def data_process(self):
normal = transforms.Normalize([0.485, 0.456, 0.406],
[0.229, 0.224, 0.225])
if cfg.use_autoaugment:
train_transform = transforms.Compose([
transforms.Resize(cfg.image_size),
transforms.CenterCrop(cfg.crop_image_size),
# transforms.RandomRotation(cfg.rotate_degree),
auto_augment.AutoAugment(dataset=cfg.autoaugment_dataset),
transforms.ToTensor(),
transforms.RandomErasing(p=0.3,
scale=(0.02, 0.1),
ratio=(0.6, 1.2)),
normal
])
else:
train_transform = transforms.Compose([
transforms.Resize(cfg.image_size),
transforms.CenterCrop(cfg.crop_image_size),
transforms.RandomRotation(cfg.rotate_degree),
transforms.ToTensor(),
transforms.RandomErasing(p=0.3,
scale=(0.02, 0.1),
ratio=(0.6, 1.2)), normal
])
test_trainsform = transforms.Compose([
transforms.Resize(cfg.image_size),
transforms.CenterCrop(cfg.crop_image_size),
transforms.ToTensor(), normal
])
return train_transform, test_trainsform
def predict(self, image_path):
'''
模型预测返回结果
:param input: 评估传入样例 {"image_path":".\/data\/input\/BeijingGarbage\/image\/0.png"}
:return: 模型预测成功之后返回给系统样例 {"label":"0"}
'''
# print(image_path)
img = Image.open(image_path).convert('RGB')
test_trainsform = transforms.Compose([
transforms.Resize(cfg.image_size),
transforms.CenterCrop(cfg.crop_image_size),
auto_augment.AutoAugment(dataset=cfg.autoaugment_dataset),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])
# output = 0
r1, r2, r3, r4, r5 = 0, 0, 0, 0, 0
for _ in range(cfg.tta_times):
tensor = test_trainsform(img)
tensor = torch.unsqueeze(tensor, dim=0).float()
v1, v2, v3, v4, v5 = self.model1(tensor.to(device))
t1, t2, t3, t4, t5 = self.model2(tensor.to(device))
# v1, v2, v3, v4, v5 = self.model1(tensor.to(device))
r1 += v1
r2 += v2
r3 += v3
r4 += v4
r5 += v5
r1 += t1
r2 += t2
r3 += t3
r4 += t4
r5 += t5
val_pred1 = torch.max(r1, 1)[1]
val_pred2 = torch.max(r2, 1)[1]
val_pred3 = torch.max(r3, 1)[1]
val_pred4 = torch.max(r4, 1)[1]
val_pred5 = torch.max(r5, 1)[1]
val_pred = torch.stack([val_pred1, val_pred2, val_pred3, val_pred4, val_pred5], 1).squeeze()
pred = list(val_pred.cpu().numpy())
pred = [i for i in pred if i != 10]
temp = ''
for v in pred:
temp += str(v)
return int(temp)
def __init__(self, args, mode='train'):
self.args = args
root_dir = args.dataset
batch_size = args.batch_size
augment = args.augment
img_size = args.img_size
self.resize_size = (int(img_size / 0.875), int(img_size / 0.875))
self.crop_size = (img_size, img_size)
self.root_dir = os.path.join(root_dir, mode)
self.img_path_list = []
self.label_list = []
self.class_names = []
for class_index, class_name in enumerate(os.listdir(self.root_dir)):
class_dir = os.path.join(self.root_dir, class_name)
for img_name in os.listdir(class_dir):
self.img_path_list.append(os.path.join(class_dir, img_name))
self.label_list.append(class_index)
self.class_names.append(class_name)
if mode == 'train':
pad_len = len(self.img_path_list) % batch_size
img_path_list_len = len(self.img_path_list)
for _ in range(pad_len):
rand_index = np.random.randint(0, img_path_list_len)
self.img_path_list.append(self.img_path_list[rand_index])
self.label_list.append(self.label_list[rand_index])
self.data_index = np.arange(0, len(self.label_list))
np.random.shuffle(self.data_index)
else:
self.data_index = np.arange(0, len(self.label_list))
self.img_path_list = np.array(self.img_path_list)
self.label_list = np.array(self.label_list)
self.augment = augment
self.mode = mode
self.batch_size = batch_size
self.eppch_index = 0
self.num_class = len(self.class_names)
self.random_crop_transform = A.Compose([
A.RandomCrop(width=self.crop_size[0], height=self.crop_size[1]),
])
self.center_crop_transform = A.Compose([
A.CenterCrop(width=self.crop_size[0], height=self.crop_size[1]),
])
self.auto_augment = auto_augment.AutoAugment()
self.rand_augment = rand_augment.RandAugment(magnitude=10.)
f"/content/tianchiorgame/df_WeatherClassify/output/{config.prefix}_train.csv",
index=None)
val_df.to_csv(
f"/content/tianchiorgame/df_WeatherClassify/output/{config.prefix}_val.csv",
index=None)
print(f"train length: {len(train_df)}, val length: {len(val_df)}")
writer = SummaryWriter(log_dir="output/run/",
filename_suffix=config.prefix)
# 数据加载
train_augmentation = transforms.Compose([
transforms.Resize(config.IMG_SIZE + 32),
transforms.RandomResizedCrop(config.IMG_SIZE),
transforms.RandomHorizontalFlip(),
auto_augment.AutoAugment(),
m_transform.RandomErasing(0.5, [0.02, 0.4], 0.3, 20),
transforms.ToTensor(),
transforms.Normalize(config.mean, config.std)
])
val_augmentation = transforms.Compose([
transforms.Resize(config.IMG_SIZE),
transforms.CenterCrop(config.IMG_SIZE),
transforms.ToTensor(),
transforms.Normalize(config.mean, config.std)
])
train_dataset = m_dataloader.IMetDataset(
train_df,
os.path.join("/content/drive/My Drive/train"),
n_classes=config.NUM_CLASSES,
transforms=train_augmentation)
#训练中记录三个指标：验证集的最佳acc和对应的loss，验证集上的最低loss
metrics = {'best_acc': 10, 'best_acc_loss': 100, 'best_loss': 100}
train_transform = transforms.Compose([
transforms.Resize((size + 32, size + 32)),
transforms.RandomChoice([
transforms.RandomCrop(size,
padding=1,
pad_if_needed=True,
padding_mode='edge'),
transforms.RandomResizedCrop(size,
scale=(resize_scale, 1.0),
ratio=(0.8, 1.2))
]),
transforms.RandomHorizontalFlip(),
auto_augment.AutoAugment(dataset='CIFAR'), #auto_augment
transforms.ToTensor(),
transforms.RandomErasing(p=erasing_prob),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])
val_transform = transforms.Compose([
transforms.Resize((size, size)),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])
for fold_idx, (train_idx, val_idx) in enumerate(
skf.split(train_df['filename'], train_df['label'])):
print('*******************fold {0}*****************'.format(fold_idx))
train_data = train_df.iloc[train_idx]
val_data = train_df.iloc[val_idx]