def single_test(params, results_logger):
acc_all = []
iter_num = 600
few_shot_params = dict(n_way=params.test_n_way, n_support=params.n_shot)
if params.dataset in ['omniglot', 'cross_char']:
assert params.model == 'Conv4' and not params.train_aug, 'omniglot only support Conv4 without augmentation'
params.model = 'Conv4S'
if params.method == 'baseline':
model = BaselineFinetune(model_dict[params.model], **few_shot_params)
elif params.method == 'baseline++':
model = BaselineFinetune(model_dict[params.model],
loss_type='dist',
**few_shot_params)
elif params.method == 'protonet':
model = ProtoNet(model_dict[params.model], **few_shot_params)
elif params.method == 'DKT':
model = DKT(model_dict[params.model], **few_shot_params)
elif params.method == 'matchingnet':
model = MatchingNet(model_dict[params.model], **few_shot_params)
elif params.method in ['relationnet', 'relationnet_softmax']:
if params.model == 'Conv4':
feature_model = backbone.Conv4NP
elif params.model == 'Conv6':
feature_model = backbone.Conv6NP
elif params.model == 'Conv4S':
feature_model = backbone.Conv4SNP
else:
feature_model = lambda: model_dict[params.model](flatten=False)
loss_type = 'mse' if params.method == 'relationnet' else 'softmax'
model = RelationNet(feature_model,
loss_type=loss_type,
**few_shot_params)
elif params.method in ['maml', 'maml_approx']:
backbone.ConvBlock.maml = True
backbone.SimpleBlock.maml = True
backbone.BottleneckBlock.maml = True
backbone.ResNet.maml = True
model = MAML(model_dict[params.model],
approx=(params.method == 'maml_approx'),
**few_shot_params)
if params.dataset in ['omniglot', 'cross_char'
]: # maml use different parameter in omniglot
model.n_task = 32
model.task_update_num = 1
model.train_lr = 0.1
else:
raise ValueError('Unknown method')
model = model.cuda()
checkpoint_dir = '%s/checkpoints/%s/%s_%s' % (
configs.save_dir, params.dataset, params.model, params.method)
if params.train_aug:
checkpoint_dir += '_aug'
if not params.method in ['baseline', 'baseline++']:
checkpoint_dir += '_%dway_%dshot' % (params.train_n_way, params.n_shot)
# modelfile = get_resume_file(checkpoint_dir)
if not params.method in ['baseline', 'baseline++']:
if params.save_iter != -1:
modelfile = get_assigned_file(checkpoint_dir, params.save_iter)
else:
modelfile = get_best_file(checkpoint_dir)
if modelfile is not None:
tmp = torch.load(modelfile)
model.load_state_dict(tmp['state'])
else:
print("[WARNING] Cannot find 'best_file.tar' in: " +
str(checkpoint_dir))
split = params.split
if params.save_iter != -1:
split_str = split + "_" + str(params.save_iter)
else:
split_str = split
if params.method in ['maml', 'maml_approx',
'DKT']: # maml do not support testing with feature
if 'Conv' in params.model:
if params.dataset in ['omniglot', 'cross_char']:
image_size = 28
else:
image_size = 84
else:
image_size = 224
datamgr = SetDataManager(image_size,
n_eposide=iter_num,
n_query=15,
**few_shot_params)
if params.dataset == 'cross':
if split == 'base':
loadfile = configs.data_dir['miniImagenet'] + 'all.json'
else:
loadfile = configs.data_dir['CUB'] + split + '.json'
elif params.dataset == 'cross_char':
if split == 'base':
loadfile = configs.data_dir['omniglot'] + 'noLatin.json'
else:
loadfile = configs.data_dir['emnist'] + split + '.json'
else:
loadfile = configs.data_dir[params.dataset] + split + '.json'
novel_loader = datamgr.get_data_loader(loadfile, aug=False)
if params.adaptation:
model.task_update_num = 100 # We perform adaptation on MAML simply by updating more times.
model.eval()
acc_mean, acc_std = model.test_loop(novel_loader, return_std=True)
else:
novel_file = os.path.join(
checkpoint_dir.replace("checkpoints",
"features"), split_str + ".hdf5"
) # defaut split = novel, but you can also test base or val classes
cl_data_file = feat_loader.init_loader(novel_file)
for i in range(iter_num):
acc = feature_evaluation(cl_data_file,
model,
n_query=15,
adaptation=params.adaptation,
**few_shot_params)
acc_all.append(acc)
acc_all = np.asarray(acc_all)
acc_mean = np.mean(acc_all)
acc_std = np.std(acc_all)
print('%d Test Acc = %4.2f%% +- %4.2f%%' %
(iter_num, acc_mean, 1.96 * acc_std / np.sqrt(iter_num)))
with open('record/results.txt', 'a') as f:
timestamp = time.strftime("%Y%m%d-%H%M%S", time.localtime())
aug_str = '-aug' if params.train_aug else ''
aug_str += '-adapted' if params.adaptation else ''
if params.method in ['baseline', 'baseline++']:
exp_setting = '%s-%s-%s-%s%s %sshot %sway_test' % (
params.dataset, split_str, params.model, params.method,
aug_str, params.n_shot, params.test_n_way)
else:
exp_setting = '%s-%s-%s-%s%s %sshot %sway_train %sway_test' % (
params.dataset, split_str, params.model, params.method,
aug_str, params.n_shot, params.train_n_way, params.test_n_way)
acc_str = '%d Test Acc = %4.2f%% +- %4.2f%%' % (
iter_num, acc_mean, 1.96 * acc_std / np.sqrt(iter_num))
f.write('Time: %s, Setting: %s, Acc: %s \n' %
(timestamp, exp_setting, acc_str))
results_logger.log("single_test_acc", acc_mean)
results_logger.log("single_test_acc_std",
1.96 * acc_std / np.sqrt(iter_num))
results_logger.log("time", timestamp)
results_logger.log("exp_setting", exp_setting)
results_logger.log("acc_str", acc_str)
return acc_mean
def get_logits_targets(params):
acc_all = []
iter_num = 600
few_shot_params = dict(n_way = params.test_n_way , n_support = params.n_shot)
if params.dataset in ['omniglot', 'cross_char']:
assert params.model == 'Conv4' and not params.train_aug ,'omniglot only support Conv4 without augmentation'
params.model = 'Conv4S'
if params.method == 'baseline':
model = BaselineFinetune( model_dict[params.model], **few_shot_params )
elif params.method == 'baseline++':
model = BaselineFinetune( model_dict[params.model], loss_type = 'dist', **few_shot_params )
elif params.method == 'protonet':
model = ProtoNet( model_dict[params.model], **few_shot_params )
elif params.method == 'DKT':
model = DKT(model_dict[params.model], **few_shot_params)
elif params.method == 'matchingnet':
model = MatchingNet( model_dict[params.model], **few_shot_params )
elif params.method in ['relationnet', 'relationnet_softmax']:
if params.model == 'Conv4':
feature_model = backbone.Conv4NP
elif params.model == 'Conv6':
feature_model = backbone.Conv6NP
elif params.model == 'Conv4S':
feature_model = backbone.Conv4SNP
else:
feature_model = lambda: model_dict[params.model]( flatten = False )
loss_type = 'mse' if params.method == 'relationnet' else 'softmax'
model = RelationNet( feature_model, loss_type = loss_type , **few_shot_params )
elif params.method in ['maml' , 'maml_approx']:
backbone.ConvBlock.maml = True
backbone.SimpleBlock.maml = True
backbone.BottleneckBlock.maml = True
backbone.ResNet.maml = True
model = MAML( model_dict[params.model], approx = (params.method == 'maml_approx') , **few_shot_params )
if params.dataset in ['omniglot', 'cross_char']: #maml use different parameter in omniglot
model.n_task = 32
model.task_update_num = 1
model.train_lr = 0.1
else:
raise ValueError('Unknown method')
model = model.cuda()
checkpoint_dir = '%s/checkpoints/%s/%s_%s' %(configs.save_dir, params.dataset, params.model, params.method)
if params.train_aug:
checkpoint_dir += '_aug'
if not params.method in ['baseline', 'baseline++'] :
checkpoint_dir += '_%dway_%dshot' %( params.train_n_way, params.n_shot)
#modelfile = get_resume_file(checkpoint_dir)
if not params.method in ['baseline', 'baseline++'] :
if params.save_iter != -1:
modelfile = get_assigned_file(checkpoint_dir,params.save_iter)
else:
modelfile = get_best_file(checkpoint_dir)
if modelfile is not None:
tmp = torch.load(modelfile)
model.load_state_dict(tmp['state'])
else:
print("[WARNING] Cannot find 'best_file.tar' in: " + str(checkpoint_dir))
split = params.split
if params.save_iter != -1:
split_str = split + "_" +str(params.save_iter)
else:
split_str = split
if params.method in ['maml', 'maml_approx', 'DKT']: #maml do not support testing with feature
if 'Conv' in params.model:
if params.dataset in ['omniglot', 'cross_char']:
image_size = 28
else:
image_size = 84
else:
image_size = 224
datamgr = SetDataManager(image_size, n_eposide = iter_num, n_query = 15 , **few_shot_params)
if params.dataset == 'cross':
if split == 'base':
loadfile = configs.data_dir['miniImagenet'] + 'all.json'
else:
loadfile = configs.data_dir['CUB'] + split +'.json'
elif params.dataset == 'cross_char':
if split == 'base':
loadfile = configs.data_dir['omniglot'] + 'noLatin.json'
else:
loadfile = configs.data_dir['emnist'] + split +'.json'
else:
loadfile = configs.data_dir[params.dataset] + split + '.json'
novel_loader = datamgr.get_data_loader( loadfile, aug = False)
if params.adaptation:
model.task_update_num = 100 #We perform adaptation on MAML simply by updating more times.
model.eval()
logits_list = list()
targets_list = list()
for i, (x,_) in enumerate(novel_loader):
logits = model.get_logits(x).detach()
targets = torch.tensor(np.repeat(range(params.test_n_way), model.n_query)).cuda()
logits_list.append(logits) #.cpu().detach().numpy())
targets_list.append(targets) #.cpu().detach().numpy())
else:
novel_file = os.path.join( checkpoint_dir.replace("checkpoints","features"), split_str +".hdf5")
cl_data_file = feat_loader.init_loader(novel_file)
logits_list = list()
targets_list = list()
n_query = 15
n_way = few_shot_params['n_way']
n_support = few_shot_params['n_support']
class_list = cl_data_file.keys()
for i in range(iter_num):
#----------------------
select_class = random.sample(class_list,n_way)
z_all = []
for cl in select_class:
img_feat = cl_data_file[cl]
perm_ids = np.random.permutation(len(img_feat)).tolist()
z_all.append( [ np.squeeze( img_feat[perm_ids[i]]) for i in range(n_support+n_query) ] ) # stack each batch
z_all = torch.from_numpy(np.array(z_all))
model.n_query = n_query
logits = model.set_forward(z_all, is_feature = True).detach()
targets = torch.tensor(np.repeat(range(n_way), n_query)).cuda()
logits_list.append(logits)
targets_list.append(targets)
#----------------------
return torch.cat(logits_list, 0), torch.cat(targets_list, 0)
backbone.BottleneckBlock.maml = True
backbone.ResNet.maml = True
model = MAML(model_dict[params.model],
approx=(params.method == 'maml_approx'),
**few_shot_params)
if params.dataset in ['omniglot', 'cross_char'
]: #maml use different parameter in omniglot
model.n_task = 32
model.task_update_num = 1
model.train_lr = 0.1
else:
raise ValueError('Unknown method')
device = params.device
with torch.cuda.device(device):
model = model.cuda()
checkpoint_dir = '%s/checkpoints/%s/%s_%s' % (
configs.save_dir, params.dataset, params.model, params.method)
if params.dataset == 'CUB' and params.train_class == "CUB_mini":
checkpoint_dir = '%s/checkpoints/%s/%s_%s' % (
configs.save_dir, "CUB_mini", params.model, params.method)
if params.train_class == "all_Imagenet":
checkpoint_dir = '%s/checkpoints/%s/%s_%s' % (
configs.save_dir, "all_Imagenet", params.model, params.method)
if params.train_aug:
checkpoint_dir += '_aug'
if params.selection_classes != -1:
checkpoint_dir += "_{0}classes".format(params.selection_classes)
backbone.ResNet.maml = True
train_model = MAML(model_dict[params.model],
approx=(params.method == 'maml_approx'),
**train_few_shot_params)
if params.dataset in [
'omniglot', 'cross_char'
]: # maml use different parameter in omniglot
train_model.n_task = 32
train_model.task_update_num = 1
train_model.train_lr = 0.1
test_model = train_model
else:
raise ValueError('Unknown method')
train_model = train_model.cuda()
test_model = test_model.cuda()
key = "%s_%s_%s_Temp-%s_Margin-%s" % (params.model, time.strftime('%c'),
params.loss_type, params.temperature,
params.margin)
if params.train_aug:
key += '_aug'
if not params.method == 'baseline':
key += '_%dway_%dshot' % (params.train_n_way, params.n_shot)
params.checkpoint_dir = '%s/checkpoints/%s/%s' % (configs.save_dir,
params.dataset, key)
if not os.path.isdir(params.checkpoint_dir):
os.makedirs(params.checkpoint_dir)
start_epoch = params.start_epoch