def make_data_loader(spec, tag=''):
if spec is None:
return None
dataset = datasets.make(spec['dataset'])
dataset = datasets.make(spec['wrapper'], args={'dataset': dataset})
log('{} dataset: size={}'.format(tag, len(dataset)))
for k, v in dataset[0].items():
log(' {}: shape={}'.format(k, tuple(v.shape)))
loader = DataLoader(dataset,
batch_size=spec['batch_size'],
shuffle=(tag == 'train'),
num_workers=8,
pin_memory=True)
return loader
def main(config):
random.seed(0)
np.random.seed(0)
torch.manual_seed(0)
torch.cuda.manual_seed(0)
# torch.backends.cudnn.deterministic = True
# torch.backends.cudnn.benchmark = False
##### Dataset #####
dataset = datasets.make(config['dataset'], **config['test'])
utils.log('meta-test set: {} (x{}), {}'.format(dataset[0][0].shape,
len(dataset),
dataset.n_classes))
loader = DataLoader(dataset,
config['test']['n_episode'],
collate_fn=datasets.collate_fn,
num_workers=1,
pin_memory=True)
##### Model #####
ckpt = torch.load(config['load'])
inner_args = utils.config_inner_args(config.get('inner_args'))
model = models.load(ckpt, load_clf=(not inner_args['reset_classifier']))
if args.efficient:
model.go_efficient()
if config.get('_parallel'):
model = nn.DataParallel(model)
utils.log('num params: {}'.format(utils.compute_n_params(model)))
##### Evaluation #####
model.eval()
aves_va = utils.AverageMeter()
va_lst = []
for epoch in range(1, config['epoch'] + 1):
for data in tqdm(loader, leave=False):
x_shot, x_query, y_shot, y_query = data
x_shot, y_shot = x_shot.cuda(), y_shot.cuda()
x_query, y_query = x_query.cuda(), y_query.cuda()
if inner_args['reset_classifier']:
if config.get('_parallel'):
model.module.reset_classifier()
else:
model.reset_classifier()
logits = model(x_shot,
x_query,
y_shot,
inner_args,
meta_train=False)
logits = logits.view(-1, config['test']['n_way'])
labels = y_query.view(-1)
pred = torch.argmax(logits, dim=1)
acc = utils.compute_acc(pred, labels)
aves_va.update(acc, 1)
va_lst.append(acc)
print('test epoch {}: acc={:.2f} +- {:.2f} (%)'.format(
epoch,
aves_va.item() * 100,
utils.mean_confidence_interval(va_lst) * 100))
def main(config):
svname = args.name
if svname is None:
svname = 'classifier_{}'.format(config['train_dataset'])
svname += '_' + config['model_args']['encoder']
clsfr = config['model_args']['classifier']
if clsfr != 'linear-classifier':
svname += '-' + clsfr
if args.tag is not None:
svname += '_' + args.tag
save_path = os.path.join('./save', svname)
utils.ensure_path(save_path)
utils.set_log_path(save_path)
writer = SummaryWriter(os.path.join(save_path, 'tensorboard'))
yaml.dump(config, open(os.path.join(save_path, 'config.yaml'), 'w'))
#### Dataset ####
# train
train_dataset = datasets.make(config['train_dataset'],
**config['train_dataset_args'])
augmentations = [
transforms.Compose([
transforms.RandomHorizontalFlip(p=0.5),
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
]),
transforms.Compose([
transforms.RandomResizedCrop(size=(80, 80),
scale=(0.08, 1.0),
ratio=(0.75, 1.3333)),
transforms.RandomHorizontalFlip(p=0.5),
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
]),
transforms.Compose([
transforms.RandomRotation(35),
transforms.RandomHorizontalFlip(p=0.5),
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
]),
transforms.Compose([
transforms.ColorJitter(0.4, 0.4, 0.4, 0.1),
transforms.RandomHorizontalFlip(p=0.5),
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
]),
transforms.Compose([
transforms.RandomResizedCrop(size=(80, 80),
scale=(0.08, 1.0),
ratio=(0.75, 1.3333)),
transforms.RandomRotation(35),
transforms.RandomHorizontalFlip(p=0.5),
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
]),
transforms.Compose([
transforms.RandomRotation(35),
transforms.ColorJitter(0.4, 0.4, 0.4, 0.1),
transforms.RandomHorizontalFlip(p=0.5),
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
]),
transforms.Compose([
transforms.RandomResizedCrop(size=(80, 80),
scale=(0.08, 1.0),
ratio=(0.75, 1.3333)),
transforms.ColorJitter(0.4, 0.4, 0.4, 0.1),
transforms.RandomHorizontalFlip(p=0.5),
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
]),
transforms.Compose([
transforms.RandomRotation(35),
transforms.RandomResizedCrop(size=(80, 80),
scale=(0.08, 1.0),
ratio=(0.75, 1.3333)),
transforms.ColorJitter(0.4, 0.4, 0.4, 0.1),
transforms.RandomHorizontalFlip(p=0.5),
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
])
]
train_dataset.transform = augmentations[int(config['_a'])]
print(train_dataset.transform)
print("_a", config['_a'])
input("Continue with these augmentations?")
train_loader = DataLoader(train_dataset,
config['batch_size'],
shuffle=True,
num_workers=0,
pin_memory=True)
utils.log('train dataset: {} (x{}), {}'.format(train_dataset[0][0].shape,
len(train_dataset),
train_dataset.n_classes))
if config.get('visualize_datasets'):
utils.visualize_dataset(train_dataset, 'train_dataset', writer)
# val
if config.get('val_dataset'):
eval_val = True
val_dataset = datasets.make(config['val_dataset'],
**config['val_dataset_args'])
val_loader = DataLoader(val_dataset,
config['batch_size'],
num_workers=0,
pin_memory=True)
utils.log('val dataset: {} (x{}), {}'.format(val_dataset[0][0].shape,
len(val_dataset),
val_dataset.n_classes))
if config.get('visualize_datasets'):
utils.visualize_dataset(val_dataset, 'val_dataset', writer)
else:
eval_val = False
# few-shot eval
if config.get('fs_dataset'):
ef_epoch = config.get('eval_fs_epoch')
if ef_epoch is None:
ef_epoch = 5
eval_fs = True
fs_dataset = datasets.make(config['fs_dataset'],
**config['fs_dataset_args'])
utils.log('fs dataset: {} (x{}), {}'.format(fs_dataset[0][0].shape,
len(fs_dataset),
fs_dataset.n_classes))
if config.get('visualize_datasets'):
utils.visualize_dataset(fs_dataset, 'fs_dataset', writer)
n_way = 5
n_query = 15
n_shots = [1, 5]
fs_loaders = []
for n_shot in n_shots:
fs_sampler = CategoriesSampler(fs_dataset.label,
200,
n_way,
n_shot + n_query,
ep_per_batch=4)
fs_loader = DataLoader(fs_dataset,
batch_sampler=fs_sampler,
num_workers=0,
pin_memory=True)
fs_loaders.append(fs_loader)
else:
eval_fs = False
########
#### Model and Optimizer ####
if config.get('load'):
model_sv = torch.load(config['load'])
model = models.load(model_sv)
else:
model = models.make(config['model'], **config['model_args'])
if eval_fs:
fs_model = models.make('meta-baseline', encoder=None)
fs_model.encoder = model.encoder
if config.get('_parallel'):
model = nn.DataParallel(model)
if eval_fs:
fs_model = nn.DataParallel(fs_model)
utils.log('num params: {}'.format(utils.compute_n_params(model)))
optimizer, lr_scheduler = utils.make_optimizer(model.parameters(),
config['optimizer'],
**config['optimizer_args'])
########
max_epoch = config['max_epoch']
save_epoch = config.get('save_epoch')
max_va = 0.
timer_used = utils.Timer()
timer_epoch = utils.Timer()
for epoch in range(1, max_epoch + 1 + 1):
if epoch == max_epoch + 1:
if not config.get('epoch_ex'):
break
train_dataset.transform = train_dataset.default_transform
print(train_dataset.transform)
train_loader = DataLoader(train_dataset,
config['batch_size'],
shuffle=True,
num_workers=0,
pin_memory=True)
timer_epoch.s()
aves_keys = ['tl', 'ta', 'vl', 'va']
if eval_fs:
for n_shot in n_shots:
aves_keys += ['fsa-' + str(n_shot)]
aves = {k: utils.Averager() for k in aves_keys}
# train
model.train()
writer.add_scalar('lr', optimizer.param_groups[0]['lr'], epoch)
for data, label in tqdm(train_loader, desc='train', leave=False):
# for data, label in train_loader:
data, label = data.cuda(), label.cuda()
logits = model(data)
loss = F.cross_entropy(logits, label)
acc = utils.compute_acc(logits, label)
optimizer.zero_grad()
loss.backward()
optimizer.step()
aves['tl'].add(loss.item())
aves['ta'].add(acc)
logits = None
loss = None
# eval
if eval_val:
model.eval()
for data, label in tqdm(val_loader, desc='val', leave=False):
data, label = data.cuda(), label.cuda()
with torch.no_grad():
logits = model(data)
loss = F.cross_entropy(logits, label)
acc = utils.compute_acc(logits, label)
aves['vl'].add(loss.item())
aves['va'].add(acc)
if eval_fs and (epoch % ef_epoch == 0 or epoch == max_epoch + 1):
fs_model.eval()
for i, n_shot in enumerate(n_shots):
np.random.seed(0)
for data, _ in tqdm(fs_loaders[i],
desc='fs-' + str(n_shot),
leave=False):
x_shot, x_query = fs.split_shot_query(data.cuda(),
n_way,
n_shot,
n_query,
ep_per_batch=4)
label = fs.make_nk_label(n_way, n_query,
ep_per_batch=4).cuda()
with torch.no_grad():
logits = fs_model(x_shot, x_query).view(-1, n_way)
acc = utils.compute_acc(logits, label)
aves['fsa-' + str(n_shot)].add(acc)
# post
if lr_scheduler is not None:
lr_scheduler.step()
for k, v in aves.items():
aves[k] = v.item()
t_epoch = utils.time_str(timer_epoch.t())
t_used = utils.time_str(timer_used.t())
t_estimate = utils.time_str(timer_used.t() / epoch * max_epoch)
if epoch <= max_epoch:
epoch_str = str(epoch)
else:
epoch_str = 'ex'
log_str = 'epoch {}, train {:.4f}|{:.4f}'.format(
epoch_str, aves['tl'], aves['ta'])
writer.add_scalars('loss', {'train': aves['tl']}, epoch)
writer.add_scalars('acc', {'train': aves['ta']}, epoch)
if eval_val:
log_str += ', val {:.4f}|{:.4f}'.format(aves['vl'], aves['va'])
writer.add_scalars('loss', {'val': aves['vl']}, epoch)
writer.add_scalars('acc', {'val': aves['va']}, epoch)
if eval_fs and (epoch % ef_epoch == 0 or epoch == max_epoch + 1):
log_str += ', fs'
for n_shot in n_shots:
key = 'fsa-' + str(n_shot)
log_str += ' {}: {:.4f}'.format(n_shot, aves[key])
writer.add_scalars('acc', {key: aves[key]}, epoch)
if epoch <= max_epoch:
log_str += ', {} {}/{}'.format(t_epoch, t_used, t_estimate)
else:
log_str += ', {}'.format(t_epoch)
utils.log(log_str)
if config.get('_parallel'):
model_ = model.module
else:
model_ = model
training = {
'epoch': epoch,
'optimizer': config['optimizer'],
'optimizer_args': config['optimizer_args'],
'optimizer_sd': optimizer.state_dict(),
}
save_obj = {
'file': __file__,
'config': config,
'model': config['model'],
'model_args': config['model_args'],
'model_sd': model_.state_dict(),
'training': training,
}
if epoch <= max_epoch:
torch.save(save_obj, os.path.join(save_path, 'epoch-last.pth'))
if (save_epoch is not None) and epoch % save_epoch == 0:
torch.save(
save_obj,
os.path.join(save_path, 'epoch-{}.pth'.format(epoch)))
if aves['va'] > max_va:
max_va = aves['va']
torch.save(save_obj, os.path.join(save_path, 'max-va.pth'))
else:
torch.save(save_obj, os.path.join(save_path, 'epoch-ex.pth'))
writer.flush()
def main(config):
# Environment setup
save_dir = config['save_dir']
utils.ensure_path(save_dir)
with open(osp.join(save_dir, 'config.yaml'), 'w') as f:
yaml.dump(config, f, sort_keys=False)
global log, writer
logger = set_logger(osp.join(save_dir, 'log.txt'))
log = logger.info
writer = SummaryWriter(osp.join(save_dir, 'tensorboard'))
os.environ['WANDB_NAME'] = config['exp_name']
os.environ['WANDB_DIR'] = config['save_dir']
if not config.get('wandb_upload', False):
os.environ['WANDB_MODE'] = 'dryrun'
t = config['wandb']
os.environ['WANDB_API_KEY'] = t['api_key']
wandb.init(project=t['project'], entity=t['entity'], config=config)
log('logging init done.')
log(f'wandb id: {wandb.run.id}')
# Dataset, model and optimizer
train_dataset = datasets.make((config['train_dataset']))
test_dataset = datasets.make((config['test_dataset']))
model = models.make(config['model'], args=None).cuda()
log(f'model #params: {utils.compute_num_params(model)}')
n_gpus = len(os.environ['CUDA_VISIBLE_DEVICES'].split(','))
if n_gpus > 1:
model = nn.DataParallel(model)
optimizer = utils.make_optimizer(model.parameters(), config['optimizer'])
train_loader = DataLoader(train_dataset, config['batch_size'], shuffle=True,
num_workers=8, pin_memory=True)
test_loader = DataLoader(test_dataset, config['batch_size'],
num_workers=8, pin_memory=True)
# Ready for training
max_epoch = config['max_epoch']
n_milestones = config.get('n_milestones', 1)
milestone_epoch = max_epoch // n_milestones
min_test_loss = 1e18
sample_batch_train = sample_data_batch(train_dataset).cuda()
sample_batch_test = sample_data_batch(test_dataset).cuda()
epoch_timer = utils.EpochTimer(max_epoch)
for epoch in range(1, max_epoch + 1):
log_text = f'epoch {epoch}'
# Train
model.train()
adjust_lr(optimizer, epoch, max_epoch, config)
log_temp_scalar('lr', optimizer.param_groups[0]['lr'], epoch)
ave_scalars = {k: utils.Averager() for k in ['loss']}
pbar = tqdm(train_loader, desc='train', leave=False)
for data in pbar:
data = data.cuda()
t = train_step(model, data, data, optimizer)
for k, v in t.items():
ave_scalars[k].add(v, len(data))
pbar.set_description(desc=f"train loss:{t['loss']:.4f}")
log_text += ', train:'
for k, v in ave_scalars.items():
v = v.item()
log_text += f' {k}={v:.4f}'
log_temp_scalar('train/' + k, v, epoch)
# Test
model.eval()
ave_scalars = {k: utils.Averager() for k in ['loss']}
pbar = tqdm(test_loader, desc='test', leave=False)
for data in pbar:
data = data.cuda()
t = eval_step(model, data, data)
for k, v in t.items():
ave_scalars[k].add(v, len(data))
pbar.set_description(desc=f"test loss:{t['loss']:.4f}")
log_text += ', test:'
for k, v in ave_scalars.items():
v = v.item()
log_text += f' {k}={v:.4f}'
log_temp_scalar('test/' + k, v, epoch)
test_loss = ave_scalars['loss'].item()
if epoch % milestone_epoch == 0:
with torch.no_grad():
pred = model(sample_batch_train).clamp(0, 1)
video_batch = torch.cat([sample_batch_train, pred], dim=0)
log_temp_videos('train/videos', video_batch, epoch)
img_batch = video_batch[:, :, 3, :, :]
log_temp_images('train/images', img_batch, epoch)
pred = model(sample_batch_test).clamp(0, 1)
video_batch = torch.cat([sample_batch_test, pred], dim=0)
log_temp_videos('test/videos', video_batch, epoch)
img_batch = video_batch[:, :, 3, :, :]
log_temp_images('test/images', img_batch, epoch)
# Summary and save
log_text += ', {} {}/{}'.format(*epoch_timer.step())
log(log_text)
model_ = model.module if n_gpus > 1 else model
model_spec = config['model']
model_spec['sd'] = model_.state_dict()
optimizer_spec = config['optimizer']
optimizer_spec['sd'] = optimizer.state_dict()
pth_file = {
'model': model_spec,
'optimizer': optimizer_spec,
'epoch': epoch,
}
if test_loss < min_test_loss:
min_test_loss = test_loss
wandb.run.summary['min_test_loss'] = min_test_loss
torch.save(pth_file, osp.join(save_dir, 'min-test-loss.pth'))
torch.save(pth_file, osp.join(save_dir, 'epoch-last.pth'))
writer.flush()
def main(config):
svname = args.name
if svname is None:
svname = 'meta_{}-{}shot'.format(config['train_dataset'],
config['n_shot'])
svname += '_' + config['model']
if config['model_args'].get('encoder'):
svname += '-' + config['model_args']['encoder']
if config['model_args'].get('prog_synthesis'):
svname += '-' + config['model_args']['prog_synthesis']
svname += '-seed' + str(args.seed)
if args.tag is not None:
svname += '_' + args.tag
save_path = os.path.join(args.save_dir, svname)
utils.ensure_path(save_path, remove=False)
utils.set_log_path(save_path)
writer = SummaryWriter(os.path.join(save_path, 'tensorboard'))
yaml.dump(config, open(os.path.join(save_path, 'config.yaml'), 'w'))
logger = utils.Logger(file_name=os.path.join(save_path, "log_sdout.txt"),
file_mode="a+",
should_flush=True)
#### Dataset ####
n_way, n_shot = config['n_way'], config['n_shot']
n_query = config['n_query']
if config.get('n_train_way') is not None:
n_train_way = config['n_train_way']
else:
n_train_way = n_way
if config.get('n_train_shot') is not None:
n_train_shot = config['n_train_shot']
else:
n_train_shot = n_shot
if config.get('ep_per_batch') is not None:
ep_per_batch = config['ep_per_batch']
else:
ep_per_batch = 1
random_state = np.random.RandomState(args.seed)
print('seed：', args.seed)
# train
train_dataset = datasets.make(config['train_dataset'],
**config['train_dataset_args'])
utils.log('train dataset: {} (x{})'.format(train_dataset[0][0].shape,
len(train_dataset)))
if config.get('visualize_datasets'):
utils.visualize_dataset(train_dataset, 'train_dataset', writer)
train_sampler = BongardSampler(train_dataset.n_tasks,
config['train_batches'], ep_per_batch,
random_state.randint(2**31))
train_loader = DataLoader(train_dataset,
batch_sampler=train_sampler,
num_workers=8,
pin_memory=True)
# tvals
tval_loaders = {}
tval_name_ntasks_dict = {
'tval': 2000,
'tval_ff': 600,
'tval_bd': 480,
'tval_hd_comb': 400,
'tval_hd_novel': 320
} # numbers depend on dataset
for tval_type in tval_name_ntasks_dict.keys():
if config.get('{}_dataset'.format(tval_type)):
tval_dataset = datasets.make(
config['{}_dataset'.format(tval_type)],
**config['{}_dataset_args'.format(tval_type)])
utils.log('{} dataset: {} (x{})'.format(tval_type,
tval_dataset[0][0].shape,
len(tval_dataset)))
if config.get('visualize_datasets'):
utils.visualize_dataset(tval_dataset, 'tval_ff_dataset',
writer)
tval_sampler = BongardSampler(
tval_dataset.n_tasks,
n_batch=tval_name_ntasks_dict[tval_type] // ep_per_batch,
ep_per_batch=ep_per_batch,
seed=random_state.randint(2**31))
tval_loader = DataLoader(tval_dataset,
batch_sampler=tval_sampler,
num_workers=8,
pin_memory=True)
tval_loaders.update({tval_type: tval_loader})
else:
tval_loaders.update({tval_type: None})
# val
val_dataset = datasets.make(config['val_dataset'],
**config['val_dataset_args'])
utils.log('val dataset: {} (x{})'.format(val_dataset[0][0].shape,
len(val_dataset)))
if config.get('visualize_datasets'):
utils.visualize_dataset(val_dataset, 'val_dataset', writer)
val_sampler = BongardSampler(val_dataset.n_tasks,
n_batch=900 // ep_per_batch,
ep_per_batch=ep_per_batch,
seed=random_state.randint(2**31))
val_loader = DataLoader(val_dataset,
batch_sampler=val_sampler,
num_workers=8,
pin_memory=True)
########
#### Model and optimizer ####
if config.get('load'):
print('loading pretrained model: ', config['load'])
model = models.load(torch.load(config['load']))
else:
model = models.make(config['model'], **config['model_args'])
if config.get('load_encoder'):
print('loading pretrained encoder: ', config['load_encoder'])
encoder = models.load(torch.load(config['load_encoder'])).encoder
model.encoder.load_state_dict(encoder.state_dict())
if config.get('load_prog_synthesis'):
print('loading pretrained program synthesis model: ',
config['load_prog_synthesis'])
prog_synthesis = models.load(
torch.load(config['load_prog_synthesis']))
model.prog_synthesis.load_state_dict(prog_synthesis.state_dict())
if config.get('_parallel'):
model = nn.DataParallel(model)
utils.log('num params: {}'.format(utils.compute_n_params(model)))
optimizer, lr_scheduler = utils.make_optimizer(model.parameters(),
config['optimizer'],
**config['optimizer_args'])
########
max_epoch = config['max_epoch']
save_epoch = config.get('save_epoch')
max_va = 0.
timer_used = utils.Timer()
timer_epoch = utils.Timer()
aves_keys = ['tl', 'ta', 'vl', 'va']
tval_tuple_lst = []
for k, v in tval_loaders.items():
if v is not None:
loss_key = 'tvl' + k.split('tval')[-1]
acc_key = ' tva' + k.split('tval')[-1]
aves_keys.append(loss_key)
aves_keys.append(acc_key)
tval_tuple_lst.append((k, v, loss_key, acc_key))
trlog = dict()
for k in aves_keys:
trlog[k] = []
for epoch in range(1, max_epoch + 1):
timer_epoch.s()
aves = {k: utils.Averager() for k in aves_keys}
# train
model.train()
if config.get('freeze_bn'):
utils.freeze_bn(model)
writer.add_scalar('lr', optimizer.param_groups[0]['lr'], epoch)
for data, label in tqdm(train_loader, desc='train', leave=False):
x_shot, x_query = fs.split_shot_query(data.cuda(),
n_train_way,
n_train_shot,
n_query,
ep_per_batch=ep_per_batch)
label_query = fs.make_nk_label(n_train_way,
n_query,
ep_per_batch=ep_per_batch).cuda()
if config['model'] == 'snail': # only use one selected label_query
query_dix = random_state.randint(n_train_way * n_query)
label_query = label_query.view(ep_per_batch, -1)[:, query_dix]
x_query = x_query[:, query_dix:query_dix + 1]
if config['model'] == 'maml': # need grad in maml
model.zero_grad()
logits = model(x_shot, x_query).view(-1, n_train_way)
loss = F.cross_entropy(logits, label_query)
acc = utils.compute_acc(logits, label_query)
optimizer.zero_grad()
loss.backward()
optimizer.step()
aves['tl'].add(loss.item())
aves['ta'].add(acc)
logits = None
loss = None
# eval
model.eval()
for name, loader, name_l, name_a in [('val', val_loader, 'vl', 'va')
] + tval_tuple_lst:
if config.get('{}_dataset'.format(name)) is None:
continue
np.random.seed(0)
for data, _ in tqdm(loader, desc=name, leave=False):
x_shot, x_query = fs.split_shot_query(
data.cuda(),
n_way,
n_shot,
n_query,
ep_per_batch=ep_per_batch)
label_query = fs.make_nk_label(
n_way, n_query, ep_per_batch=ep_per_batch).cuda()
if config[
'model'] == 'snail': # only use one randomly selected label_query
query_dix = random_state.randint(n_train_way)
label_query = label_query.view(ep_per_batch, -1)[:,
query_dix]
x_query = x_query[:, query_dix:query_dix + 1]
if config['model'] == 'maml': # need grad in maml
model.zero_grad()
logits = model(x_shot, x_query, eval=True).view(-1, n_way)
loss = F.cross_entropy(logits, label_query)
acc = utils.compute_acc(logits, label_query)
else:
with torch.no_grad():
logits = model(x_shot, x_query,
eval=True).view(-1, n_way)
loss = F.cross_entropy(logits, label_query)
acc = utils.compute_acc(logits, label_query)
aves[name_l].add(loss.item())
aves[name_a].add(acc)
# post
if lr_scheduler is not None:
lr_scheduler.step()
for k, v in aves.items():
aves[k] = v.item()
trlog[k].append(aves[k])
t_epoch = utils.time_str(timer_epoch.t())
t_used = utils.time_str(timer_used.t())
t_estimate = utils.time_str(timer_used.t() / epoch * max_epoch)
log_str = 'epoch {}, train {:.4f}|{:.4f}, val {:.4f}|{:.4f}'.format(
epoch, aves['tl'], aves['ta'], aves['vl'], aves['va'])
for tval_name, _, loss_key, acc_key in tval_tuple_lst:
log_str += ', {} {:.4f}|{:.4f}'.format(tval_name, aves[loss_key],
aves[acc_key])
writer.add_scalars('loss', {tval_name: aves[loss_key]}, epoch)
writer.add_scalars('acc', {tval_name: aves[acc_key]}, epoch)
log_str += ', {} {}/{}'.format(t_epoch, t_used, t_estimate)
utils.log(log_str)
writer.add_scalars('loss', {
'train': aves['tl'],
'val': aves['vl'],
}, epoch)
writer.add_scalars('acc', {
'train': aves['ta'],
'val': aves['va'],
}, epoch)
if config.get('_parallel'):
model_ = model.module
else:
model_ = model
training = {
'epoch': epoch,
'optimizer': config['optimizer'],
'optimizer_args': config['optimizer_args'],
'optimizer_sd': optimizer.state_dict(),
}
save_obj = {
'file': __file__,
'config': config,
'model': config['model'],
'model_args': config['model_args'],
'model_sd': model_.state_dict(),
'training': training,
}
torch.save(save_obj, os.path.join(save_path, 'epoch-last.pth'))
torch.save(trlog, os.path.join(save_path, 'trlog.pth'))
if (save_epoch is not None) and epoch % save_epoch == 0:
torch.save(save_obj,
os.path.join(save_path, 'epoch-{}.pth'.format(epoch)))
if aves['va'] > max_va:
max_va = aves['va']
torch.save(save_obj, os.path.join(save_path, 'max-va.pth'))
writer.flush()
print('finished training!')
logger.close()
def main(config):
svname = args.name
if svname is None:
svname = 'meta_{}-{}shot'.format(
config['train_dataset'], config['n_shot'])
svname += '_' + config['model'] + '-' + config['model_args']['encoder']
if args.tag is not None:
svname += '_' + args.tag
save_path = os.path.join('./save', svname)
utils.ensure_path(save_path)
utils.set_log_path(save_path)
writer = SummaryWriter(os.path.join(save_path, 'tensorboard'))
yaml.dump(config, open(os.path.join(save_path, 'config.yaml'), 'w'))
#### Dataset ####
n_way, n_shot = config['n_way'], config['n_shot']
n_query = config['n_query']
if config.get('n_train_way') is not None:
n_train_way = config['n_train_way']
else:
n_train_way = n_way
if config.get('n_train_shot') is not None:
n_train_shot = config['n_train_shot']
else:
n_train_shot = n_shot
if config.get('ep_per_batch') is not None:
ep_per_batch = config['ep_per_batch']
else:
ep_per_batch = 1
# train
train_dataset = datasets.make(config['train_dataset'],
**config['train_dataset_args'])
utils.log('train dataset: {} (x{}), {}'.format(
train_dataset[0][0].shape, len(train_dataset),
train_dataset.n_classes))
if config.get('visualize_datasets'):
utils.visualize_dataset(train_dataset, 'train_dataset', writer)
train_sampler = CategoriesSampler(
train_dataset.label, config['train_batches'],
n_train_way, n_train_shot + n_query,
ep_per_batch=ep_per_batch)
train_loader = DataLoader(train_dataset, batch_sampler=train_sampler,
num_workers=8, pin_memory=True)
# tval
if config.get('tval_dataset'):
tval_dataset = datasets.make(config['tval_dataset'],
**config['tval_dataset_args'])
utils.log('tval dataset: {} (x{}), {}'.format(
tval_dataset[0][0].shape, len(tval_dataset),
tval_dataset.n_classes))
if config.get('visualize_datasets'):
utils.visualize_dataset(tval_dataset, 'tval_dataset', writer)
tval_sampler = CategoriesSampler(
tval_dataset.label, 200,
n_way, n_shot + n_query,
ep_per_batch=4)
tval_loader = DataLoader(tval_dataset, batch_sampler=tval_sampler,
num_workers=8, pin_memory=True)
else:
tval_loader = None
# val
val_dataset = datasets.make(config['val_dataset'],
**config['val_dataset_args'])
utils.log('val dataset: {} (x{}), {}'.format(
val_dataset[0][0].shape, len(val_dataset),
val_dataset.n_classes))
if config.get('visualize_datasets'):
utils.visualize_dataset(val_dataset, 'val_dataset', writer)
val_sampler = CategoriesSampler(
val_dataset.label, 200,
n_way, n_shot + n_query,
ep_per_batch=4)
val_loader = DataLoader(val_dataset, batch_sampler=val_sampler,
num_workers=8, pin_memory=True)
########
#### Model and optimizer ####
if config.get('load'):
model_sv = torch.load(config['load'])
model = models.load(model_sv)
else:
model = models.make(config['model'], **config['model_args'])
if config.get('load_encoder'):
encoder = models.load(torch.load(config['load_encoder'])).encoder
model.encoder.load_state_dict(encoder.state_dict())
if config.get('_parallel'):
model = nn.DataParallel(model)
utils.log('num params: {}'.format(utils.compute_n_params(model)))
optimizer, lr_scheduler = utils.make_optimizer(
model.parameters(),
config['optimizer'], **config['optimizer_args'])
########
max_epoch = config['max_epoch']
save_epoch = config.get('save_epoch')
max_va = 0.
timer_used = utils.Timer()
timer_epoch = utils.Timer()
aves_keys = ['tl', 'ta', 'tvl', 'tva', 'vl', 'va']
trlog = dict()
for k in aves_keys:
trlog[k] = []
for epoch in range(1, max_epoch + 1):
timer_epoch.s()
aves = {k: utils.Averager() for k in aves_keys}
# train
model.train()
if config.get('freeze_bn'):
utils.freeze_bn(model)
writer.add_scalar('lr', optimizer.param_groups[0]['lr'], epoch)
np.random.seed(epoch)
for data, _ in tqdm(train_loader, desc='train', leave=False):
x_shot, x_query = fs.split_shot_query(
data.cuda(), n_train_way, n_train_shot, n_query,
ep_per_batch=ep_per_batch)
label = fs.make_nk_label(n_train_way, n_query,
ep_per_batch=ep_per_batch).cuda()
logits = model(x_shot, x_query).view(-1, n_train_way)
loss = F.cross_entropy(logits, label)
acc = utils.compute_acc(logits, label)
optimizer.zero_grad()
loss.backward()
optimizer.step()
aves['tl'].add(loss.item())
aves['ta'].add(acc)
logits = None; loss = None
# eval
model.eval()
for name, loader, name_l, name_a in [
('tval', tval_loader, 'tvl', 'tva'),
('val', val_loader, 'vl', 'va')]:
if (config.get('tval_dataset') is None) and name == 'tval':
continue
np.random.seed(0)
for data, _ in tqdm(loader, desc=name, leave=False):
x_shot, x_query = fs.split_shot_query(
data.cuda(), n_way, n_shot, n_query,
ep_per_batch=4)
label = fs.make_nk_label(n_way, n_query,
ep_per_batch=4).cuda()
with torch.no_grad():
logits = model(x_shot, x_query).view(-1, n_way)
loss = F.cross_entropy(logits, label)
acc = utils.compute_acc(logits, label)
aves[name_l].add(loss.item())
aves[name_a].add(acc)
_sig = int(_[-1])
# post
if lr_scheduler is not None:
lr_scheduler.step()
for k, v in aves.items():
aves[k] = v.item()
trlog[k].append(aves[k])
t_epoch = utils.time_str(timer_epoch.t())
t_used = utils.time_str(timer_used.t())
t_estimate = utils.time_str(timer_used.t() / epoch * max_epoch)
utils.log('epoch {}, train {:.4f}|{:.4f}, tval {:.4f}|{:.4f}, '
'val {:.4f}|{:.4f}, {} {}/{} (@{})'.format(
epoch, aves['tl'], aves['ta'], aves['tvl'], aves['tva'],
aves['vl'], aves['va'], t_epoch, t_used, t_estimate, _sig))
writer.add_scalars('loss', {
'train': aves['tl'],
'tval': aves['tvl'],
'val': aves['vl'],
}, epoch)
writer.add_scalars('acc', {
'train': aves['ta'],
'tval': aves['tva'],
'val': aves['va'],
}, epoch)
if config.get('_parallel'):
model_ = model.module
else:
model_ = model
training = {
'epoch': epoch,
'optimizer': config['optimizer'],
'optimizer_args': config['optimizer_args'],
'optimizer_sd': optimizer.state_dict(),
}
save_obj = {
'file': __file__,
'config': config,
'model': config['model'],
'model_args': config['model_args'],
'model_sd': model_.state_dict(),
'training': training,
}
torch.save(save_obj, os.path.join(save_path, 'epoch-last.pth'))
torch.save(trlog, os.path.join(save_path, 'trlog.pth'))
if (save_epoch is not None) and epoch % save_epoch == 0:
torch.save(save_obj,
os.path.join(save_path, 'epoch-{}.pth'.format(epoch)))
if aves['va'] > max_va:
max_va = aves['va']
torch.save(save_obj, os.path.join(save_path, 'max-va.pth'))
writer.flush()
def main(config):
# dataset
dataset = datasets.make(config['dataset'], **config['dataset_args'])
utils.log('dataset: {} (x{}), {}'.format(dataset[0][0].shape, len(dataset),
dataset.n_classes))
if not args.sauc:
n_way = 5
else:
n_way = 2
n_shot, n_unlabel, n_query = args.shot, 30, 15
n_batch = 200
ep_per_batch = 4
batch_sampler = CategoriesSampler_Semi(dataset.label,
n_batch,
n_way,
n_shot,
n_unlabel,
n_query,
ep_per_batch=ep_per_batch)
loader = DataLoader(dataset,
batch_sampler=batch_sampler,
num_workers=8,
pin_memory=True)
# model
if config.get('load') is None:
model = models.make('meta-baseline', encoder=None)
else:
model = models.load(torch.load(config['load']))
if config.get('load_encoder') is not None:
encoder = models.load(torch.load(config['load_encoder'])).encoder
model.encoder = encoder
if config.get('_parallel'):
model = nn.DataParallel(model)
model.eval()
utils.log('num params: {}'.format(utils.compute_n_params(model)))
# testing
aves_keys = ['vl', 'va']
aves = {k: utils.Averager() for k in aves_keys}
test_epochs = args.test_epochs
np.random.seed(0)
va_lst = []
for epoch in range(1, test_epochs + 1):
for data, _ in tqdm(loader, leave=False):
x_shot, x_unlabel, x_query = fs.split_shot_query_semi(
data.cuda(),
n_way,
n_shot,
n_unlabel,
n_query,
ep_per_batch=ep_per_batch)
with torch.no_grad():
if not args.sauc:
logits = model(x_shot, x_unlabel, x_query).view(-1, n_way)
label = fs.make_nk_label(n_way,
n_query,
ep_per_batch=ep_per_batch).cuda()
loss = F.cross_entropy(logits, label)
acc = utils.compute_acc(logits, label)
aves['vl'].add(loss.item(), len(data))
aves['va'].add(acc, len(data))
va_lst.append(acc)
else:
x_shot = x_shot[:, 0, :, :, :, :].contiguous()
shot_shape = x_shot.shape[:-3]
img_shape = x_shot.shape[-3:]
bs = shot_shape[0]
p = model.encoder(x_shot.view(-1, *img_shape)).reshape(
*shot_shape, -1).mean(dim=1, keepdim=True)
q = model.encoder(x_query.view(-1, *img_shape)).view(
bs, -1, p.shape[-1])
p = F.normalize(p, dim=-1)
q = F.normalize(q, dim=-1)
s = torch.bmm(q, p.transpose(2, 1)).view(bs, -1).cpu()
for i in range(bs):
k = s.shape[1] // 2
y_true = [1] * k + [0] * k
acc = roc_auc_score(y_true, s[i])
aves['va'].add(acc, len(data))
va_lst.append(acc)
print('test epoch {}: acc={:.2f} +- {:.2f} (%), loss={:.4f} (@{})'.
format(epoch, aves['va'].item() * 100,
mean_confidence_interval(va_lst) * 100, aves['vl'].item(),
_[-1]))
def main(config, args):
random.seed(0)
np.random.seed(0)
torch.manual_seed(0)
torch.cuda.manual_seed(0)
# torch.backends.cudnn.deterministic = True
# torch.backends.cudnn.benchmark = False
wandb_auth()
try:
__IPYTHON__
wandb.init(project="NAS", group=f"maml")
except:
wandb.init(project="NAS", group=f"maml", config=config)
ckpt_name = args.name
if ckpt_name is None:
ckpt_name = config['encoder']
ckpt_name += '_' + config['dataset'].replace('meta-', '')
ckpt_name += '_{}_way_{}_shot'.format(config['train']['n_way'],
config['train']['n_shot'])
if args.tag is not None:
ckpt_name += '_' + args.tag
ckpt_path = os.path.join('./save', ckpt_name)
utils.ensure_path(ckpt_path)
utils.set_log_path(ckpt_path)
writer = SummaryWriter(os.path.join(ckpt_path, 'tensorboard'))
yaml.dump(config, open(os.path.join(ckpt_path, 'config.yaml'), 'w'))
##### Dataset #####
# meta-train
train_set = datasets.make(config['dataset'], **config['train'])
utils.log('meta-train set: {} (x{}), {}'.format(train_set[0][0].shape,
len(train_set),
train_set.n_classes))
# meta-val
eval_val = False
if config.get('val'):
eval_val = True
val_set = datasets.make(config['dataset'], **config['val'])
utils.log('meta-val set: {} (x{}), {}'.format(val_set[0][0].shape,
len(val_set),
val_set.n_classes))
val_loader = DataLoader(val_set,
config['val']['n_episode'],
collate_fn=datasets.collate_fn,
num_workers=1,
pin_memory=True)
# if args.split == "traintrain" and config.get('val'): # TODO I dont think this is what they meant by train-train :D
# train_set = torch.utils.data.ConcatDataset([train_set, val_set])
train_loader = DataLoader(train_set,
config['train']['n_episode'],
collate_fn=datasets.collate_fn,
num_workers=1,
pin_memory=True)
##### Model and Optimizer #####
inner_args = utils.config_inner_args(config.get('inner_args'))
if config.get('load') or (args.load is True and
os.path.exists(ckpt_path + '/epoch-last.pth')):
if config.get('load') is None:
config['load'] = ckpt_path + '/epoch-last.pth'
ckpt = torch.load(config['load'])
config['encoder'] = ckpt['encoder']
config['encoder_args'] = ckpt['encoder_args']
config['classifier'] = ckpt['classifier']
config['classifier_args'] = ckpt['classifier_args']
model = models.load(ckpt,
load_clf=(not inner_args['reset_classifier']))
optimizer, lr_scheduler = optimizers.load(ckpt, model.parameters())
start_epoch = ckpt['training']['epoch'] + 1
max_va = ckpt['training']['max_va']
else:
config['encoder_args'] = config.get('encoder_args') or dict()
config['classifier_args'] = config.get('classifier_args') or dict()
config['encoder_args']['bn_args']['n_episode'] = config['train'][
'n_episode']
config['classifier_args']['n_way'] = config['train']['n_way']
model = models.make(config['encoder'], config['encoder_args'],
config['classifier'], config['classifier_args'])
optimizer, lr_scheduler = optimizers.make(config['optimizer'],
model.parameters(),
**config['optimizer_args'])
start_epoch = 1
max_va = 0.
if args.efficient:
model.go_efficient()
if config.get('_parallel'):
model = nn.DataParallel(model)
utils.log('num params: {}'.format(utils.compute_n_params(model)))
timer_elapsed, timer_epoch = utils.Timer(), utils.Timer()
##### Training and evaluation #####
# 'tl': meta-train loss
# 'ta': meta-train accuracy
# 'vl': meta-val loss
# 'va': meta-val accuracy
aves_keys = ['tl', 'ta', 'vl', 'va']
trlog = dict()
for k in aves_keys:
trlog[k] = []
for epoch in tqdm(range(start_epoch, config['epoch'] + 1),
desc="Iterating over epochs"):
timer_epoch.start()
aves = {k: utils.AverageMeter() for k in aves_keys}
# meta-train
model.train()
writer.add_scalar('lr', optimizer.param_groups[0]['lr'], epoch)
np.random.seed(epoch)
all_sotls = 0
all_sovls = 0
for data_idx, data in enumerate(
tqdm(train_loader, desc='meta-train', leave=False)):
x_shot, x_query, y_shot, y_query = data
x_shot, y_shot = x_shot.cuda(), y_shot.cuda()
x_query, y_query = x_query.cuda(), y_query.cuda()
if inner_args['reset_classifier']:
if config.get('_parallel'):
model.module.reset_classifier()
else:
model.reset_classifier()
if args.split == "traintrain":
x_query = x_shot
y_query = y_shot
logits, sotl, all_losses = model(x_shot,
x_query,
y_shot,
inner_args,
meta_train=True)
# print("HAHHA", data_idx, all_losses)
# sotl = sum([l[-1] for l in all_losses])
# for l in all_losses[:-1]:
# for i in range(len(l)-1):
# l[i] = l[i].detach()
logits = logits.flatten(0, 1)
labels = y_query.flatten()
all_sotls += sotl
pred = torch.argmax(logits, dim=-1)
acc = utils.compute_acc(pred, labels)
loss = F.cross_entropy(logits, labels)
# all_sovls += loss # TODO I think this causes blowup because it creates new tensors that never get discarded and it maintains the computational graph after?
if args.split == "trainval" or (
args.split == "sovl"
and not data_idx % args.sotl_freq == 0):
aves['tl'].update(loss.item(), 1)
aves['ta'].update(acc, 1)
optimizer.zero_grad()
loss.backward()
for param in optimizer.param_groups[0]['params']:
nn.utils.clip_grad_value_(param, 10)
optimizer.step()
elif args.split == "traintrain":
aves['tl'].update(loss.item(), 1)
aves['ta'].update(acc, 1)
# sotl = sum(sotl) + loss
optimizer.zero_grad()
# sotl.backward()
loss.backward()
for param in optimizer.param_groups[0]['params']:
nn.utils.clip_grad_value_(param, 10)
optimizer.step()
elif args.split == "sotl" and data_idx % args.sotl_freq == 0:
# TODO doesnt work whatsoever
aves['tl'].update(loss.item(), 1)
aves['ta'].update(acc, 1)
optimizer.zero_grad()
all_sotls.backward()
for param in optimizer.param_groups[0]['params']:
nn.utils.clip_grad_value_(param, 10)
optimizer.step()
all_sotls = 0 # detach
elif args.split == "sovl" and data_idx % args.sotl_freq == 0:
# TODO doesnt work whatsoever
aves['tl'].update(loss.item(), 1)
aves['ta'].update(acc, 1)
optimizer.zero_grad()
all_sovls.backward()
for param in optimizer.param_groups[0]['params']:
nn.utils.clip_grad_value_(param, 10)
optimizer.step()
all_sovls = 0 # detach
# meta-val
if eval_val:
model.eval()
np.random.seed(0)
for data in tqdm(val_loader, desc='meta-val', leave=False):
x_shot, x_query, y_shot, y_query = data
x_shot, y_shot = x_shot.cuda(), y_shot.cuda()
x_query, y_query = x_query.cuda(), y_query.cuda()
if inner_args['reset_classifier']:
if config.get('_parallel'):
model.module.reset_classifier()
else:
model.reset_classifier()
logits, sotl, all_losses = model(x_shot,
x_query,
y_shot,
inner_args,
meta_train=False)
logits = logits.flatten(0, 1)
labels = y_query.flatten()
pred = torch.argmax(logits, dim=-1)
acc = utils.compute_acc(pred, labels)
loss = F.cross_entropy(logits, labels)
aves['vl'].update(loss.item(), 1)
aves['va'].update(acc, 1)
if lr_scheduler is not None:
lr_scheduler.step()
for k, avg in aves.items():
aves[k] = avg.item()
trlog[k].append(aves[k])
t_epoch = utils.time_str(timer_epoch.end())
t_elapsed = utils.time_str(timer_elapsed.end())
t_estimate = utils.time_str(timer_elapsed.end() /
(epoch - start_epoch + 1) *
(config['epoch'] - start_epoch + 1))
# formats output
log_str = 'epoch {}, meta-train {:.4f}|{:.4f}'.format(
str(epoch), aves['tl'], aves['ta'])
writer.add_scalars('loss', {'meta-train': aves['tl']}, epoch)
writer.add_scalars('acc', {'meta-train': aves['ta']}, epoch)
if eval_val:
log_str += ', meta-val {:.4f}|{:.4f}'.format(
aves['vl'], aves['va'])
writer.add_scalars('loss', {'meta-val': aves['vl']}, epoch)
writer.add_scalars('acc', {'meta-val': aves['va']}, epoch)
wandb.log({
"train_loss": aves['tl'],
"train_acc": aves['ta'],
"val_loss": aves['vl'],
"val_acc": aves['va']
})
log_str += ', {} {}/{}'.format(t_epoch, t_elapsed, t_estimate)
utils.log(log_str)
# saves model and meta-data
if config.get('_parallel'):
model_ = model.module
else:
model_ = model
training = {
'epoch':
epoch,
'max_va':
max(max_va, aves['va']),
'optimizer':
config['optimizer'],
'optimizer_args':
config['optimizer_args'],
'optimizer_state_dict':
optimizer.state_dict(),
'lr_scheduler_state_dict':
lr_scheduler.state_dict() if lr_scheduler is not None else None,
}
ckpt = {
'file': __file__,
'config': config,
'encoder': config['encoder'],
'encoder_args': config['encoder_args'],
'encoder_state_dict': model_.encoder.state_dict(),
'classifier': config['classifier'],
'classifier_args': config['classifier_args'],
'classifier_state_dict': model_.classifier.state_dict(),
'training': training,
}
# 'epoch-last.pth': saved at the latest epoch
# 'max-va.pth': saved when validation accuracy is at its maximum
torch.save(ckpt, os.path.join(ckpt_path, 'epoch-last.pth'))
torch.save(trlog, os.path.join(ckpt_path, 'trlog.pth'))
if aves['va'] > max_va:
max_va = aves['va']
torch.save(ckpt, os.path.join(ckpt_path, 'max-va.pth'))
writer.flush()
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('--config')
parser.add_argument('--model')
parser.add_argument('--gpu', default='0')
args = parser.parse_args()
os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu
with open(args.config, 'r') as f:
config = yaml.load(f, Loader=yaml.FullLoader)
spec = config['test_dataset']
dataset = datasets.make(spec['dataset'])
dataset = datasets.make(spec['wrapper'], args={'dataset': dataset})
loader = DataLoader(dataset,
batch_size=spec['batch_size'],
num_workers=8,
pin_memory=True)
model_spec = torch.load(args.model)['model']
model = models.make(model_spec, load_sd=True).cuda()
res = eval_psnr(loader,
model,
data_norm=config.get('data_norm'),
eval_type=config.get('eval_type'),
eval_bsize=config.get('eval_bsize'),
verbose=True)
def main(config):
random.seed(0)
np.random.seed(0)
torch.manual_seed(0)
torch.cuda.manual_seed(0)
# torch.backends.cudnn.deterministic = True
# torch.backends.cudnn.benchmark = False
ckpt_name = args.name
if ckpt_name is None:
ckpt_name = config['encoder']
ckpt_name += '_' + config['dataset'].replace('meta-', '')
ckpt_name += '_{}_way_{}_shot'.format(config['train']['n_way'],
config['train']['n_shot'])
if args.tag is not None:
ckpt_name += '_' + args.tag
ckpt_path = os.path.join('./save', ckpt_name)
utils.ensure_path(ckpt_path)
utils.set_log_path(ckpt_path)
writer = SummaryWriter(os.path.join(ckpt_path, 'tensorboard'))
yaml.dump(config, open(os.path.join(ckpt_path, 'config.yaml'), 'w'))
##### Dataset #####
# meta-train
train_set = datasets.make(config['dataset'], **config['train'])
utils.log('meta-train set: {} (x{}), {}'.format(train_set[0][0].shape,
len(train_set),
train_set.n_classes))
train_loader = DataLoader(train_set,
config['train']['n_episode'],
collate_fn=datasets.collate_fn,
num_workers=1,
pin_memory=True)
# meta-val
eval_val = False
if config.get('val'):
eval_val = True
val_set = datasets.make(config['dataset'], **config['val'])
utils.log('meta-val set: {} (x{}), {}'.format(val_set[0][0].shape,
len(val_set),
val_set.n_classes))
val_loader = DataLoader(val_set,
config['val']['n_episode'],
collate_fn=datasets.collate_fn,
num_workers=1,
pin_memory=True)
##### Model and Optimizer #####
inner_args = utils.config_inner_args(config.get('inner_args'))
if config.get('load'):
ckpt = torch.load(config['load'])
config['encoder'] = ckpt['encoder']
config['encoder_args'] = ckpt['encoder_args']
config['classifier'] = ckpt['classifier']
config['classifier_args'] = ckpt['classifier_args']
model = models.load(ckpt,
load_clf=(not inner_args['reset_classifier']))
optimizer, lr_scheduler = optimizers.load(ckpt, model.parameters())
start_epoch = ckpt['training']['epoch'] + 1
max_va = ckpt['training']['max_va']
else:
config['encoder_args'] = config.get('encoder_args') or dict()
config['classifier_args'] = config.get('classifier_args') or dict()
config['encoder_args']['bn_args']['n_episode'] = config['train'][
'n_episode']
config['classifier_args']['n_way'] = config['train']['n_way']
model = models.make(config['encoder'], config['encoder_args'],
config['classifier'], config['classifier_args'])
optimizer, lr_scheduler = optimizers.make(config['optimizer'],
model.parameters(),
**config['optimizer_args'])
start_epoch = 1
max_va = 0.
if args.efficient:
model.go_efficient()
if config.get('_parallel'):
model = nn.DataParallel(model)
utils.log('num params: {}'.format(utils.compute_n_params(model)))
timer_elapsed, timer_epoch = utils.Timer(), utils.Timer()
##### Training and evaluation #####
# 'tl': meta-train loss
# 'ta': meta-train accuracy
# 'vl': meta-val loss
# 'va': meta-val accuracy
aves_keys = ['tl', 'ta', 'vl', 'va']
trlog = dict()
for k in aves_keys:
trlog[k] = []
for epoch in range(start_epoch, config['epoch'] + 1):
timer_epoch.start()
aves = {k: utils.AverageMeter() for k in aves_keys}
# meta-train
model.train()
writer.add_scalar('lr', optimizer.param_groups[0]['lr'], epoch)
np.random.seed(epoch)
for data in tqdm(train_loader, desc='meta-train', leave=False):
x_shot, x_query, y_shot, y_query = data
x_shot, y_shot = x_shot.cuda(), y_shot.cuda()
x_query, y_query = x_query.cuda(), y_query.cuda()
if inner_args['reset_classifier']:
if config.get('_parallel'):
model.module.reset_classifier()
else:
model.reset_classifier()
logits = model(x_shot,
x_query,
y_shot,
inner_args,
meta_train=True)
logits = logits.flatten(0, 1)
labels = y_query.flatten()
pred = torch.argmax(logits, dim=-1)
acc = utils.compute_acc(pred, labels)
loss = F.cross_entropy(logits, labels)
aves['tl'].update(loss.item(), 1)
aves['ta'].update(acc, 1)
optimizer.zero_grad()
loss.backward()
for param in optimizer.param_groups[0]['params']:
nn.utils.clip_grad_value_(param, 10)
optimizer.step()
# meta-val
if eval_val:
model.eval()
np.random.seed(0)
for data in tqdm(val_loader, desc='meta-val', leave=False):
x_shot, x_query, y_shot, y_query = data
x_shot, y_shot = x_shot.cuda(), y_shot.cuda()
x_query, y_query = x_query.cuda(), y_query.cuda()
if inner_args['reset_classifier']:
if config.get('_parallel'):
model.module.reset_classifier()
else:
model.reset_classifier()
logits = model(x_shot,
x_query,
y_shot,
inner_args,
meta_train=False)
logits = logits.flatten(0, 1)
labels = y_query.flatten()
pred = torch.argmax(logits, dim=-1)
acc = utils.compute_acc(pred, labels)
loss = F.cross_entropy(logits, labels)
aves['vl'].update(loss.item(), 1)
aves['va'].update(acc, 1)
if lr_scheduler is not None:
lr_scheduler.step()
for k, avg in aves.items():
aves[k] = avg.item()
trlog[k].append(aves[k])
t_epoch = utils.time_str(timer_epoch.end())
t_elapsed = utils.time_str(timer_elapsed.end())
t_estimate = utils.time_str(timer_elapsed.end() /
(epoch - start_epoch + 1) *
(config['epoch'] - start_epoch + 1))
# formats output
log_str = 'epoch {}, meta-train {:.4f}|{:.4f}'.format(
str(epoch), aves['tl'], aves['ta'])
writer.add_scalars('loss', {'meta-train': aves['tl']}, epoch)
writer.add_scalars('acc', {'meta-train': aves['ta']}, epoch)
if eval_val:
log_str += ', meta-val {:.4f}|{:.4f}'.format(
aves['vl'], aves['va'])
writer.add_scalars('loss', {'meta-val': aves['vl']}, epoch)
writer.add_scalars('acc', {'meta-val': aves['va']}, epoch)
log_str += ', {} {}/{}'.format(t_epoch, t_elapsed, t_estimate)
utils.log(log_str)
# saves model and meta-data
if config.get('_parallel'):
model_ = model.module
else:
model_ = model
training = {
'epoch':
epoch,
'max_va':
max(max_va, aves['va']),
'optimizer':
config['optimizer'],
'optimizer_args':
config['optimizer_args'],
'optimizer_state_dict':
optimizer.state_dict(),
'lr_scheduler_state_dict':
lr_scheduler.state_dict() if lr_scheduler is not None else None,
}
ckpt = {
'file': __file__,
'config': config,
'encoder': config['encoder'],
'encoder_args': config['encoder_args'],
'encoder_state_dict': model_.encoder.state_dict(),
'classifier': config['classifier'],
'classifier_args': config['classifier_args'],
'classifier_state_dict': model_.classifier.state_dict(),
'training': training,
}
# 'epoch-last.pth': saved at the latest epoch
# 'max-va.pth': saved when validation accuracy is at its maximum
torch.save(ckpt, os.path.join(ckpt_path, 'epoch-last.pth'))
torch.save(trlog, os.path.join(ckpt_path, 'trlog.pth'))
if aves['va'] > max_va:
max_va = aves['va']
torch.save(ckpt, os.path.join(ckpt_path, 'max-va.pth'))
writer.flush()
def main(config):
svname = args.name
if svname is None:
svname = 'moco_{}'.format(config['train_dataset'])
svname += '_' + config['model_args']['encoder']
out_dim = config['model_args']['encoder_args']['out_dim']
svname += '-out_dim' + str(out_dim)
svname += '-seed' + str(args.seed)
if args.tag is not None:
svname += '_' + args.tag
save_path = os.path.join(args.save_dir, svname)
utils.ensure_path(save_path, remove=False)
utils.set_log_path(save_path)
writer = SummaryWriter(os.path.join(save_path, 'tensorboard'))
yaml.dump(config, open(os.path.join(save_path, 'config.yaml'), 'w'))
random_state = np.random.RandomState(args.seed)
print('seed：', args.seed)
logger = utils.Logger(file_name=os.path.join(save_path, "log_sdout.txt"),
file_mode="a+",
should_flush=True)
#### Dataset ####
# train
train_dataset = datasets.make(config['train_dataset'],
**config['train_dataset_args'])
train_loader = DataLoader(train_dataset,
config['batch_size'],
shuffle=True,
num_workers=8,
pin_memory=True,
drop_last=True)
utils.log('train dataset: {} (x{})'.format(train_dataset[0][0][0].shape,
len(train_dataset)))
if config.get('visualize_datasets'):
utils.visualize_dataset(train_dataset, 'train_dataset', writer)
# val
if config.get('val_dataset'):
eval_val = True
val_dataset = datasets.make(config['val_dataset'],
**config['val_dataset_args'])
val_loader = DataLoader(val_dataset,
config['batch_size'],
num_workers=8,
pin_memory=True,
drop_last=True)
utils.log('val dataset: {} (x{})'.format(val_dataset[0][0][0].shape,
len(val_dataset)))
if config.get('visualize_datasets'):
utils.visualize_dataset(val_dataset, 'val_dataset', writer)
else:
eval_val = False
# few-shot eval
if config.get('eval_fs'):
ef_epoch = config.get('eval_fs_epoch')
if ef_epoch is None:
ef_epoch = 5
eval_fs = True
n_way = 2
n_query = 1
n_shot = 6
if config.get('ep_per_batch') is not None:
ep_per_batch = config['ep_per_batch']
else:
ep_per_batch = 1
# tvals
fs_loaders = {}
tval_name_ntasks_dict = {
'tval': 2000,
'tval_ff': 600,
'tval_bd': 480,
'tval_hd_comb': 400,
'tval_hd_novel': 320
} # numbers depend on dataset
for tval_type in tval_name_ntasks_dict.keys():
if config.get('{}_dataset'.format(tval_type)):
tval_dataset = datasets.make(
config['{}_dataset'.format(tval_type)],
**config['{}_dataset_args'.format(tval_type)])
utils.log('{} dataset: {} (x{})'.format(
tval_type, tval_dataset[0][0][0].shape, len(tval_dataset)))
if config.get('visualize_datasets'):
utils.visualize_dataset(tval_dataset, 'tval_ff_dataset',
writer)
tval_sampler = BongardSampler(
tval_dataset.n_tasks,
n_batch=tval_name_ntasks_dict[tval_type] // ep_per_batch,
ep_per_batch=ep_per_batch,
seed=random_state.randint(2**31))
tval_loader = DataLoader(tval_dataset,
batch_sampler=tval_sampler,
num_workers=8,
pin_memory=True)
fs_loaders.update({tval_type: tval_loader})
else:
fs_loaders.update({tval_type: None})
else:
eval_fs = False
########
#### Model and Optimizer ####
if config.get('load'):
model_sv = torch.load(config['load'])
model = models.load(model_sv)
else:
model = models.make(config['model'], **config['model_args'])
if eval_fs:
fs_model = models.make('meta-baseline', encoder=None)
fs_model.encoder = model.encoder
if config.get('_parallel'):
model = nn.DataParallel(model)
if eval_fs:
fs_model = nn.DataParallel(fs_model)
utils.log('num params: {}'.format(utils.compute_n_params(model)))
optimizer, lr_scheduler = utils.make_optimizer(model.parameters(),
config['optimizer'],
**config['optimizer_args'])
########
max_epoch = config['max_epoch']
save_epoch = config.get('save_epoch')
max_va = 0.
timer_used = utils.Timer()
timer_epoch = utils.Timer()
for epoch in range(1, max_epoch + 1 + 1):
timer_epoch.s()
aves_keys = ['tl', 'ta', 'vl', 'va', 'tvl', 'tva']
if eval_fs:
for k, v in fs_loaders.items():
if v is not None:
aves_keys += ['fsa' + k.split('tval')[-1]]
aves = {ave_k: utils.Averager() for ave_k in aves_keys}
# train
model.train()
writer.add_scalar('lr', optimizer.param_groups[0]['lr'], epoch)
for data, _ in tqdm(train_loader, desc='train', leave=False):
logits, label = model(im_q=data[0].cuda(), im_k=data[1].cuda())
loss = F.cross_entropy(logits, label)
acc = utils.compute_acc(logits, label)
optimizer.zero_grad()
loss.backward()
optimizer.step()
aves['tl'].add(loss.item())
aves['ta'].add(acc)
logits = None
loss = None
# val
if eval_val:
model.eval()
for data, _ in tqdm(val_loader, desc='val', leave=False):
with torch.no_grad():
logits, label = model(im_q=data[0].cuda(),
im_k=data[1].cuda())
loss = F.cross_entropy(logits, label)
acc = utils.compute_acc(logits, label)
aves['vl'].add(loss.item())
aves['va'].add(acc)
if eval_fs and (epoch % ef_epoch == 0 or epoch == max_epoch + 1):
fs_model.eval()
for k, v in fs_loaders.items():
if v is not None:
ave_key = 'fsa' + k.split('tval')[-1]
np.random.seed(0)
for data, _ in tqdm(v, desc=ave_key, leave=False):
x_shot, x_query = fs.split_shot_query(
data[0].cuda(),
n_way,
n_shot,
n_query,
ep_per_batch=ep_per_batch)
label_query = fs.make_nk_label(
n_way, n_query, ep_per_batch=ep_per_batch).cuda()
with torch.no_grad():
logits = fs_model(x_shot, x_query).view(-1, n_way)
acc = utils.compute_acc(logits, label_query)
aves[ave_key].add(acc)
# post
if lr_scheduler is not None:
lr_scheduler.step()
for k, v in aves.items():
aves[k] = v.item()
t_epoch = utils.time_str(timer_epoch.t())
t_used = utils.time_str(timer_used.t())
t_estimate = utils.time_str(timer_used.t() / epoch * max_epoch)
if epoch <= max_epoch:
epoch_str = str(epoch)
else:
epoch_str = 'ex'
log_str = 'epoch {}, train {:.4f}|{:.4f}'.format(
epoch_str, aves['tl'], aves['ta'])
writer.add_scalars('loss', {'train': aves['tl']}, epoch)
writer.add_scalars('acc', {'train': aves['ta']}, epoch)
if eval_val:
log_str += ', val {:.4f}|{:.4f}, tval {:.4f}|{:.4f}'.format(
aves['vl'], aves['va'], aves['tvl'], aves['tva'])
writer.add_scalars('loss', {'val': aves['vl']}, epoch)
writer.add_scalars('loss', {'tval': aves['tvl']}, epoch)
writer.add_scalars('acc', {'val': aves['va']}, epoch)
writer.add_scalars('acc', {'tval': aves['tva']}, epoch)
if eval_fs and (epoch % ef_epoch == 0 or epoch == max_epoch + 1):
log_str += ', fs'
for ave_key in aves_keys:
if 'fsa' in ave_key:
log_str += ' {}: {:.4f}'.format(ave_key, aves[ave_key])
writer.add_scalars('acc', {ave_key: aves[ave_key]}, epoch)
if epoch <= max_epoch:
log_str += ', {} {}/{}'.format(t_epoch, t_used, t_estimate)
else:
log_str += ', {}'.format(t_epoch)
utils.log(log_str)
if config.get('_parallel'):
model_ = model.module
else:
model_ = model
training = {
'epoch': epoch,
'optimizer': config['optimizer'],
'optimizer_args': config['optimizer_args'],
'optimizer_sd': optimizer.state_dict(),
}
save_obj = {
'file': __file__,
'config': config,
'model': config['model'],
'model_args': config['model_args'],
'model_sd': model_.state_dict(),
'training': training,
}
if epoch <= max_epoch:
torch.save(save_obj, os.path.join(save_path, 'epoch-last.pth'))
if (save_epoch is not None) and epoch % save_epoch == 0:
torch.save(
save_obj,
os.path.join(save_path, 'epoch-{}.pth'.format(epoch)))
if aves['va'] > max_va:
max_va = aves['va']
torch.save(save_obj, os.path.join(save_path, 'max-va.pth'))
else:
torch.save(save_obj, os.path.join(save_path, 'epoch-ex.pth'))
writer.flush()
print('finished training!')
logger.close()