caffe solver 配置详解
caffe solver通过协调网络前向推理和反向梯度传播来进行模型优化,并通过权重参数更新来改善网络损失求解最优算法,而solver学习的任务被划分为:监督优化和参数更新,生成损失并计算梯度。caffe solver是caffe中的核心,它定义着整个模型如何运转,不管是命令行方式还是pycaffe接口方式进行网络训练或测试,都是需要一个solver配置文件的,而solver的配置参数总共有42个,罗列如下: net weight_decay net_param regularization_type train_net stepsize test_net stepvalue train_net_param clip_gradients test_net_param snapshot train_state snapshot_prefix test_state snapshot_diff test_iter snapshot_format test_interval solver_mode test_compute_loss device_id test_initialization random_seed base_lr type display delta average_loss momentum2 max_iter rms_decay iter_size debug