Ubiquant_CharacterHunter / post_train.py
Facepalm0's picture
Upload post_train.py with huggingface_hub
b034760 verified
raw
history blame
8.06 kB
import torch
from torch.utils.data import DataLoader
from models.resnet import resnet18, resnet34, resnet50
from models.openmax import OpenMax
from models.metamax import MetaMax
from train import GameDataset
from utils.eval_utils import evaluate_openmax, evaluate_metamax
from torchvision import transforms
from utils.data_stats import load_dataset_stats
from pprint import pprint
def prepare_data_and_model(model_path='models/best_model.pth', model_type='resnet18', batch_size=400):
"""准备数据和模型"""
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
# 加载数据集统计信息和准备数据
mean, std = load_dataset_stats()
transform = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize(mean=mean, std=std)
])
# 加载训练集和验证集
train_dataset = GameDataset('jk_zfls/round0_train', num_labels=20, transform=transform)
val_dataset = GameDataset('jk_zfls/round0_eval', num_labels=21, transform=transform)
train_loader = DataLoader(train_dataset, batch_size=batch_size, shuffle=False,
num_workers=4, pin_memory=True)
val_loader = DataLoader(val_dataset, batch_size=batch_size, shuffle=False,
num_workers=4, pin_memory=True)
# 加载预训练模型
if model_type == 'resnet18':
model = resnet18(num_classes=20)
elif model_type == 'resnet34':
model = resnet34(num_classes=20)
elif model_type == 'resnet50':
model = resnet50(num_classes=20)
checkpoint = torch.load(model_path)
model.load_state_dict(checkpoint['model_state_dict'])
model = model.to(device)
model.eval()
return model, train_loader, val_loader, device
def collect_features(model, loader, device, return_logits=False):
"""收集特征和标签"""
features_list = []
logits_list = []
labels_list = []
with torch.no_grad():
for images, labels, paths in loader:
images = images.to(device)
if return_logits:
logits, features = model(images, return_features=True)
logits_list.append(logits.cpu())
else:
_, features = model(images, return_features=True)
features_list.append(features.cpu())
labels_list.append(labels)
if return_logits:
return torch.cat(features_list), torch.cat(logits_list), torch.cat(labels_list)
else:
return torch.cat(features_list), torch.cat(labels_list)
def train_openmax(features,labels, model, val_loader, device, fraction=0.2):
"""训练和评估OpenMax模型
fraction: 未知类别比例
"""
# OpenMax特定的超参数搜索空间
# alpha_range = [3, 5, 8, 12, 16, 20]
alpha_range = [12, 13, 14, 15, 16, 17, 18, 19, 20]
# tailsize_range = [10, 15, 20, 25, 30]
tailsize_range = [5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15]
# multiplier_range = [0.5, 0.75, 1, 1.25, 1.5]
multiplier_range = [0.5, 0.6, 0.7, 0.8, 0.9, 1.0]
best_params = {
'alpha': None,
'tailsize': None,
'multiplier': None,
'accuracy': .0,
'model': None
}
val_features, val_logits, val_labels = collect_features(model, val_loader, device, return_logits=True)
print("\n=== Training OpenMax ===")
for alpha in alpha_range:
for tailsize in tailsize_range:
print(f"\nTraining OpenMax with alpha={alpha}, tailsize={tailsize}")
openmax = OpenMax(num_classes=20, tailsize=tailsize, alpha=alpha)
openmax.fit(features, labels)
print(f"Training finished, evaluating...")
for multiplier in multiplier_range:
overall_acc, known_acc, unknown_acc = evaluate_openmax(
openmax, val_features, val_logits, val_labels, multiplier=multiplier, fraction=fraction, verbose=False
)
if overall_acc > best_params['accuracy']:
best_params.update({
'alpha': alpha,
'tailsize': tailsize,
'multiplier': multiplier,
'accuracy': overall_acc,
'model': openmax
})
print(f"\nNew best OpenMax parameters found:")
print(f"Alpha: {alpha}")
print(f"Tailsize: {tailsize}")
print(f"Multiplier: {multiplier}")
print(f"Overall Accuracy: {overall_acc:.2f}%")
print(f"Known Classes Accuracy: {known_acc:.2f}%")
print(f"Unknown Class Accuracy: {unknown_acc:.2f}%")
elif overall_acc > 95.0:
print(f"Alpha: {alpha}")
print(f"Tailsize: {tailsize}")
print(f"Multiplier: {multiplier}")
print(f"Overall Accuracy: {overall_acc:.2f}%")
print(f"Known Classes Accuracy: {known_acc:.2f}%")
print(f"Unknown Class Accuracy: {unknown_acc:.2f}%")
return best_params
def train_metamax(features, labels, model, val_loader, device):
"""训练和评估MetaMax模型"""
# MetaMax特定的超参数搜索空间
meta_ratio_range = [0.05, 0.1, 0.15, 0.2, 0.25]
threshold_range = [0.1, 0.2, 0.3, 0.4, 0.5]
best_params = {
'meta_ratio': None,
'threshold': None,
'accuracy': .0,
'model': None
}
print("\n=== Training MetaMax ===")
for meta_ratio in meta_ratio_range:
print(f"\nTesting MetaMax with meta_ratio={meta_ratio}")
metamax = MetaMax(num_classes=20, meta_ratio=meta_ratio)
metamax.fit(features, labels)
for threshold in threshold_range:
overall_acc, known_acc, unknown_acc = evaluate_metamax(
metamax, model, val_loader, device, threshold=threshold, verbose=False
)
if overall_acc > best_params['accuracy']:
best_params.update({
'meta_ratio': meta_ratio,
'threshold': threshold,
'accuracy': overall_acc,
'model': metamax
})
if overall_acc > 90.0:
print(f"\nNew best MetaMax parameters found:")
print(f"Meta Ratio: {meta_ratio}")
print(f"Threshold: {threshold}")
print(f"Overall Accuracy: {overall_acc:.2f}%")
print(f"Known Classes Accuracy: {known_acc:.2f}%")
print(f"Unknown Class Accuracy: {unknown_acc:.2f}%")
return best_params
if __name__ == '__main__':
# 准备数据和模型
model, train_loader, val_loader, device = prepare_data_and_model(model_path='models/resnet50_99.92.pth', model_type='resnet50', batch_size=128)
# 收集特征
features, labels = collect_features(model, train_loader, device, return_logits=False)
# 训练OpenMax
best_openmax_params = train_openmax(features, labels, model, val_loader, device)
print("\nSaving OpenMax model...")
pprint(best_openmax_params)
torch.save(best_openmax_params['model'], f'models/resnet50_openmax_{best_openmax_params["accuracy"]:.2f}.pth')
print(f"OpenMax model saved to models/resnet50_openmax_{best_openmax_params['accuracy']:.2f}.pth")
# 训练MetaMax
# best_metamax_params = train_metamax(features, labels, model, val_loader, device)
# print("\nSaving MetaMax model...")
# pprint(best_metamax_params)
# torch.save(best_metamax_params['model'], 'models/best_metamax.pth')
# print(f"MetaMax model saved to models/best_metamax.pth")