528 lines
22 KiB
Python
528 lines
22 KiB
Python
import os
|
||
import torch
|
||
import torch.nn as nn
|
||
from datasets import load_dataset
|
||
from tqdm import tqdm
|
||
import numpy as np
|
||
import argparse
|
||
from train_utils import get_last_non_padded_token_rep, compute_ot_loss_cos, update_centroids_ema, update_centroids_ema_hard, get_ex_data, collate_fn
|
||
from transformers import AutoTokenizer, AutoModelForCausalLM
|
||
from llm_layers import add_sv_layers
|
||
from sklearn.metrics import roc_auc_score
|
||
from torch.cuda.amp import autocast, GradScaler
|
||
import torch.nn.functional as F
|
||
from sinkhorn_knopp import SinkhornKnopp_imb
|
||
import logging
|
||
from utils import load_npy_shapes,split_indices
|
||
|
||
|
||
def seed_everything(seed: int):
|
||
import random, os
|
||
import numpy as np
|
||
import torch
|
||
|
||
random.seed(seed)
|
||
os.environ['PYTHONHASHSEED'] = str(seed)
|
||
np.random.seed(seed)
|
||
torch.manual_seed(seed)
|
||
torch.cuda.manual_seed(seed)
|
||
torch.backends.cudnn.deterministic = True
|
||
torch.backends.cudnn.benchmark = True
|
||
|
||
|
||
def train_model(model, optimizer, device, prompts, labels, args):
|
||
"""
|
||
对应论文里的两阶段训练：
|
||
- Phase 1: 使用人工标注的 exemplar 做 initial training（式 (3)(4)(5)(6)）
|
||
- Phase 2: 使用 OT + Sinkhorn 选出来的伪标签数据做 self-training（Sec.3.3 pseudo-labeling）
|
||
|
||
参数说明：
|
||
- model: 冻结好的 LLM + 已经插入 TSV 层的模型（只训练 TSV）
|
||
- optimizer: 只优化 TSV 的 AdamW
|
||
- device: cuda
|
||
- prompts: [test_prompts, train_prompts, exemplar_prompts]
|
||
- labels: 对应三块的标签 [test_labels, train_labels, exemplar_labels]
|
||
- args: 超参数（学习率、epoch 数、Sinkhorn 参数等）
|
||
"""
|
||
|
||
layer_number = -1 # 使用最后一层 hidden state（这里 -1 当索引）
|
||
|
||
# ========= 日志 & 结果保存目录 =========
|
||
dir_name = f"SV_{args.model_name}_{args.dataset_name}/{args.component}/{args.str_layer}/{args.lam}"
|
||
log_dir = f"/{dir_name}/"
|
||
log_file = os.path.join(log_dir, f"log.txt")
|
||
os.makedirs(dir_name, exist_ok=True)
|
||
|
||
logging.basicConfig(
|
||
filename=log_file,
|
||
filemode="w",
|
||
level=logging.INFO,
|
||
format="%(asctime)s - %(levelname)s - %(message)s",
|
||
)
|
||
|
||
logging.info("Starting training")
|
||
logging.info(
|
||
f"component={args.component}, str_layer={args.str_layer}"
|
||
)
|
||
|
||
# 解包数据：test / wild train / exemplar
|
||
test_prompts, train_prompts = prompts[0], prompts[1]
|
||
test_labels, train_labels = labels[0], labels[1]
|
||
|
||
batch_size = args.batch_size
|
||
losses = []
|
||
best_test_auroc = -1
|
||
|
||
scaler = GradScaler() # 混合精度的梯度缩放器
|
||
|
||
num_trains = args.num_train
|
||
|
||
|
||
|
||
# 用 exemplar 的标签估计类分布 w（truthful/hallu 的先验）
|
||
# ex_hallu = P(hallucinated), ex_true = P(truthful)
|
||
ex_hallu = (num_exemplars - exemplar_labels[:num_exemplars].sum()) / num_exemplars
|
||
ex_true = (exemplar_labels[:num_exemplars].sum()) / num_exemplars
|
||
cls_dist = torch.tensor([ex_hallu, ex_true]).float().cuda()
|
||
cls_dist = cls_dist.view(-1, 1) # 形状 [2, 1]
|
||
|
||
# 实例化带类别边缘约束的 Sinkhorn（实现论文里的 OT 问题）
|
||
sinkhorn = SinkhornKnopp_imb(args, cls_dist)
|
||
|
||
# ========= 初始化两个类别的“原型向量” μ_c =========
|
||
# 这里 centroids 就是论文中球面上的 μ_truthful, μ_hallucinated
|
||
centroids = torch.randn((2, model.config.hidden_size)).half().cuda()
|
||
centroids = F.normalize(centroids, p=2, dim=1)
|
||
|
||
# 把 exemplar 的 prompt/label 整理成一个大 batch（padding）
|
||
train_prompts_, train_labels_ = train_prompts, train_labels
|
||
train_prompts, train_labels = collate_fn(train_prompts, train_labels)
|
||
|
||
# ========= Phase 1: Initial training on exemplars =========
|
||
num_epochs = args.init_num_epochs
|
||
|
||
for epoch in range(num_epochs):
|
||
running_loss = 0.0
|
||
total = 0
|
||
num_samples = num_trains
|
||
|
||
# 在 exemplar 上分 batch 训练（对应论文中 D_E 只含人工标注数据）
|
||
for batch_start in tqdm(
|
||
range(0, num_samples, batch_size),
|
||
desc=f"Epoch {epoch+1}/{num_epochs} Batches",
|
||
leave=False,
|
||
):
|
||
batch_prompts = train_prompts[batch_start: batch_start + batch_size]
|
||
batch_labels = train_labels[batch_start: batch_start + batch_size]
|
||
|
||
# attention_mask: 1 = valid token, 0 = padding
|
||
attention_mask = (batch_prompts != 0).half()
|
||
|
||
batch_prompts = batch_prompts.to(device)
|
||
batch_labels = batch_labels.to(device)
|
||
attention_mask = attention_mask.to(batch_prompts.device)
|
||
|
||
# ======= 前向传播：取最后一层最后非 padding token 的表示 r_v =======
|
||
with autocast(dtype=torch.float16):
|
||
output = model(
|
||
batch_prompts.squeeze(),
|
||
attention_mask=attention_mask.squeeze(),
|
||
output_hidden_states=True,
|
||
)
|
||
|
||
hidden_states = output.hidden_states # tuple(len=L+1)
|
||
hidden_states = torch.stack(hidden_states, dim=0).squeeze()
|
||
last_layer_hidden_state = hidden_states[layer_number] # [B, T, H]
|
||
|
||
# 对应论文里 r_v：最后非 padding token 的 hidden（包含 SV steer）
|
||
last_token_rep = get_last_non_padded_token_rep(
|
||
last_layer_hidden_state, attention_mask.squeeze()
|
||
)
|
||
|
||
# 把标签变成 one-hot（2 维）
|
||
batch_labels_oh = torch.nn.functional.one_hot(
|
||
batch_labels, num_classes=-1 # 这里 -1 应该在 utils 里特殊处理成 2
|
||
)
|
||
|
||
# compute_ot_loss_cos：对应 Eq.(5)，不过是用 OT 的版本：
|
||
# - 通过与 centroids 的余弦距离计算类似 softmax 的分类损失
|
||
# - 里面会用 args.cos_temp 做温度缩放
|
||
ot_loss, similarities = compute_ot_loss_cos(
|
||
last_token_rep, centroids, batch_labels_oh, batch_size, args
|
||
)
|
||
|
||
loss = ot_loss
|
||
total += batch_labels.size(0)
|
||
|
||
# ====== 更新类别原型 μ_c（EMA）对应论文 Eq.(6) ======
|
||
with torch.no_grad():
|
||
centroids = update_centroids_ema_hard(
|
||
centroids, last_token_rep, batch_labels_oh, args
|
||
)
|
||
|
||
# ====== 只对 TSV 反传，LLM 本体是冻结的 ======
|
||
scaler.scale(loss).backward()
|
||
scaler.step(optimizer)
|
||
scaler.update()
|
||
optimizer.zero_grad()
|
||
|
||
running_loss += loss.item() * batch_labels.size(0)
|
||
|
||
# 一个 epoch 的平均 loss
|
||
epoch_loss = running_loss / total
|
||
|
||
# ====== 在 test 上评估，记录 AUROC ======
|
||
if (epoch + 1) % 1 == 0:
|
||
test_labels_ = test_labels
|
||
test_predictions, test_labels_combined = test_model(
|
||
model, centroids, test_prompts, test_labels_, device, batch_size, layer_number
|
||
)
|
||
|
||
test_auroc = roc_auc_score(
|
||
test_labels_combined.cpu().numpy(), test_predictions.cpu().numpy()
|
||
)
|
||
|
||
print(f"Epoch [{epoch+1}/{num_epochs}], Loss: {epoch_loss:.4f}")
|
||
logging.info(f"Epoch [{epoch+1}/{num_epochs}], Loss: {epoch_loss:.4f}")
|
||
losses.append(epoch_loss)
|
||
|
||
if test_auroc > best_test_auroc:
|
||
best_test_auroc = test_auroc
|
||
best_test_epoch = epoch
|
||
print(f"Best test AUROC: {best_test_auroc:.4f}, at epoch: {best_test_epoch}")
|
||
logging.info(
|
||
f"Best test AUROC: {best_test_auroc:.4f}, at epoch: {best_test_epoch}"
|
||
)
|
||
|
||
logging.info(
|
||
f"Epoch [{epoch+1}/{num_epochs}], Train Loss: {epoch_loss:.4f}, "
|
||
)
|
||
logging.info(f"Test AUROC: {test_auroc:.4f}")
|
||
print(f"Epoch [{epoch+1}/{num_epochs}],Test AUROC: {test_auroc:.4f}")
|
||
|
||
# ========= 进入 Phase 2：伪标签 + 自训练 =========
|
||
logging.info(f"SS Learning Starts")
|
||
|
||
# 1) get_ex_data: 在 wild train 上做 TSV 推断 + Sinkhorn OT，选出高置信伪标签样本
|
||
with torch.no_grad():
|
||
selected_indices, selected_labels_soft = get_ex_data(
|
||
model,
|
||
train_prompts, # wild 区间的样本
|
||
train_labels, # 这里一般是无标签 or dummy，这里暂时没用
|
||
batch_size,
|
||
centroids, # 当前原型
|
||
sinkhorn, # OT + Sinkhorn 对象，用来求 Q
|
||
args.num_selected_data, # 选多少个伪标签样本
|
||
cls_dist,
|
||
args,
|
||
)
|
||
|
||
num_samples = len(selected_indices) + args.num_exemplars
|
||
|
||
num_epochs = args.aug_num_epochs # 自训练的 epoch 数
|
||
|
||
exemplar_label = torch.tensor(exemplar_labels).cuda()
|
||
|
||
# 2) 构造“增强后的训练集”：伪标签样本 + 原来的 exemplar
|
||
selected_prompts = [train_prompts[i] for i in selected_indices]
|
||
selected_labels = selected_labels_soft # 这里已经是 soft label（OT 输出的 q）
|
||
|
||
augmented_prompts = selected_prompts + exemplar_prompts_
|
||
exemplar_labels = torch.nn.functional.one_hot(
|
||
exemplar_label.to(torch.int64), num_classes=2
|
||
)
|
||
augmented_labels = torch.concat(
|
||
(selected_labels, torch.tensor(exemplar_labels).clone().cuda())
|
||
)
|
||
|
||
augmented_prompts_train = augmented_prompts
|
||
augmented_labels_label = augmented_labels
|
||
|
||
num_samples = len(augmented_prompts_train)
|
||
|
||
# 3) 在“伪标签 + exemplar”上再训练一轮 TSV（自训练）
|
||
with autocast(dtype=torch.float16):
|
||
for epoch in range(num_epochs):
|
||
running_loss = 0.0
|
||
total = 0
|
||
all_labels = []
|
||
|
||
for batch_start in tqdm(
|
||
range(0, num_samples, batch_size),
|
||
desc=f"Epoch {epoch+1}/{num_epochs} Batches",
|
||
leave=False,
|
||
):
|
||
batch_prompts = augmented_prompts_train[batch_start: batch_start + batch_size]
|
||
batch_labels = augmented_labels_label[batch_start: batch_start + batch_size]
|
||
|
||
batch_prompts, batch_labels = collate_fn(batch_prompts, batch_labels)
|
||
|
||
attention_mask = (batch_prompts != 0).half()
|
||
|
||
batch_prompts = batch_prompts.to(device)
|
||
batch_labels = batch_labels.to(device)
|
||
attention_mask = attention_mask.to(batch_prompts.device)
|
||
|
||
output = model(
|
||
batch_prompts.squeeze(),
|
||
attention_mask=attention_mask.squeeze(),
|
||
output_hidden_states=True,
|
||
)
|
||
|
||
hidden_states = output.hidden_states
|
||
hidden_states = torch.stack(hidden_states, dim=0).squeeze()
|
||
last_layer_hidden_state = hidden_states[layer_number]
|
||
|
||
last_token_rep = get_last_non_padded_token_rep(
|
||
last_layer_hidden_state, attention_mask.squeeze()
|
||
)
|
||
|
||
# 这里 compute_ot_loss_cos 接收的是 soft label (OT 得到的 q)
|
||
ot_loss, similarities = compute_ot_loss_cos(
|
||
last_token_rep, centroids, batch_labels, batch_size, args
|
||
)
|
||
|
||
loss = ot_loss
|
||
|
||
with torch.no_grad():
|
||
# 自训练阶段，原型更新用的是 soft label 版的 EMA
|
||
centroids = update_centroids_ema(
|
||
centroids, last_token_rep, batch_labels.half(), args
|
||
)
|
||
all_labels.append(batch_labels.cpu())
|
||
total += batch_labels.size(0)
|
||
|
||
scaler.scale(loss).backward()
|
||
scaler.step(optimizer)
|
||
scaler.update()
|
||
optimizer.zero_grad()
|
||
|
||
running_loss += loss.item() * batch_labels.size(0)
|
||
|
||
epoch_loss = running_loss / total
|
||
|
||
# ====== 用 test set 评估 AUROC ======
|
||
with torch.no_grad():
|
||
all_labels = torch.cat(all_labels).numpy()
|
||
test_labels_ = test_labels
|
||
|
||
if epoch % 1 == 0:
|
||
test_predictions, test_labels_combined = test_model(
|
||
model,
|
||
centroids,
|
||
test_prompts,
|
||
test_labels_,
|
||
device,
|
||
batch_size,
|
||
layer_number,
|
||
)
|
||
test_auroc = roc_auc_score(test_labels_combined, test_predictions)
|
||
|
||
print(f"Epoch [{epoch+1}/{num_epochs}], Loss: {epoch_loss:.4f}")
|
||
losses.append(epoch_loss)
|
||
|
||
if test_auroc > best_test_auroc:
|
||
best_test_auroc = test_auroc
|
||
best_test_epoch = epoch + args.init_num_epochs
|
||
print(
|
||
f"Best test AUROC: {best_test_auroc:.4f}, at epoch: {best_test_epoch}"
|
||
)
|
||
logging.info(
|
||
f"Best test AUROC: {best_test_auroc:.4f}, at epoch: {best_test_epoch}"
|
||
)
|
||
|
||
logging.info(
|
||
f"Epoch [{epoch+1}/{num_epochs}], Train Loss: {epoch_loss:.4f}, "
|
||
)
|
||
logging.info(
|
||
f"Best test AUROC: {best_test_auroc:.4f}, at epoch: {best_test_epoch}"
|
||
)
|
||
|
||
return best_test_auroc
|
||
|
||
|
||
|
||
def test_model(model, centroids, test_prompts, test_labels, device, batch_size, layer_number):
|
||
"""
|
||
在论文里相当于：
|
||
- 对 test 集算当前 TSV steer 后的 r_v
|
||
- 计算与两个原型 μ_c 的余弦相似度
|
||
- 用 softmax 得到属于 truthful 的概率 p(c=truthful | r_v)
|
||
- 用这个概率做 AUROC 评估
|
||
"""
|
||
model.eval()
|
||
val_predictions = [] # 保存 p_truthful
|
||
val_labels_combined = [] # 对应的 gt（二分类标签）
|
||
|
||
num_val_samples = len(test_prompts)
|
||
|
||
with torch.no_grad():
|
||
with autocast(dtype=torch.float16):
|
||
for batch_start in range(0, num_val_samples, batch_size):
|
||
batch_prompts = test_prompts[batch_start:batch_start + batch_size]
|
||
batch_labels = test_labels[batch_start:batch_start + batch_size]
|
||
batch_prompts, batch_labels = collate_fn(batch_prompts, batch_labels)
|
||
|
||
attention_mask = (batch_prompts != 0).half().to(device)
|
||
batch_prompts = batch_prompts.to(device)
|
||
batch_labels = batch_labels.to(device)
|
||
|
||
# 直接 forward + 拿最后一个非 padding token 的 hidden 表示作为 r_v
|
||
output = model(
|
||
batch_prompts.squeeze(),
|
||
attention_mask=attention_mask.squeeze(),
|
||
output_hidden_states=True,
|
||
)
|
||
hidden_states = output.hidden_states
|
||
hidden_states = torch.stack(hidden_states, dim=0).squeeze()
|
||
last_layer_hidden_state = hidden_states[layer_number]
|
||
last_token_rep = get_last_non_padded_token_rep(
|
||
last_layer_hidden_state, attention_mask.squeeze()
|
||
)
|
||
|
||
# 归一化到球面 & 原型也归一化
|
||
last_token_rep = F.normalize(last_token_rep, p=2, dim=-1)
|
||
centroids = F.normalize(centroids, p=2, dim=-1)
|
||
|
||
with autocast(dtype=torch.float16):
|
||
similarities = torch.matmul(last_token_rep, centroids.T) # [B, 2]
|
||
|
||
# 相似度 / 温度 -> softmax -> 取“真”的那一维作为概率
|
||
similarity_scores = torch.softmax(similarities / 0.1, dim=-1)
|
||
similarity_scores = similarity_scores[:, 1] # 假设 index 1 = truthful
|
||
|
||
val_predictions.append(similarity_scores.cpu())
|
||
val_labels_combined.append(batch_labels.cpu())
|
||
|
||
val_predictions = torch.cat(val_predictions)
|
||
val_labels_combined = torch.cat(val_labels_combined)
|
||
return val_predictions, val_labels_combined
|
||
|
||
|
||
|
||
HF_NAMES = {
|
||
'llama3.1-8B': 'meta-llama/Meta-Llama-3.1-8B',
|
||
'qwen2.5-7B': 'Qwen/Qwen2.5-7B' ,
|
||
'llama_7B': 'baffo32/decapoda-research-llama-7B-hf',
|
||
'honest_llama_7B': 'validation/results_dump/llama_7B_seed_42_top_48_heads_alpha_15',
|
||
'alpaca_7B': 'circulus/alpaca-7b',
|
||
'vicuna_7B': 'AlekseyKorshuk/vicuna-7b',
|
||
'llama2_chat_7B': 'models/Llama-2-7b-chat-hf',
|
||
'llama2_chat_13B': 'models/Llama-2-13b-chat-hf',
|
||
'llama2_chat_70B': 'meta-llama/Llama-2-70b-chat-hf',
|
||
}
|
||
|
||
def main():
|
||
|
||
# ======================
|
||
# 1. 解析命令行参数
|
||
# ======================
|
||
parser = argparse.ArgumentParser()
|
||
parser.add_argument('--model_name', type=str, default='alpaca_7B')
|
||
parser.add_argument('--num_gene', type=int, default=1) # 每个问题生成多少个答案
|
||
parser.add_argument('--train_sv', type=bool, default=True) # 是否执行“生成答案”阶段（1=生成+保存答案，0=不生成）
|
||
parser.add_argument('--steer_place', type=str, default='layer', help='where to steer (layer, mlp and head)') # 是否执行“生成答案”阶段（1=生成+保存答案，0=不生成）
|
||
parser.add_argument('--dataset_name', type=str, default='AdvBench') # 使用哪个数据集
|
||
parser.add_argument('--device', type=int, default=0) # GPU id（没怎么用，device_map="auto" 为主）
|
||
parser.add_argument("--model_dir", type=str, default=None, help='local directory with model data')
|
||
parser.add_argument("--batch_size", type=int, default=64)
|
||
parser.add_argument("--cos_temp", type=float, default=0.1) # 余弦相似度的 softmax 温度
|
||
parser.add_argument("--ema_decay", type=float, default=0.99) # 原型向量 EMA 衰减系数
|
||
parser.add_argument("--lr", type=float, default=0.005) # SV 的学习率
|
||
parser.add_argument("--str_layer", type=int, default=9) # 插 SV 的层号（第几层 hidden）
|
||
parser.add_argument("--component", type=str, default='res') # SV 以何种方式注入（残差等）
|
||
parser.add_argument("--lam", type=float, default=5) # SV 强度 λ
|
||
parser.add_argument("--init_num_epochs", type=int, default=20) # SV训练轮数
|
||
parser.add_argument("--optimizer", type=str, default='AdamW')
|
||
parser.add_argument('--train_ratio', type=float, default=0.8)
|
||
parser.add_argument('--val_ratio', type=float, default=0.1)
|
||
|
||
args = parser.parse_args()
|
||
|
||
# ======================
|
||
# 2. 确定 HuggingFace 模型名
|
||
# ======================
|
||
# HF_NAMES 在文件前面定义：
|
||
# HF_NAMES = {'llama3.1-8B': 'meta-llama/Meta-Llama-3.1-8B', 'qwen2.5-7B': 'Qwen/Qwen2.5-7B'}
|
||
# model_prefix 可以拼接一些前缀，这里默认是空
|
||
model_name_or_path = HF_NAMES[args.model_prefix + args.model_name]
|
||
|
||
# ======================
|
||
# 3. 加载不同的数据集
|
||
# ======================
|
||
if args.dataset_name == "AdvBench":
|
||
dataset = load_dataset("LeeWlving/Safety_Datasets", "advbench")['test']
|
||
elif args.dataset_name == 'HarmBench':
|
||
dataset = load_dataset("LeeWlving/Safety_Datasets", "harmbench")['standard']
|
||
elif args.dataset_name == 'JBB':
|
||
dataset = load_dataset("LeeWlving/Safety_Datasets", "JBB")['test']
|
||
else:
|
||
assert "Not supported dataset name!"
|
||
|
||
|
||
target_directory = f"save_for_eval/{args.dataset_name}/{args.model_name}_jailbreak"
|
||
benign_data, adverse_data = load_npy_shapes(target_directory, args.steer_place)
|
||
y_zero=np.zeros(len(benign_data))
|
||
y_one=np.ones(len(adverse_data))
|
||
|
||
data_embedding = np.concatenate((benign_data, adverse_data), axis=0)
|
||
gts = np.concatenate((y_zero, y_one), axis=0)
|
||
|
||
train_index, val_index, test_index=split_indices(len(data_embedding), args.train_ratio, args.val_ratio)
|
||
|
||
|
||
|
||
|
||
# ============================================================
|
||
# 6. 模式三：SV 训练阶段（gene=0 且 generate_gt=0）
|
||
# ============================================================
|
||
if args.train_sv:
|
||
|
||
device = torch.device("cuda")
|
||
# 加载基础 LLM（之后会在其上插 TSV）
|
||
model = AutoModelForCausalLM.from_pretrained(
|
||
model_name_or_path,
|
||
low_cpu_mem_usage=True,
|
||
torch_dtype=torch.float16,
|
||
device_map="auto",
|
||
token = ''
|
||
)
|
||
|
||
train_data = [data_embedding[i] for i in train_index]
|
||
train_labels = [gts[i] for i in train_index]
|
||
|
||
test_data = [data_embedding[i] for i in test_index]
|
||
test_labels = [gts[i] for i in test_index]
|
||
|
||
# ====== 6.4 冻结 LLM，只训练 SV 参数 ======
|
||
num_layers = model.config.num_hidden_layers
|
||
hidden_size = model.config.hidden_size
|
||
|
||
for param in model.parameters():
|
||
param.requires_grad = False
|
||
|
||
# 每一层一个 SV 向量（虽然实际只在某几层生效）
|
||
sv = nn.ParameterList(
|
||
[nn.Parameter(torch.zeros(hidden_size), requires_grad=True) for _ in range(num_layers)]
|
||
)
|
||
sv.to(device)
|
||
|
||
# 在模型对应层里插入 SV（add_sv_layers 会改 forward）
|
||
add_sv_layers(model, sv, [args.lam], args)
|
||
|
||
# 只把 SV 的参数丢给优化器
|
||
optimizer = torch.optim.AdamW(list(sv.parameters()), lr=args.lr)
|
||
|
||
# ====== 6.5 调用 train_model，进入两阶段训练流程 ======
|
||
train_model(model, optimizer, device, train_data, train_labels, args=args)
|
||
|
||
else:
|
||
print("Skip training steer vectors.")
|
||
raise NotImplementedError("Only training SV is implemented in this script.")
|
||
|
||
|
||
if __name__ == '__main__':
|
||
# 为了结果可复现，固定随机种子
|
||
seed_everything(42)
|
||
main()
|