一个单机多卡训练模型的例子

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

186

187

188

189

# """My demo train script."""
 
import argparse
import logging
import os
import random
import time
import numpy as np
import torch
 
from torch import nn, optim, Tensor
from torch.utils.data import DataLoader, Dataset
 
 
def parse_args() -> argparse.Namespace:
    """Parse arguments."""
    parser = argparse.ArgumentParser(description="Training")
    parser.add_argument("--seed", type=int, help="Fix random seed", default=123)
    parser.add_argument(
        "--log_file", type=str, help="Log file", default="test_train.log"
    )
    parser.add_argument(
        "--log_path", type=str, help="Model path", default="./training_log/"
    )
    parser.add_argument(
        "--train_epochs", type=int, help="Epochs of training", default=5
    )
    parser.add_argument("--batch_size", type=int, help="Batch size", default=32)
    parser.add_argument(
        "--learning_rate",
        type=float,
        help="Learning rate",
        default=1e-3,
    )
    parser.add_argument("--device", type=str, help="Run on which device", default="cpu")
    parser.add_argument(
        "--cuda_visible_devices", type=str, help="Cuda visible devices", default="0"
    )
    return parser.parse_args()
 
 
def init_logging(log_file: str, level: str = "INFO") -> None:
    """Initialize logging."""
    logging.basicConfig(
        filename=log_file,
        filemode="w",
        format="%(asctime)s - %(name)s - %(levelname)s - %(message)s",
        level=level,
    )
    logging.getLogger().addHandler(logging.StreamHandler())
 
 
def set_seed(seed: int) -> None:
    """Set seed for reproducibility."""
    os.environ["PYTHONHASHSEED"] = str(seed)
    os.environ["CUBLAS_WORKSPACE_CONFIG"] = ":4096:8"
    random.seed(seed)
    np.random.seed(seed)
    torch.manual_seed(seed)
    torch.cuda.manual_seed_all(seed)
    torch.backends.cudnn.benchmark = False
    torch.backends.cudnn.deterministic = True
    torch.use_deterministic_algorithms(True)
 
 
def seed_worker(work_id: int) -> None:
    """Set seed for worker."""
    np.random.seed(work_id)
    random.seed(work_id)
 
 
class DatasetClass(Dataset):
    """My demo dataset class."""
 
    def __init__(self):
        self.input = np.random.rand(1000000, 2).astype(np.float32)
        # self.input[:, 1] = 0.0
        self.target = np.zeros([1000000, 1])
        self.target[:, 0] = self.input[:, 0] + 1.0
 
    def __len__(self):
        return len(self.input)
 
    def __getitem__(self, idx: int) -> tuple:
        return self.input[idx], self.target[idx]
 
 
class ModelClass(torch.nn.Module):
    """My demo model class."""
 
    def __init__(self):
        super().__init__()
        self.my_layer = nn.Linear(2, 1)
 
    def forward(self, inputs: Tensor) -> Tensor:
        """My demo forward function."""
        outputs = self.my_layer(inputs)
        return outputs
 
 
def get_loss(model_output: Tensor, target: Tensor) -> Tensor:
    """My demo loss function."""
    loss = torch.norm(model_output - target, dim=-1).sum()
    return loss
 
 
def training() -> None:
    """My demo training function."""
    train_set = DatasetClass()
    g = torch.Generator()
    g.manual_seed(args.seed)
    train_loader = DataLoader(
        dataset=train_set,
        batch_size=args.batch_size,
        shuffle=True,
        num_workers=os.cpu_count(),
        pin_memory=True,
        worker_init_fn=seed_worker,
        generator=g,
    )
    model = ModelClass()
    if args.device == "cuda":
        model = nn.DataParallel(model)
    model.to(args.device)
    optimizer = optim.AdamW(model.parameters(), lr=args.learning_rate)
    for epoch in range(args.train_epochs):
        model.train()
        for batch_index, (features, labels) in enumerate(train_loader):
            features = features.to(args.device)
            labels = labels.to(args.device)
            model_outputs = model(features)
            optimizer.zero_grad(set_to_none=True)
            loss = get_loss(model_outputs, labels)
            loss.backward()
            optimizer.step()
            if batch_index % 1000 == 0:
                logging.info(
                    "Epoch: %s, Batch index: %s, Loss: %s",
                    epoch,
                    batch_index,
                    loss.item(),
                )
    torch.save(model.state_dict(), f"{args.log_path}/trained_model.pth")
 
 
def testing() -> None:
    """My demo testing function."""
    test_set = DatasetClass()
    g = torch.Generator()
    g.manual_seed(args.seed)
    test_loader = DataLoader(
        dataset=test_set,
        batch_size=args.batch_size,
        shuffle=True,
        num_workers=os.cpu_count(),
        pin_memory=True,
        worker_init_fn=seed_worker,
        generator=g,
    )
    model = ModelClass()
    if args.device == "cuda":
        model = nn.DataParallel(model)
    model.load_state_dict(torch.load(f"{args.log_path}/trained_model.pth"))
    model.to(args.device)
    model.eval()
    with torch.no_grad():
        for batch_index, (features, labels) in enumerate(test_loader):
            features = features.to(args.device)
            labels = labels.to(args.device)
            model_outputs = model(features)
            loss = get_loss(model_outputs, labels)
            if batch_index % 1000 == 0:
                logging.info(
                    "Batch index: %s, Loss: %s",
                    batch_index,
                    loss.item() / args.batch_size,
                )
 
 
if __name__ == "__main__":
    args = parse_args()
    set_seed(args.seed)
    init_logging(args.log_file)
    os.environ["CUDA_VISIBLE_DEVICES"] = args.cuda_visible_devices
    main_start_time = time.time()
    training()
    main_end_time = time.time()
    logging.info("Main time: %s", main_end_time - main_start_time)
    testing()