File size: 7,424 Bytes
485127c | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 | # -*- coding: utf-8 -*-
from torch.utils.data import DataLoader
import tqdm
from torch.cuda.amp import GradScaler, autocast
import torch.nn.functional as F
from torch import nn
import torch
import numpy as np
from torch.utils.tensorboard import SummaryWriter
import datetime
import os
import json
from metrics import get_roc_metrics, get_precision_recall_metrics
from torch.optim.lr_scheduler import CosineAnnealingLR
import time
from utils import GpuMem
try:
from transformers import AdamW
except:
from torch.optim import AdamW
def evaluate_model_SPO(model, data, DEVICE):
model.to(DEVICE)
model.eval()
loss = 0
eval_loader = DataLoader(data, batch_size=1, shuffle=False)
epoch_crit_train_original, epoch_crit_train_sampled = [],[]
start_time = time.time()
with torch.no_grad():
for batch in tqdm.tqdm(eval_loader, desc="Evaluating"):
text = batch
output = model(text)
loss += output['loss'].item()
epoch_crit_train_original.extend(output['crit'][1].tolist())
epoch_crit_train_sampled.extend(output['crit'][3].tolist())
print(f"Total time: {time.time() - start_time:.4f}s")
avg_loss = loss / len(eval_loader)
fpr, tpr, roc_auc = get_roc_metrics(epoch_crit_train_original, epoch_crit_train_sampled)
p, r, pr_auc = get_precision_recall_metrics(epoch_crit_train_original, epoch_crit_train_sampled)
# print(f"val_loss: {avg_loss:.6f}")
print(f"val_ROC_AUC: {roc_auc:.4f}, PR AUC: {pr_auc:.4f}")
print(f"val_Real_mean/std: {np.mean(epoch_crit_train_original):.2f}/{np.std(epoch_crit_train_original):.2f}, val_Samples_mean/std: {np.mean(epoch_crit_train_sampled):.2f}/{np.std(epoch_crit_train_sampled):.2f}")
print("="*10)
results_dict = {
"name": "imbd",
'info': {'n_samples': len(epoch_crit_train_original)},
'predictions': {'real': epoch_crit_train_original,
'samples': epoch_crit_train_sampled},
'metrics': {'roc_auc': roc_auc, 'fpr': fpr, 'tpr': tpr},
'pr_metrics': {'pr_auc': pr_auc, 'precision': p, 'recall': r},
}
return results_dict
def fine_tune_ours(model, data, DEVICE, ckpt_dir='./ckpt', args=None):
current_time = datetime.datetime.now().strftime("%Y-%m-%d_%H-%M-%S")
writer = SummaryWriter(log_dir=f"./scripts/ImBD/logs/{args.task_name}_spo_lr_{args.lr}_beta_{args.beta}_a_{args.a}_{current_time}/train_ai_detection")
train_loader = DataLoader(data[0], batch_size=1, shuffle=True)
epochs = args.epochs
optimizer = AdamW(model.parameters(), lr=args.lr)
scheduler = CosineAnnealingLR(optimizer, T_max=len(train_loader) * epochs, eta_min=0,
last_epoch=-1)
scaler = GradScaler()
model.to(DEVICE)
# Number of iterations for gradient accumulation
accumulation_steps = args.a
epoch_losses, i, loss = [], 0, torch.tensor(0.0).to(DEVICE)
epoch_crit_train_original, epoch_crit_train_sampled = [],[]
start_time = time.time()
for epoch in range(epochs):
optimizer.zero_grad()
start_time = time.time()
for batch in tqdm.tqdm(train_loader, desc=f"Fine-tuning: {epoch} epoch"):
text = batch
scheduler.step()
with autocast():
outputs_1 = model(text)
epoch_crit_train_original.extend(outputs_1['crit'][1].tolist())
epoch_crit_train_sampled.extend(outputs_1['crit'][3].tolist())
loss += (outputs_1['loss'].to(torch.float32)) / accumulation_steps
if ((i + 1) % accumulation_steps) == 0:
scaler.scale(loss).backward()
scaler.step(optimizer)
optimizer.zero_grad()
scaler.update()
writer.add_scalar('Loss/train', loss.item(), i)
epoch_losses.append(loss.item())
loss = torch.tensor(0.0).to(DEVICE)
epoch_losses.append(loss.item())
i += 1
print(f"Total time: {time.time() - start_time:.4f}s")
fpr, tpr, roc_auc = get_roc_metrics(epoch_crit_train_original, epoch_crit_train_sampled)
p, r, pr_auc = get_precision_recall_metrics(epoch_crit_train_original, epoch_crit_train_sampled)
print(f"ROC AUC: {roc_auc:.4f}, PR AUC: {pr_auc:.4f}")
print(f"Real mean/std: {np.mean(epoch_crit_train_original):.2f}/{np.std(epoch_crit_train_original):.2f}, Samples mean/std: {np.mean(epoch_crit_train_sampled):.2f}/{np.std(epoch_crit_train_sampled):.2f}")
epoch_avg_loss = np.mean(epoch_losses)
writer.add_scalar('Loss/epoch', epoch_avg_loss, epoch)
writer.add_scalar('ROC_AUC/epoch', roc_auc, epoch)
writer.add_scalar('PR_AUC/epoch', pr_auc, epoch)
writer.add_scalar('Real_mean/epoch',np.mean(epoch_crit_train_original),epoch)
writer.add_scalar('Real_std/epoch',np.std(epoch_crit_train_original),epoch)
writer.add_scalar('Sampled_mean/epoch',np.mean(epoch_crit_train_sampled),epoch)
writer.add_scalar('Sampled_std/epoch',np.std(epoch_crit_train_sampled),epoch)
epoch_crit_train_original, epoch_crit_train_sampled = [],[] # reset crit
print(f"\nAverage Loss for Epoch {epoch}: {epoch_avg_loss}")
# if not os.path.exists(ckpt_dir):
# os.makedirs(ckpt_dir)
# model.save_pretrained(ckpt_dir)
# print(f"Saved finetuned model to {os.path.join(ckpt_dir, 'ours-finetuned.pth')}")
writer.close()
return model
def run(
model,
data,
DEVICE,
args,
ckpt_dir='./ckpt',
):
if args.ebt or args.eval_only:
print("Evaluating model before tuning...")
d = evaluate_model_SPO(model, data[1], DEVICE)
if args.SPOtrained:
output_path = f"{args.output_file}.imbd.json"
else:
method_name=args.base_model.split("_")[-1]
output_path = f"{args.output_file}.{method_name}.json"
with open(output_path, "w") as j:
json.dump(d,j)
print(f"Results saved to {output_path}.")
if args.eval_only:
return
tracker = GpuMem()
print('Fine-tuning model...')
start = time.perf_counter()
with tracker:
model = fine_tune_ours(
model,
data,
DEVICE=DEVICE,
ckpt_dir=ckpt_dir,
args=args
)
pre_time = time.perf_counter() - start
pre_memory = tracker.memory_usage()
if args.eval_after_train:
print("Evaluating model after tuning...")
start = time.perf_counter()
with tracker:
d = evaluate_model_SPO(model, data[1], DEVICE)
eval_time = time.perf_counter() - start
eval_time = eval_time / (len(data[1]) << 1)
eval_memory = tracker.memory_usage()
d['compute_info'] = {'pre_time': pre_time, 'eval_time': eval_time,
'pre_memory': pre_memory, 'eval_memory': eval_memory,}
if args.SPOtrained:
output_path = f"{args.output_file}.imbd.json"
else:
method_name=args.base_model.split("_")[-1]
output_path = f"{args.output_file}.{method_name}.json"
with open(output_path, "w") as j:
json.dump(d, j)
print(f"Results saved to {output_path}.")
|