Created by: backyes

settings(
    #learning_rate_decay_a = 1e-05,
    #learning_rate_decay_b = 0.0,
    learning_rate = 0.1,
    batch_size = batch_size,
    algorithm = 'sgd',
        #l1weight=0.1,
    #num_batches_per_send_parameter = 1,
    learning_method = 'rmsprop',
)

Generally, l1weigh is specified for owlqn, but with setting above, l1weigh indeed has influence on convergence .

With l1weight=0.1

I /home/wangyanfei/baidu/idl/paddle/paddle/trainer/Tester.cpp:111]  Test samples=10877807 cost=0.0634097 Eval: __auc_evaluator_0__=0.641067
..............I /home/wangyanfei/baidu/idl/paddle/paddle/trainer/TrainerInternal.cpp:179]  Pass=0 Batch=1094 samples=43721520 AvgCost=0.0665462 Eval: __auc_evaluator_0__=0.620773
I /home/wangyanfei/baidu/idl/paddle/paddle/trainer/Tester.cpp:111]  Test samples=10877807 cost=0.0635014 Eval: __auc_evaluator_0__=0.634032

With l1weigh=0.1 is removed:

I /home/wangyanfei/baidu/idl/paddle/paddle/trainer/Tester.cpp:111]  Test samples=10877807 cost=0.0629021 Eval: __auc_evaluator_0__=0.65267
..............I /home/wangyanfei/baidu/idl/paddle/paddle/trainer/TrainerInternal.cpp:179]  Pass=0 Batch=1094 samples=43721520 AvgCost=0.0658823 Eval: __auc_evaluator_0__=0.619865
I /home/wangyanfei/baidu/idl/paddle/paddle/trainer/Tester.cpp:111]  Test samples=10877807 cost=0.0640271 Eval: __auc_evaluator_0__=0.600817

need more TEST to verify this conclusion...