def check_np_traning():
traverse_list = ['a9a']
for name in traverse_list:
n_features = _config.datasets[name]['n_features']
is_multi = _config.datasets[name]['is_multi']
has_test = _config.datasets[name]['has_test']
try:
X_train, y_train, X_test, y_test = get_libsvm_data(
name, n_features, is_multi, has_test)
except (Exception):
continue
# -------------------
# Train np model
# -------------------
# svc
kernels = ['linear', 'rbf', 'poly', 'sigmoid']
for kernel in kernels:
print('---{} svc---'.format(kernel))
svc = construct_np_model(model_type='svc', kernel=kernel)
svc = train_np_model(svc, X_train, y_train)
eval_np_model(svc, X_train, y_train, X_test, y_test)
alpha_i = np.abs(svc.dual_coef_[0])
print('max_alpha_i:{}'.format(np.max(alpha_i)))
print('min_alpha_i:{}'.format(np.min(alpha_i)))
def check_kernel(kernel, dataset_name):
name = dataset_name
n_features = _config.datasets[name]['n_features']
is_multi = _config.datasets[name]['is_multi']
has_test = _config.datasets[name]['has_test']
X_train, y_train, X_test, y_test = get_libsvm_data(name,
n_features,
is_multi,
has_test,
show=False)
svc = construct_np_model(model_type='svc', kernel=kernel)
svc = train_np_model(svc, X_train, y_train)
X_fit = X_train[svc.support_]
coef = svc.dual_coef_
intercept = svc.intercept_
params = svc.get_params()
if 'gamma' in params:
params['gamma'] = 1 / (
n_features *
X_train.var()) if params['gamma'] == 'scale' else params['gamma']
print('----show kernel metrics---')
km_interface = cal_km(params, X_fit, X_train[:10], type='interface')
km_realize = cal_km(params, X_fit, X_train[:10], type='realize')
print('\n sklearn:')
print(km_interface)
print(km_interface.shape)
print('\n realize:')
print(km_realize)
print(km_realize.shape)
print('')
print('---show hand-10 fx ---')
fx_interface = cal_fx(km_interface, coef, intercept)
fx_realize = cal_fx(km_realize, coef, intercept)
print('\n fx_interface:')
print(fx_interface)
print('\n fx_realize:')
print(fx_realize)
print('\n decsion function:')
print(svc.decision_function(X_train[:10]))
print('\n y_pred:')
print(svc.predict(X_train[:10]))
print('')
print('---show composite fx---')
km1 = cal_km(params, X_fit[:5000], X_train[:10], type='realize')
fx1 = cal_fx(km1, coef[:, :5000], 0)
km2 = cal_km(params, X_fit[5000:], X_train[:10], type='realize')
fx2 = cal_fx(km2, coef[:, 5000:], 0)
print('\n fx1-5000 ins')
print(fx1)
print('\n fx2-others')
print(fx2)
print('\n batch fx = fx1 + fx2 + intercept')
print(fx1 + fx2 + intercept)
print('')
def conv_check(kernel, dataset_name):
# 训练svc
name = dataset_name
n_features = _config.datasets[name]['n_features']
is_multi = _config.datasets[name]['is_multi']
has_test = _config.datasets[name]['has_test']
X_train, y_train, X_test, y_test = get_libsvm_data(name,
n_features,
is_multi,
has_test,
show=False)
svc = construct_np_model(model_type='svc', kernel=kernel)
svc = train_np_model(svc, X_train, y_train)
X_fit = X_train[svc.support_]
y_fit = y_train[svc.support_]
coef = svc.dual_coef_
intercept = svc.intercept_
params = svc.get_params()
if 'gamma' in params:
params['gamma'] = 1 / (
n_features *
X_train.var()) if params['gamma'] == 'scale' else params['gamma']
# 数据分类别分批次输入
map_size = 128
n_blocks = [2, 2, 2, 2]
# 根据coef对X_fit等升序排序
sorted_idx = np.argsort(coef[0])
sorted_coef = coef[0][sorted_idx].reshape(-1, 1)
sorted_X_fit = X_fit[sorted_idx]
sorted_y_fit = y_fit[sorted_idx]
# 舍弃低权重的instance
fit_dataloaders, n_fit = get_fit_dataloaders(sorted_X_fit, sorted_coef,
sorted_y_fit, _config.labels,
map_size)
# 展示原分布
print('----data distribution----')
print('train data - num:{}, distribution:{}'.format(
X_train.shape[0], sorted(Counter(y_train))))
print('fit data - num:{}, distribution:{}'.format(X_fit.shape[0],
sorted(Counter(y_fit))))
print(' ')
# 确定压缩率(如果map_size过小，90%会报错)
rate_list = [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9]
for rate in rate_list:
# 计算每个map压缩的数量
n_compressed = X_fit.shape[0] * (1 - rate)
rate = n_compressed / n_fit
n_generate = int(rate * map_size)
conv_module = CNN_Res18(Basic_Block, n_blocks, n_generate)
if torch.cuda.is_available():
conv_module = conv_module.cuda()
X_gen = None
y_gen = None
is_none = True
for label in _config.labels:
fit_dataloader = fit_dataloaders[label]
for data in fit_dataloader:
X_map, coef_map = data
X_map = Variable(X_map)
if torch.cuda.is_available():
X_map = X_map.cuda()
X_gen_temp = conv_module(X_map)
y_gen_temp = np.full((X_gen_temp.shape[0], ), label)
if is_none:
print(X_gen_temp.size())
X_gen = X_gen_temp.cpu().detach().numpy()
y_gen = y_gen_temp
is_none = False
else:
X_gen = np.concatenate(
(X_gen, X_gen_temp.cpu().detach().numpy()), axis=0)
y_gen = np.concatenate((y_gen, y_gen_temp), axis=0)
#展示压缩后分布
print('{:.2f}% data - num:{}, distribution:{}'.format(
rate * 100, X_gen.shape[0], sorted(Counter(y_gen).items())))
def datamodule_check(kernel, dataset_name):
# 训练svc
name = dataset_name
n_features = _config.datasets[name]['n_features']
is_multi = _config.datasets[name]['is_multi']
has_test = _config.datasets[name]['has_test']
X_train, y_train, X_test, y_test = get_libsvm_data(name,
n_features,
is_multi,
has_test,
show=False)
svc = construct_np_model(model_type='svc', kernel=kernel)
svc = train_np_model(svc, X_train, y_train)
X_fit = X_train[svc.support_]
y_fit = y_train[svc.support_]
coef = svc.dual_coef_
intercept = svc.intercept_
params = svc.get_params()
if 'gamma' in params:
params['gamma'] = 1 / (
n_features *
X_train.var()) if params['gamma'] == 'scale' else params['gamma']
# 数据分类别分批次输入
map_size = 128
# 根据coef对X_fit等升序排序
sorted_idx = np.argsort(coef[0])
sorted_coef = coef[0][sorted_idx].reshape(-1, 1)
sorted_X_fit = X_fit[sorted_idx]
sorted_y_fit = y_fit[sorted_idx]
# 舍弃低权重的instance
fit_dataloaders, n_fit = get_fit_dataloaders(sorted_X_fit, sorted_coef,
sorted_y_fit, _config.labels,
map_size)
for label in _config.labels:
fit_dataloader = fit_dataloaders[label]
for data in fit_dataloader:
X_map, coef_map = data
coef_map = coef_map.view(1, -1)
gdm = GeneralDataModule(_config.datasets[name]['n_features'],
_config.n_classes,
_config.labels,
batch_size=64)
gdm.setup(X_train, X_test, params, X_map[0][0].numpy(),
coef_map.numpy(), label)
train_loader = gdm.train_dataloader()
val_loader = gdm.val_dataloader()
test_loader = gdm.test_dataloader()
print('train data')
for train_data in train_loader:
X_batch, fx_batch = train_data
print('X size:', X_batch.size())
print('fx size:', fx_batch.size())
print(fx_batch)
break
print('val data')
for val_data in val_loader:
X_batch, fx_batch = val_data
print(X_batch.size())
print('fx size:', fx_batch.size())
print(fx_batch)
break
print('test data')
for test_data in test_loader:
X_batch, fx_batch = test_data
print(X_batch.size())
print('fx size:', fx_batch.size())
print(fx_batch)
break
break
is_multi = datasets[name]['is_multi']
has_test = datasets[name]['has_test']
# -------------------
# Load Data
# -------------------
print('...Load data')
X_train, y_train, X_test, y_test = get_libsvm_data(name, n_features, is_multi, has_test)
# -------------------
# Train np model
# -------------------
print('\n... Train np model')
np_model = construct_np_model(model_type=model_type, kernel=kernel)
np_model = train_np_model(np_model, X_train, y_train)
eval_np_model(np_model, X_train, y_train, X_test, y_test)
# 获取参数并计算fx
X_fit = X_train[np_model.support_]
y_fit = y_train[np_model.support_]
params = np_model.get_params()
if 'gamma' in params:
params['gamma'] = 1 / (n_features * X_train.var()) if params['gamma'] == 'scale' else params['gamma']
coef = np_model.dual_coef_
intercept = np_model.intercept_
km = cal_km(params, X_fit, X_test[:10], type='realize')
fx = cal_fx(km, coef, intercept)
print('\nfx-realize:')