def run_exp(params, verbose):
print_params(params)
n_episodes = 10000
episode_params = dict(n_way=params.n_way,
n_support=params.n_shot,
n_query=params.n_query,
n_semi=params.n_semi)
model = ProtoSemi(opt=params)
model = model.cuda()
train_mean, cl_data_file = load_features(params)
acc_list = []
start_time = time.perf_counter()
for i in range(1, n_episodes + 1):
acc = run_episode(train_mean, cl_data_file, model, **episode_params)
acc_list += acc
if i % verbose == 0:
print_msg(i, n_episodes, start_time, acc_list, acc)
res = [avg(acc_list[ind::len(acc)]) for ind in range(len(acc))]
ci = [ci_95(acc_list[ind::len(acc)]) for ind in range(len(acc))]
return res, ci
def run_exp(params, verbose):
print_params(params)
n_episodes = 10000
few_shot_params = dict(n_way=params.n_way, n_support=params.n_shot)
model = ProtoMSP(opt=params)
model = model.cuda()
train_mean, cl_data_file = load_features(params)
time_list = []
name = [
'baseline',
'sub reg',
'sub',
'pca',
'ica',
'cluster',
'pca cluster',
'ica cluster',
'msp',
'pca msp',
'ica msp',
]
for i in range(1, n_episodes + 1):
times = run_episode(train_mean,
cl_data_file,
model,
n_query=params.n_query,
**few_shot_params)
time_list.append(times)
if i % verbose == 0:
tl = sum(time_list) / len(time_list)
msg = ''
for k, v in zip(name, tl):
msg += f'{k}: {v:.2e}, '
print(msg)
return