def get_xp(args, optimizer):
# various useful information to store
args.command_line = 'python ' + ' '.join(sys.argv)
args.pid = os.getpid()
args.cwd = os.getcwd()
args.hostname = socket.gethostname()
xp = logger.Experiment(args.xp_name,
use_visdom=args.visdom,
visdom_opts={'server': 'http://localhost',
'port': args.port},
time_indexing=False, xlabel='Epoch')
xp.SumMetric(name='epoch', to_plot=False)
xp.AvgMetric(name='error', tag='train')
xp.AvgMetric(name='error', tag='val')
xp.AvgMetric(name='error', tag='test')
xp.TimeMetric(name='timer', tag='train')
xp.TimeMetric(name='timer', tag='val')
xp.TimeMetric(name='timer', tag='test')
xp.AvgMetric(name='obj', tag='train')
xp.log_config(vars(args))
return xp
def get_xp(args, model, optimizer):
# various useful information to store
args.command_line = 'python ' + ' '.join(sys.argv)
args.pid = os.getpid()
args.cwd = os.getcwd()
args.hostname = socket.gethostname()
xp = logger.Experiment(args.xp_name,
use_visdom=args.visdom,
visdom_opts={
'server': args.server,
'port': args.port
},
time_indexing=False,
xlabel='Epoch')
xp.SumMetric(name='epoch', to_plot=False)
xp.AvgMetric(name='acc', tag='train')
xp.AvgMetric(name='acc', tag='val')
xp.AvgMetric(name='acc', tag='test')
xp.BestMetric(name='acc', tag='valbest')
xp.TimeMetric(name='timer', tag='train')
xp.TimeMetric(name='timer', tag='val')
xp.TimeMetric(name='timer', tag='test')
xp.AvgMetric(name='loss', tag='train')
xp.AvgMetric(name='obj', tag='train')
xp.AvgMetric(name='reg')
xp.log_config(vars(args))
xp.AvgMetric(name="gamma")
xp.SimpleMetric(name='eta')
if args.log:
# log at each epoch
xp.Epoch.add_hook(
lambda: xp.to_json('{}/results.json'.format(xp.name_and_dir)))
# log after final evaluation on test set
xp.Acc_Test.add_hook(
lambda: xp.to_json('{}/results.json'.format(xp.name_and_dir)))
# save with different names at each epoch if needed
if args.dump_epoch:
filename = lambda: '{}-{}/model.pkl'.format(
xp.name_and_dir, int(xp.Epoch.value))
else:
filename = lambda: '{}/model.pkl'.format(xp.name_and_dir)
xp.Epoch.add_hook(lambda: save_state(model, optimizer, filename()))
# save results and model for best validation performance
xp.Acc_Valbest.add_hook(
lambda: xp.to_json('{}/best_results.json'.format(xp.name_and_dir)))
xp.Acc_Valbest.add_hook(lambda: save_state(
model, optimizer, '{}/best_model.pkl'.format(xp.name_and_dir)))
return xp
def logger_from_dict(config, use_visdom=True):
"""
Create a logger from a dictionary
:param config: Dictionary containing the experiment hyper-parameters
:param use_visdom: Wether to use Visdom
:return: Logger
"""
experiment_name = datetime.datetime.now().strftime("%Y-%m-%d-")
ordered_config = collections.OrderedDict(sorted(config.items()))
for key, value in ordered_config.items():
key = helpers.to_camel_case(key.replace(' ', '_'))
if value is str:
value = helpers.to_camel_case(value.replace(' ', '_'))
if key == 'Transforms' or key == 'trainType':
continue
if value is not None and key is not None:
experiment_name += "{}_{}-".format(key, value)
# Delete last dash
experiment_name = experiment_name[:-1]
# Create logger
log = logger.Experiment(name=experiment_name,
use_visdom=use_visdom,
visdom_opts={
'server': 'http://localhost',
'port': 8097
},
time_indexing=False,
xlabel='Epoch')
log.log_config(config)
# create parent metric for training metrics (easier interface)
log.ParentWrapper(
tag='train',
name='parent',
children=[log.AvgMetric(name='loss'),
log.AvgMetric(name='acc')])
# same for validation metrics (note all children inherit tag from parent)
log.ParentWrapper(
tag='val',
name='parent',
children=[log.AvgMetric(name='loss'),
log.AvgMetric(name='acc')])
# Add a best metric for the validation accuracy
log.ParentWrapper(tag='best',
name='parent',
children=[log.BestMetric(name='acc')])
return log
def create_experiment(args):
xp = logger.Experiment(args.out_name,
use_visdom=args.visdom,
visdom_opts=dict(server=args.server,
port=args.port),
time_indexing=False,
xlabel='Epoch')
xp.ParentWrapper(tag='train',
name='parent',
children=(xp.AvgMetric(name='loss'),
xp.AvgMetric(name='acc1'),
xp.AvgMetric(name='acck'),
xp.SimpleMetric(name='obj'),
xp.TimeMetric(name='timer')))
xp.ParentWrapper(tag='val',
name='parent',
children=(xp.AvgMetric(name='acck'),
xp.AvgMetric(name='acc1'),
xp.TimeMetric(name='timer')))
xp.ParentWrapper(tag='test',
name='parent',
children=(xp.AvgMetric(name='acc1'),
xp.AvgMetric(name='acck'),
xp.TimeMetric(name='timer')))
xp.SumMetric(name='epoch', to_plot=False)
xp.DynamicMetric(name='learning_rate')
xp.DynamicMetric(name='temperature', to_plot=False)
xp.DynamicMetric(name='mu', to_plot=False)
xp.BestMetric(tag='val_best', name='acc1', mode='max')
xp.BestMetric(tag='val_best', name='acck', mode='max')
if args.visdom:
xp.plotter.set_win_opts(name='acc1', opts={'title': 'Accuracy@1'})
xp.plotter.set_win_opts(name='acck', opts={'title': 'Accuracy@k'})
xp.plotter.set_win_opts(name='loss', opts={'title': 'Loss Function'})
xp.plotter.set_win_opts(name='obj',
opts={'title': 'Objective Function'})
xp.plotter.set_win_opts(name='learning_rate',
opts={'title': 'Learning Rate'})
xp.plotter.set_win_opts(name='Timer',
opts={'title': 'Time (s) / epoch'})
xp.log_config(vars(args))
return xp
def init_visdom(env_name, config):
assert type(config) == dict
visdom_opts = {"server": 'http://localhost', "port": 8787}
stats = logger.Experiment(env_name,
log_git_hash=False,
use_visdom=True,
visdom_opts=visdom_opts,
time_indexing=False)
val_metrics = stats.ParentWrapper(
tag="validation",
name="parent",
children=(stats.SimpleMetric(name='loss'),
stats.SimpleMetric(name="accuracy")))
train_metrics = stats.ParentWrapper(
tag="training",
name="parent",
children=(stats.AvgMetric(name='loss'),
stats.AvgMetric(name="accuracy")))
stats.log_config(config)
def update_metrics(loss, acc, key='train'):
if key == 'train':
train_metrics.update(loss=loss, accuracy=acc)
elif key == "val":
val_metrics.update(loss=loss, accuracy=acc)
def log_metrics():
stats.log_metric(train_metrics)
stats.log_metric(val_metrics)
train_metrics.reset()
val_metrics.reset()
#
norm = stats.ParentWrapper(tag="norm",
name="parent2",
children=(stats.SimpleMetric(name="norm"), ))
def plot_norm(val):
norm.update(norm=val)
stats.log_metric(norm)
norm.reset()
return update_metrics, log_metrics, plot_norm
def set_defaults(args):
# force no logging in debug mode
if args.debug:
args.visdom = False
args.log = False
args.out_name = '../xp/debug'
# remove previous log in debug mode
if os.path.exists('../xp/debug_log.txt'):
os.remove('../xp/debug_log.txt')
elif args.eval:
args.visdom = False
args.log = False
args.out_name = '../xp/{}'.format(args.dataset)
args.epochs = 0
# find settings of experiment
_xp = logger.Experiment("")
_xp.from_json(args.load_model.replace(".pkl", ".json"))
for k in ('topk', 'model'):
setattr(args, k, _xp.config[k])
if args.multiple_crops:
args.test_batch_size = 32
assert args.dataset in ('imagenet', 'cifar100')
if args.dataset == 'imagenet':
my_dict = imagenet_defaults
if args.dataset == 'cifar100':
my_dict = cifar100_defaults
# set number of classes
args.num_classes = my_dict['num_classes']
# replace None values with default ones
for (k, v) in my_dict.items():
if getattr(args, k) is None:
setattr(args, k, v)
# store full command line in args
args.command_line = ' '.join(sys.argv)
# store cuurent directory and hostname in args
args.cwd = os.getcwd()
args.hostname = socket.gethostname()
def create_xp(args):
logger.set_default_indexing('increment')
xp_name = "{}_{}_{}".format(args.opt, args.lr, args.weight_decay)
if args.proj is not None:
xp_name += "_proj{}".format(args.proj)
setproctitle.setproctitle(xp_name)
plotter = logger.Plotter(visdom_opts={
'server': 'http://atlas.robots.ox.ac.uk',
'port': 9006,
'env': xp_name
},
mode='automatic')
xp = logger.Experiment(name=xp_name, plotter=plotter, track_git=True)
# log the hyperparameters of the experiment
xp.config.update(**vars(args))
xp.config.record()
xp.gamma = logger.SimpleMetric()
xp.gamma_unclipped = logger.SimpleMetric()
xp.epsilon = logger.SimpleMetric()
return xp
acck = np.random.rand() + 90
return loss, acc1, acck
# some hyper-parameters of the experiment
lr = 0.01
n_epochs = 10
#----------------------------------------------------------
# Prepare logging
#----------------------------------------------------------
# create Experiment
xp = logger.Experiment("xp_name", use_visdom=True,
visdom_opts={'server': 'http://localhost', 'port': 8097},
time_indexing=False, xlabel='Epoch')
# log the hyperparameters of the experiment
xp.log_config({'lr': lr, 'n_epochs': n_epochs})
# create parent metric for training metrics (easier interface)
xp.ParentWrapper(tag='train', name='parent',
children=(xp.AvgMetric(name='loss'),
xp.AvgMetric(name='acc1'),
xp.AvgMetric(name='acck')))
# same for validation metrics (note all children inherit tag from parent)
xp.ParentWrapper(tag='val', name='parent',
children=(xp.AvgMetric(name='loss'),
xp.AvgMetric(name='acc1'),
xp.AvgMetric(name='acck')))
best1 = xp.BestMetric(tag="val-best", name="acc1")
bestk = xp.BestMetric(tag="val-best", name="acck")
np.random.seed(opt.seed)
torch.manual_seed(opt.seed)
if opt.cuda:
cudnn.benchmark = True
torch.cuda.manual_seed_all(opt.seed)
# create logger
LOG_DIR = '{0}/logger'.format(opt.experiment)
# some hyperparameters we wish to save for this experiment
hyperparameters = dict(regularization=1, n_epochs=opt.epochs)
# options for the remote visualization backend
visdom_opts = dict(server='http://localhost', port=8097)
# create logger for visdom
xp = logger.Experiment('xp_name', use_visdom=True, visdom_opts=visdom_opts)
# log the hyperparameters of the experiment
xp.log_config(hyperparameters)
# create parent metric for training metrics (easier interface)
train_metrics = xp.ParentWrapper(tag='train',
name='parent',
children=(xp.AvgMetric(name='lossD'),
xp.AvgMetric(name='lossG')))
if torch.cuda.is_available() and not opt.cuda:
print(
"WARNING: You have a CUDA device, so you should probably run with --cuda"
)
# setup transformations
transforms = transforms.Compose([
if len(real_paths[0]) > 0:
sg = model.graph.subgraph(real_paths[0])
drawer.draw(sg, vis_opts={'title': 'Real', **draw_ops}, weights=1.0, vis_win='Clean')
for ce_name, ce in cost_evaluators.items():
if ce_name.startswith('parallel'):
n_steps = max(ce.node_steps[0].values())
drawer.draw(sg, vis_opts={'title': 'Alloc ' + ce_name, **draw_ops},
weights=ce.node_steps[0],
colormap=drawing_utils.get_colormap(n_steps),
vis_win='Alloc ' + ce_name)
drawer.draw(model.graph, vis_opts={'title': 'Mean', **draw_ops},
weights=path_recorder.get_posterior_weights(), vis_win='mean')
logger.info('EPOCH DONE')
if __name__ == '__main__':
logger.info('Executing main from {}'.format(os.getcwd()))
exp_name = 'Direct Launch'
args = vars(argument_parser())
vis_conf = external.get_visdom_conf()
xp_logger = data_logger.Experiment(exp_name, use_visdom=True,
visdom_opts={'server': vis_conf['url'], 'port': vis_conf['port'],
'env': args['draw_env']},
time_indexing=False, xlabel='Epoch')
main(args, xp_logger)
import logger
import os
for root, dirs, files in os.walk("./runs"):
for file in files:
if file.endswith('.json') and 'archive' not in root:
file_path = os.path.join(root, file)
print('Loading: ', file_path)
xp = logger.Experiment('dummy_name')
xp.from_json(file_path)
for opts in xp.visdom_win_opts.values():
if 'legend' in opts:
opts.pop('legend')
xp.to_visdom(visdom_opts={'server': 'http://localhost', 'port': 8098})
print('Success!')
def main(_run, nepochs, device, use_visdom, visdom_conf, n_classes,
lambda_reward, r_beta, r_gamma, _config):
exp_name = format_exp_name(_run._id, _config)
if use_visdom:
visdom_conf.update(env=exp_name)
_run.info['visdom_server'] = "{server}:{port}/env/{env}".format(
**visdom_conf)
else:
_run.info['visdom_server'] = "No visdom"
_run.info['exp_name'] = exp_name
front = _run.info['front'] = {}
xp_logger = data_logger.Experiment(exp_name,
use_visdom=use_visdom,
visdom_opts=visdom_conf,
time_indexing=False,
xlabel='Epoch',
log_git_hash=False)
xp_logger.add_log_hook(_run.log_scalar)
if use_visdom:
xp_logger.plotter.windows_opts = defaultdict(
lambda: dict(showlegend=True))
viz = Visdom(**visdom_conf) if use_visdom else None
# Dataset creation
logger.info('### Dataset ###')
ds, batch_first, class_w = create_dataset()
_run.info['class_weights'] = class_w.tolist()
confusion_matrix_opts = {
'columnnames': ds['train'].dataset.ordered_class_names,
'rownames': ds['train'].dataset.ordered_class_names
}
# Model Creation
logger.info('### Model ###')
adaptive_model = create_model()
adaptive_model.loss = torch.nn.CrossEntropyLoss(weight=class_w,
reduction='none',
ignore_index=-7)
path_recorder = PathRecorder(adaptive_model.stochastic_model)
cost_evaluator = ComputationCostEvaluator(
node_index=path_recorder.node_index, bw=False)
# cost_evaluator = SimpleEdgeCostEvaluator(node_index=path_recorder.node_index, bw=False)
cost_evaluator.init_costs(adaptive_model.stochastic_model)
logger.info('Cost: {:.5E}'.format(cost_evaluator.total_cost))
adaptive_model.to(device)
# Optim Creation
logger.info('### Optim ###')
optimizer, schedulder = create_optim(
params=adaptive_model.get_param_groups())
# Check the param_groups order, to be sure to get the learning rates in the right order for logging
assert [pg['name'] for pg in optimizer.param_groups
] == ['arch_params', 'pred_params']
def optim_closure(loss):
optimizer.zero_grad()
loss.backward()
optimizer.step()
# Logger creation
splits = ['train', 'validation', 'test']
metrics = [
'classif_loss', 'arch_loss', 'reward', 'lambda_reward',
'silence_ratio', 'accuracy', 'average_cost', 'learning_rate_pred',
'learning_rate_arch'
]
for split in splits:
xp_logger.ParentWrapper(tag=split,
name='parent'.format(split),
children=[
xp_logger.SimpleMetric(name=metric)
for metric in metrics
])
train_cost_loggers = dict(
(i, xp_logger.AvgMetric(name='train_cost', tag=name))
for i, name in enumerate(ds['train'].dataset.ordered_class_names))
train_cost_loggers_perc = dict(
(i, xp_logger.AvgMetric(name='train_cost_perceived', tag=name))
for i, name in enumerate(ds['train'].dataset.ordered_class_names))
node_names = adaptive_model.stochastic_model.ordered_node_names
# entropy_loggers = [xp_logger.SimpleMetric(name='entropy', tag=name) for name in node_names]
entropy_loggers = OrderedDict(
(i, xp_logger.SimpleMetric(name='entropy_per_node', tag=name))
for i, name in enumerate(node_names))
# proba_loggers = [xp_logger.SimpleMetric(name='proba', tag=name) for name in node_names]
proba_loggers = OrderedDict(
(i, xp_logger.SimpleMetric(name='proba_per_node', tag=name))
for i, name in enumerate(node_names))
val_cost_loggers = dict(
(i, xp_logger.AvgMetric(name='val_cost', tag=name))
for i, name in enumerate(ds['validation'].dataset.ordered_class_names))
val_cost_loggers_perc = dict(
(i, xp_logger.AvgMetric(name='val_cost_perceived', tag=name))
for i, name in enumerate(ds['validation'].dataset.ordered_class_names))
test_cost_loggers = dict(
(i, xp_logger.AvgMetric(name='test_cost', tag=name))
for i, name in enumerate(ds['test'].dataset.ordered_class_names))
test_cost_loggers_perc = dict(
(i, xp_logger.AvgMetric(name='test_cost_perceived', tag=name))
for i, name in enumerate(ds['test'].dataset.ordered_class_names))
if use_visdom:
print_properties(viz, _config)
print_properties(viz, _run.info)
ema_reward = EMA(r_beta) # Init the exponential moving average
for n in range(1, nepochs + 1):
logger.info('### Sarting epoch n°{} ### {}'.format(
n, _run.info['visdom_server']))
logger.info(' '.join(sys.argv))
if schedulder:
schedulder.step(n)
arch_lr, pred_lr = schedulder.get_lr()
xp_logger.Parent_Train.update(learning_rate_pred=pred_lr,
learning_rate_arch=arch_lr)
# Training
adaptive_model.train()
train_cm, train_costcm, train_costcm_norm, train_cost_per_step, logs, train_cost_per_signal_level, train_stats = evaluate_model(
adaptive_model,
ds['train'],
batch_first,
device,
path_recorder,
cost_evaluator,
train_cost_loggers,
train_cost_loggers_perc,
n_classes,
lambda_reward,
ema_reward,
r_gamma,
optim_closure,
name='Train')
xp_logger.Parent_Train.update(**dict(
(k, v.value()[0]) for k, v in logs.items()))
for node_idx, ent in train_stats['en'].items():
entropy_loggers[node_idx].update(ent.value()[0])
for node_idx, prob in train_stats['pn'].items():
proba_loggers[node_idx].update(prob.value()[0])
# Evaluation
adaptive_model.eval()
val_cm, val_costcm, val_costcm_norm, val_cost_per_step, logs, val_cost_per_signal_level, val_stats = evaluate_model(
adaptive_model,
ds['validation'],
batch_first,
device,
path_recorder,
cost_evaluator,
val_cost_loggers,
val_cost_loggers_perc,
n_classes,
lambda_reward,
ema_reward,
r_gamma,
name='Validation')
xp_logger.Parent_Validation.update(**dict(
(k, v.value()[0]) for k, v in logs.items()))
test_cm, test_costcm, test_costcm_norm, test_cost_per_step, logs, test_cost_per_signal_level, test_stats = evaluate_model(
adaptive_model,
ds['test'],
batch_first,
device,
path_recorder,
cost_evaluator,
test_cost_loggers,
test_cost_loggers_perc,
n_classes,
lambda_reward,
ema_reward,
r_gamma,
name='Test')
xp_logger.Parent_Test.update(**dict(
(k, v.value()[0]) for k, v in logs.items()))
if use_visdom:
# Log
plot_(viz,
train_stats['es'],
node_names,
f'Entropy per step {n} - Train',
'train_eps',
log_func=_run.log_scalar)
plot_(viz,
train_stats['ps'],
node_names,
f'Probability per step {n} - Train',
'train_pps',
log_func=_run.log_scalar)
try:
viz.heatmap(train_cm,
win='train_cm',
opts={
**confusion_matrix_opts, 'title':
'Train Confusion matrix'
})
viz.heatmap(val_cm,
win='val_cm',
opts={
**confusion_matrix_opts, 'title':
'Val Confusion matrix'
})
viz.heatmap(test_cm,
win='test_cm',
opts={
**confusion_matrix_opts, 'title':
'Test Confusion matrix'
})
# viz.heatmap(train_costcm, win='train_cost_matrix',
# opts={**confusion_matrix_opts, 'title': 'Train cost matrix'})
# viz.heatmap(val_costcm, win='val_cost_matrix', opts={**confusion_matrix_opts, 'title': 'Val cost matrix'})
# viz.heatmap(test_costcm, win='test_cost_matrix',
# opts={**confusion_matrix_opts, 'title': 'Test cost matrix'})
viz.heatmap(train_costcm_norm,
win='train_cost_matrix_norm',
opts={
**confusion_matrix_opts, 'title':
'Train cost matrix Normalized'
})
viz.heatmap(val_costcm_norm,
win='val_cost_matrix_norm',
opts={
**confusion_matrix_opts, 'title':
'Val cost matrix Normalized'
})
viz.heatmap(test_costcm_norm,
win='test_cost_matrix_norm',
opts={
**confusion_matrix_opts, 'title':
'Test cost matrix Normalized'
})
except ConnectionError as err:
logger.warning('Error in heatmaps:')
logger.warning(err)
traceback.print_exc()
plot_meters(viz,
train_cost_per_step,
'train_cps',
'Cost per step {}'.format(n),
win='cps',
log_func=_run.log_scalar)
plot_meters(viz,
val_cost_per_step,
'val_cps',
win='cps',
log_func=_run.log_scalar)
plot_meters(viz,
test_cost_per_step,
'test_cps',
win='cps',
log_func=_run.log_scalar)
plot_meters(viz,
train_cost_per_signal_level,
'cost/sig_train',
'Cost per signal {}'.format(n),
win='cpsig',
error_bars=False,
log_func=_run.log_scalar)
plot_meters(viz,
val_cost_per_signal_level,
'cost/sig_val',
win='cpsig',
error_bars=False,
log_func=_run.log_scalar)
plot_meters(viz,
test_cost_per_signal_level,
'cost/sig_test',
win='cpsig',
error_bars=False,
log_func=_run.log_scalar)
xp_logger.log_with_tag(tag='*', reset=True)
msg = 'Losses: {:.3f}({:.3E})-{:.3f}-{:.3f}, Accuracies: {:.3f}-{:.3f}-{:.3f}, Avg cost: {:.3E}-{:.3E}-{:.3E}'
msg = msg.format(xp_logger.classif_loss_train, xp_logger.reward_train,
xp_logger.classif_loss_validation,
xp_logger.classif_loss_test, xp_logger.accuracy_train,
xp_logger.accuracy_validation,
xp_logger.accuracy_test, xp_logger.average_cost_train,
xp_logger.average_cost_validation,
xp_logger.average_cost_test)
logger.info(msg)
pareto_data = {
'cost': xp_logger.logged['average_cost_validation'].values(),
'acc': xp_logger.logged['accuracy_validation'].values(),
'_orig_': xp_logger.logged['average_cost_validation'].keys()
}
pareto = paretize_exp(pareto_data, x_name='cost', crit_name='acc')
if n in pareto['_orig_']:
logger.info('New on front !')
front.update(**pareto)
save_checkpoint(adaptive_model, ex, n)
elif n > 0 and n % 50 == 0:
logger.info('Checkpointing')
save_checkpoint(adaptive_model, ex, n)
logger.info(pareto['_orig_'])
best_epoch = pareto['_orig_'][-1]
logger.info('Best \tVal: {:.3f} - Test: {:.3f} (Epoch {})\n'.format(
xp_logger.logged['accuracy_validation'][best_epoch],
xp_logger.logged['accuracy_test'][best_epoch], best_epoch))
def train(cfg, writer, logger_old, args):
# Setup seeds
torch.manual_seed(cfg.get('seed', 1337))
torch.cuda.manual_seed(cfg.get('seed', 1337))
np.random.seed(cfg.get('seed', 1337))
random.seed(cfg.get('seed', 1337))
# Setup device
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
# Setup Augmentations
augmentations = cfg['training'].get('augmentations', None)
data_aug = get_composed_augmentations(augmentations)
# Setup Dataloader
data_loader = get_loader(cfg['data']['dataset'])
data_path = cfg['data']['path']
if isinstance(cfg['training']['loss']['superpixels'], int):
use_superpixels = True
cfg['data']['train_split'] = 'train_super'
cfg['data']['val_split'] = 'val_super'
setup_superpixels(cfg['training']['loss']['superpixels'])
elif cfg['training']['loss']['superpixels'] is not None:
raise Exception(
"cfg['training']['loss']['superpixels'] is of the wrong type")
else:
use_superpixels = False
t_loader = data_loader(data_path,
is_transform=True,
split=cfg['data']['train_split'],
superpixels=cfg['training']['loss']['superpixels'],
img_size=(cfg['data']['img_rows'],
cfg['data']['img_cols']),
augmentations=data_aug)
v_loader = data_loader(
data_path,
is_transform=True,
split=cfg['data']['val_split'],
superpixels=cfg['training']['loss']['superpixels'],
img_size=(cfg['data']['img_rows'], cfg['data']['img_cols']),
)
n_classes = t_loader.n_classes
trainloader = data.DataLoader(t_loader,
batch_size=cfg['training']['batch_size'],
num_workers=cfg['training']['n_workers'],
shuffle=True)
valloader = data.DataLoader(v_loader,
batch_size=cfg['training']['batch_size'],
num_workers=cfg['training']['n_workers'])
# Setup Metrics
running_metrics_val = runningScore(n_classes)
running_metrics_train = runningScore(n_classes)
# Setup Model
model = get_model(cfg['model'], n_classes).to(device)
model = torch.nn.DataParallel(model,
device_ids=range(torch.cuda.device_count()))
# Setup optimizer, lr_scheduler and loss function
optimizer_cls = get_optimizer(cfg)
optimizer_params = {
k: v
for k, v in cfg['training']['optimizer'].items() if k != 'name'
}
optimizer = optimizer_cls(model.parameters(), **optimizer_params)
logger_old.info("Using optimizer {}".format(optimizer))
scheduler = get_scheduler(optimizer, cfg['training']['lr_schedule'])
loss_fn = get_loss_function(cfg)
logger_old.info("Using loss {}".format(loss_fn))
start_iter = 0
if cfg['training']['resume'] is not None:
if os.path.isfile(cfg['training']['resume']):
logger_old.info(
"Loading model and optimizer from checkpoint '{}'".format(
cfg['training']['resume']))
checkpoint = torch.load(cfg['training']['resume'])
model.load_state_dict(checkpoint["model_state"])
optimizer.load_state_dict(checkpoint["optimizer_state"])
scheduler.load_state_dict(checkpoint["scheduler_state"])
start_iter = checkpoint["epoch"]
logger_old.info("Loaded checkpoint '{}' (iter {})".format(
cfg['training']['resume'], checkpoint["epoch"]))
else:
logger_old.info("No checkpoint found at '{}'".format(
cfg['training']['resume']))
val_loss_meter = averageMeter()
train_loss_meter = averageMeter()
time_meter = averageMeter()
train_len = t_loader.train_len
val_static = 0
best_iou = -100.0
i = start_iter
j = 0
flag = True
# Prepare logging
xp_name = cfg['model']['arch'] + '_' + \
cfg['training']['loss']['name'] + '_' + args.name
xp = logger.Experiment(xp_name,
use_visdom=True,
visdom_opts={
'server': 'http://localhost',
'port': 8098
},
time_indexing=False,
xlabel='Epoch')
# log the hyperparameters of the experiment
xp.log_config(flatten(cfg))
# create parent metric for training metrics (easier interface)
xp.ParentWrapper(tag='train',
name='parent',
children=(xp.AvgMetric(name="loss"),
xp.AvgMetric(name='acc'),
xp.AvgMetric(name='acccls'),
xp.AvgMetric(name='fwavacc'),
xp.AvgMetric(name='meaniu')))
xp.ParentWrapper(tag='val',
name='parent',
children=(xp.AvgMetric(name="loss"),
xp.AvgMetric(name='acc'),
xp.AvgMetric(name='acccls'),
xp.AvgMetric(name='fwavacc'),
xp.AvgMetric(name='meaniu')))
best_loss = xp.BestMetric(tag='val-best', name='loss', mode='min')
best_acc = xp.BestMetric(tag='val-best', name='acc')
best_acccls = xp.BestMetric(tag='val-best', name='acccls')
best_fwavacc = xp.BestMetric(tag='val-best', name='fwavacc')
best_meaniu = xp.BestMetric(tag='val-best', name='meaniu')
xp.plotter.set_win_opts(name="loss", opts={'title': 'Loss'})
xp.plotter.set_win_opts(name="acc", opts={'title': 'Micro-Average'})
xp.plotter.set_win_opts(name="acccls", opts={'title': 'Macro-Average'})
xp.plotter.set_win_opts(name="fwavacc", opts={'title': 'FreqW Accuracy'})
xp.plotter.set_win_opts(name="meaniu", opts={'title': 'Mean IoU'})
it_per_step = cfg['training']['acc_batch_size']
eff_batch_size = cfg['training']['batch_size'] * it_per_step
while i <= train_len * (cfg['training']['epochs']) and flag:
for (images, labels, labels_s, masks) in trainloader:
i += 1
j += 1
start_ts = time.time()
scheduler.step()
model.train()
images = images.to(device)
labels = labels.to(device)
labels_s = labels_s.to(device)
masks = masks.to(device)
outputs = model(images)
if use_superpixels:
outputs_s, labels_s, sizes = convert_to_superpixels(
outputs, labels_s, masks)
loss = loss_fn(input=outputs_s, target=labels_s, size=sizes)
outputs = convert_to_pixels(outputs_s, outputs, masks)
else:
loss = loss_fn(input=outputs, target=labels)
# accumulate train metrics during train
pred = outputs.data.max(1)[1].cpu().numpy()
gt = labels.data.cpu().numpy()
running_metrics_train.update(gt, pred)
train_loss_meter.update(loss.item())
if args.evaluate:
decoded = t_loader.decode_segmap(np.squeeze(pred, axis=0))
misc.imsave("./{}.png".format(i), decoded)
image_save = np.transpose(
np.squeeze(images.data.cpu().numpy(), axis=0), (1, 2, 0))
misc.imsave("./{}.jpg".format(i), image_save)
# accumulate gradients based on the accumulation batch size
if i % it_per_step == 1 or it_per_step == 1:
optimizer.zero_grad()
grad_rescaling = torch.tensor(1. / it_per_step).type_as(loss)
loss.backward(grad_rescaling)
if (i + 1) % it_per_step == 1 or it_per_step == 1:
optimizer.step()
optimizer.zero_grad()
time_meter.update(time.time() - start_ts)
# training logs
if (j + 1) % (cfg['training']['print_interval'] *
it_per_step) == 0:
fmt_str = "Epoch [{}/{}] Iter [{}/{:d}] Loss: {:.4f} Time/Image: {:.4f}"
total_iter = int(train_len / eff_batch_size)
total_epoch = int(cfg['training']['epochs'])
current_epoch = ceil((i + 1) / train_len)
current_iter = int((j + 1) / it_per_step)
print_str = fmt_str.format(
current_epoch, total_epoch, current_iter, total_iter,
loss.item(),
time_meter.avg / cfg['training']['batch_size'])
print(print_str)
logger_old.info(print_str)
writer.add_scalar('loss/train_loss', loss.item(), i + 1)
time_meter.reset()
# end of epoch evaluation
if (i + 1) % train_len == 0 or \
(i + 1) == train_len * (cfg['training']['epochs']):
optimizer.step()
optimizer.zero_grad()
model.eval()
with torch.no_grad():
for i_val, (images_val, labels_val, labels_val_s,
masks_val) in tqdm(enumerate(valloader)):
images_val = images_val.to(device)
labels_val = labels_val.to(device)
labels_val_s = labels_val_s.to(device)
masks_val = masks_val.to(device)
outputs = model(images_val)
if use_superpixels:
outputs_s, labels_val_s, sizes_val = convert_to_superpixels(
outputs, labels_val_s, masks_val)
val_loss = loss_fn(input=outputs_s,
target=labels_val_s,
size=sizes_val)
outputs = convert_to_pixels(
outputs_s, outputs, masks_val)
else:
val_loss = loss_fn(input=outputs,
target=labels_val)
pred = outputs.data.max(1)[1].cpu().numpy()
gt = labels_val.data.cpu().numpy()
running_metrics_val.update(gt, pred)
val_loss_meter.update(val_loss.item())
writer.add_scalar('loss/val_loss', val_loss_meter.avg, i + 1)
writer.add_scalar('loss/train_loss', train_loss_meter.avg,
i + 1)
logger_old.info("Epoch %d Val Loss: %.4f" % (int(
(i + 1) / train_len), val_loss_meter.avg))
logger_old.info("Epoch %d Train Loss: %.4f" % (int(
(i + 1) / train_len), train_loss_meter.avg))
score, class_iou = running_metrics_train.get_scores()
print("Training metrics:")
for k, v in score.items():
print(k, v)
logger_old.info('{}: {}'.format(k, v))
writer.add_scalar('train_metrics/{}'.format(k), v, i + 1)
for k, v in class_iou.items():
logger_old.info('{}: {}'.format(k, v))
writer.add_scalar('train_metrics/cls_{}'.format(k), v,
i + 1)
xp.Parent_Train.update(loss=train_loss_meter.avg,
acc=score['Overall Acc: \t'],
acccls=score['Mean Acc : \t'],
fwavacc=score['FreqW Acc : \t'],
meaniu=score['Mean IoU : \t'])
score, class_iou = running_metrics_val.get_scores()
print("Validation metrics:")
for k, v in score.items():
print(k, v)
logger_old.info('{}: {}'.format(k, v))
writer.add_scalar('val_metrics/{}'.format(k), v, i + 1)
for k, v in class_iou.items():
logger_old.info('{}: {}'.format(k, v))
writer.add_scalar('val_metrics/cls_{}'.format(k), v, i + 1)
xp.Parent_Val.update(loss=val_loss_meter.avg,
acc=score['Overall Acc: \t'],
acccls=score['Mean Acc : \t'],
fwavacc=score['FreqW Acc : \t'],
meaniu=score['Mean IoU : \t'])
xp.Parent_Val.log_and_reset()
xp.Parent_Train.log_and_reset()
best_loss.update(xp.loss_val).log()
best_acc.update(xp.acc_val).log()
best_acccls.update(xp.acccls_val).log()
best_fwavacc.update(xp.fwavacc_val).log()
best_meaniu.update(xp.meaniu_val).log()
visdir = os.path.join('runs', cfg['training']['loss']['name'],
args.name, 'plots.json')
xp.to_json(visdir)
val_loss_meter.reset()
train_loss_meter.reset()
running_metrics_val.reset()
running_metrics_train.reset()
j = 0
if score["Mean IoU : \t"] >= best_iou:
val_static = 0
best_iou = score["Mean IoU : \t"]
state = {
"epoch": i + 1,
"model_state": model.state_dict(),
"optimizer_state": optimizer.state_dict(),
"scheduler_state": scheduler.state_dict(),
"best_iou": best_iou,
}
save_path = os.path.join(
writer.file_writer.get_logdir(),
"{}_{}_best_model.pkl".format(cfg['model']['arch'],
cfg['data']['dataset']))
torch.save(state, save_path)
else:
val_static += 1
if (i + 1) == train_len * (
cfg['training']['epochs']) or val_static == 10:
flag = False
break
return best_iou