = opt.state["step_size"]
score_dict["n_forwards"] = opt.state["n_forwards"]
score_dict["n_backwards"] = opt.state["n_backwards"]
score_dict["batch_size"] = train_loader.batch_size
score_dict["train_epoch_time"] = e_time - s_time
score_list += [score_dict]
# Report and save
print(pd.DataFrame(score_list).tail())
hu.save_pkl(score_list_path, score_list)
hu.torch_save(model_path, model.state_dict())
hu.torch_save(opt_path, opt.state_dict())
print("Saved: %s" % savedir)
with torch.nograd():
print('Current distance: %f',
metrics.computedistance(minimum, model))
print('Experiment completed')
def trainval(exp_dict,
savedir_base,
reset,
metrics_flag=True,
datadir=None,
cuda=False):
# bookkeeping
# ---------------
# get experiment directory
exp_id = hu.hash_dict(exp_dict)
savedir = os.path.join(savedir_base, exp_id)
if reset:
# delete and backup experiment
hc.delete_experiment(savedir, backup_flag=True)
# create folder and save the experiment dictionary
os.makedirs(savedir, exist_ok=True)
hu.save_json(os.path.join(savedir, 'exp_dict.json'), exp_dict)
print(pprint.pprint(exp_dict))
print('Experiment saved in %s' % savedir)
# set seed
# ==================
seed = 42 + exp_dict['runs']
np.random.seed(seed)
torch.manual_seed(seed)
if cuda:
device = 'cuda'
torch.cuda.manual_seed_all(seed)
else:
device = 'cpu'
print('Running on device: %s' % device)
# Dataset
# ==================
train_set = datasets.get_dataset(dataset_name=exp_dict["dataset"],
train_flag=True,
datadir=datadir,
exp_dict=exp_dict)
train_loader = DataLoader(train_set,
drop_last=True,
shuffle=True,
sampler=None,
batch_size=exp_dict["batch_size"])
# Load Val Dataset
val_set = datasets.get_dataset(dataset_name=exp_dict["dataset"],
train_flag=False,
datadir=datadir,
exp_dict=exp_dict)
# Model
# ==================
use_backpack = exp_dict['opt'].get("backpack", False)
model = models.get_model(exp_dict["model"],
train_set=train_set,
backpack=use_backpack).to(device=device)
if use_backpack:
assert exp_dict['opt']['name'] in ['nus_wrapper', 'adaptive_second']
from backpack import extend
model = extend(model)
# Choose loss and metric function
loss_function = metrics.get_metric_function(exp_dict["loss_func"])
# Load Optimizer
# ==============
n_batches_per_epoch = len(train_set) / float(exp_dict["batch_size"])
opt = optimizers.get_optimizer(opt=exp_dict["opt"],
params=model.parameters(),
n_batches_per_epoch=n_batches_per_epoch,
n_train=len(train_set),
train_loader=train_loader,
model=model,
loss_function=loss_function,
exp_dict=exp_dict,
batch_size=exp_dict["batch_size"])
# Checkpointing
# =============
score_list_path = os.path.join(savedir, "score_list.pkl")
model_path = os.path.join(savedir, "model_state_dict.pth")
opt_path = os.path.join(savedir, "opt_state_dict.pth")
if os.path.exists(score_list_path):
# resume experiment
score_list = ut.load_pkl(score_list_path)
if use_backpack:
model.load_state_dict(torch.load(model_path), strict=False)
else:
model.load_state_dict(torch.load(model_path))
opt.load_state_dict(torch.load(opt_path))
s_epoch = score_list[-1]["epoch"] + 1
else:
# restart experiment
score_list = []
s_epoch = 0
# Start Training
# ==============
n_train = len(train_loader.dataset)
n_batches = len(train_loader)
batch_size = train_loader.batch_size
for epoch in range(s_epoch, exp_dict["max_epoch"]):
# Set seed
seed = epoch + exp_dict['runs']
np.random.seed(seed)
torch.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
score_dict = {"epoch": epoch}
# Validate
# --------
if metrics_flag:
# 1. Compute train loss over train set
score_dict["train_loss"] = metrics.compute_metric_on_dataset(
model,
train_set,
metric_name=exp_dict["loss_func"],
batch_size=exp_dict['batch_size'])
# 2. Compute val acc over val set
score_dict["val_acc"] = metrics.compute_metric_on_dataset(
model,
val_set,
metric_name=exp_dict["acc_func"],
batch_size=exp_dict['batch_size'])
# Train
# -----
model.train()
print("%d - Training model with %s..." %
(epoch, exp_dict["loss_func"]))
s_time = time.time()
train_on_loader(model, train_set, train_loader, opt, loss_function,
epoch, use_backpack)
e_time = time.time()
# Record step size and batch size
score_dict["step"] = opt.state.get("step",
0) / int(n_batches_per_epoch)
score_dict["step_size"] = opt.state.get("step_size", {})
score_dict["step_size_avg"] = opt.state.get("step_size_avg", {})
score_dict["n_forwards"] = opt.state.get("n_forwards", {})
score_dict["n_backwards"] = opt.state.get("n_backwards", {})
score_dict["grad_norm"] = opt.state.get("grad_norm", {})
score_dict["batch_size"] = batch_size
score_dict["train_epoch_time"] = e_time - s_time
score_dict.update(opt.state["gv_stats"])
# Add score_dict to score_list
score_list += [score_dict]
# Report and save
print(pd.DataFrame(score_list).tail())
ut.save_pkl(score_list_path, score_list)
ut.torch_save(model_path, model.state_dict())
ut.torch_save(opt_path, opt.state_dict())
print("Saved: %s" % savedir)
return score_list
def trainval(exp_dict, savedir, datadir, metrics_flag=True):
# TODO: Do we get similar results with different seeds?
# Set seed
np.random.seed(42)
torch.manual_seed(42)
torch.cuda.manual_seed_all(42)
pprint.pprint(exp_dict)
# Load Train Dataset
train_set = datasets.get_dataset(dataset_name=exp_dict["dataset"],
train_flag=True,
datadir=datadir,
exp_dict=exp_dict)
train_loader = DataLoader(train_set,
drop_last=True,
shuffle=True,
batch_size=exp_dict["batch_size"])
# Load Val Dataset
val_set = datasets.get_dataset(dataset_name=exp_dict["dataset"],
train_flag=False,
datadir=datadir,
exp_dict=exp_dict)
# Load model
model = models.get_model(exp_dict["model"], train_set=train_set).cuda()
# Choose loss and metric function
loss_function = metrics.get_metric_function(exp_dict["loss_func"])
# Load Optimizer
n_batches_per_epoch = len(train_set) / float(exp_dict["batch_size"])
opt = optimizers.get_optimizer(opt=exp_dict["opt"],
params=model.parameters(),
n_batches_per_epoch=n_batches_per_epoch)
# Resume from last saved state_dict
if (not os.path.exists(savedir + "/run_dict.pkl")
or not os.path.exists(savedir + "/score_list.pkl")):
ut.save_pkl(savedir + "/run_dict.pkl", {"running": 1})
score_list = []
s_epoch = 0
else:
score_list = ut.load_pkl(savedir + "/score_list.pkl")
model.load_state_dict(torch.load(savedir + "/model_state_dict.pth"))
opt.load_state_dict(torch.load(savedir + "/opt_state_dict.pth"))
s_epoch = score_list[-1]["epoch"] + 1
for epoch in range(s_epoch, exp_dict["max_epoch"]):
# Set seed
np.random.seed(epoch)
torch.manual_seed(epoch)
torch.cuda.manual_seed_all(epoch)
score_dict = {"epoch": epoch}
if metrics_flag:
# 1. Compute train loss over train set
score_dict["train_loss"] = metrics.compute_metric_on_dataset(
model, train_set, metric_name=exp_dict["loss_func"])
# 2. Compute val acc over val set
score_dict["val_acc"] = metrics.compute_metric_on_dataset(
model, val_set, metric_name=exp_dict["acc_func"])
# 3. Train over train loader
model.train()
print("%d - Training model with %s..." %
(epoch, exp_dict["loss_func"]))
s_time = time.time()
for images, labels in tqdm.tqdm(train_loader):
images, labels = images.cuda(), labels.cuda()
opt.zero_grad()
if exp_dict["opt"]["name"] in exp_configs.ours_opt_list + ["l4"]:
closure = lambda: loss_function(
model, images, labels, backwards=False)
opt.step(closure)
else:
loss = loss_function(model, images, labels)
loss.backward()
opt.step()
e_time = time.time()
# Record step size and batch size
score_dict["step_size"] = opt.state["step_size"]
score_dict["n_forwards"] = opt.state["n_forwards"]
score_dict["n_backwards"] = opt.state["n_backwards"]
score_dict["batch_size"] = train_loader.batch_size
score_dict["train_epoch_time"] = e_time - s_time
# Add score_dict to score_list
score_list += [score_dict]
# Report and save
print(pd.DataFrame(score_list).tail())
ut.save_pkl(savedir + "/score_list.pkl", score_list)
ut.torch_save(savedir + "/model_state_dict.pth", model.state_dict())
ut.torch_save(savedir + "/opt_state_dict.pth", opt.state_dict())
print("Saved: %s" % savedir)
return score_list