def train_net(train_set, valid_set, test_set, in_dim, out_dim, cfg):
# 在完整数据集上训练模型
train_loader, valid_loader, test_loader = make_loader(
train_set,
valid_set,
test_set,
seq_len=cfg['seq_len'],
num_shift=cfg['num_shift'],
bt_sz=cfg['bt_sz'])
# 创建模型实例
net = RNN_Net(in_dim,
cfg['hidden_dim'],
out_dim,
batchsize=cfg['bt_sz'],
rnn_type=cfg['rnn_type'],
num_layers=cfg['num_layers'],
dropout=cfg['dropout'],
bidirectional=cfg['bidirectional'])
# 优化器定义
# opt = optim.RMSprop(net.parameters(), lr=cfg['lr_rate'], weight_decay=cfg['weight_decay'])
# opt = AdaBoundW(net.parameters(), lr=cfg['lr_rate'], weight_decay=cfg['weight_decay'])
opt = AdaBound(net.parameters(),
lr=cfg['lr_rate'],
weight_decay=cfg['weight_decay'])
# 损失定义
criterion = nn.MSELoss()
# 封装
model = Modeler(net, opt, criterion, set_device())
# 训练
for epoch in range(cfg['num_epoch']):
train_loss = model.train_model(train_loader) # 当前 epoch 的训练损失
valid_loss = model.evaluate_model(valid_loader) # 当前 epoch 的验证损失
print(
f"[{epoch + 1}/{cfg['num_epoch']}] ===>> train_loss: {train_loss: .4f} | valid_loss: {valid_loss: .4f}"
)
# 训练结束 预测
prediction, err = model.predit_point_by_point(test_loader.dataset.data,
test_loader.dataset.target)
# 可视化预测效果
# for ch in range(prediction.shape[-1]):
# plt.figure(figsize=(12, 5))
# plt.plot(np.c_[prediction.cpu().numpy()[:, ch], test_loader.dataset.target.cpu().numpy()[:, ch]])
# plt.legend([f'prediction channel{ch + 1}', f'label channel{ch + 1}'])
# if cfg['vis']:
# plt.show()
return prediction, err
seqdata_all = get_mat_data(f'dev_seq_data/{signal_type}.mat',
f'{signal_type}') # 读取数据
# 在完整数据集上训练模型
train_loader, valid_loader, test_loader = make_loader(seqdata_all,
tt_split=0.7,
tv_split=0.8,
seq_len=seq_length,
bt_sz=32)
loaders = {'train': train_loader, 'valid': valid_loader}
err_all = train_valid(2, 15, 2, test_loader.dataset.get_tensor_data(),
loaders)
# 基本设置
device = set_device()
set1 = {
'longlag_nonlinear_signals': [3, 5, 8, 10, 12, 15, 18, 20],
'nonlinear_signals': [3, 5, 8, 10, 13, 15, 18, 20]
}
all_signal_type = ['nonlinear_signals', 'longlag_nonlinear_signals']
time_spend = {'longlag_nonlinear_signals': [], 'nonlinear_signals': []}
for signal_type in all_signal_type:
for seq_len in set1[signal_type]:
print(f'{signal_type}, seq length: {seq_len}')
timer = Timer()
timer.start()
main(seq_len)
# 计时结束
b = timer.stop()
