Created by: zhaoyang1708

损失函数之前为cross_entropy，训练正常，更改为dice_loss后，本地2000条ins是可以过拟合的，但集群训练一直为0.5 sim_ctr = fluid.layers.cos_sim(user_fc_ctr, ad_fc_ctr) scale_ctr = fluid.layers.scale(x=sim_ctr, scale=5.0) predict_ctr = fluid.layers.fc(name="output2", input=scale_ctr, size=CLASS_SIZE, act="softmax") predict_ctr_one = fluid.layers.slice(predict_ctr, axes=[1],starts=[1], ends=[2]) binary_predict_ctr = predict_ctr auc_var, cur_auc, auc_states = fluid.layers.auc(input=binary_predict_ctr, label=label, slide_steps = 20) cost = fluid.layers.dice_loss(input=binary_predict_ctr, label=label) avg_cost = fluid.layers.reduce_sum(cost) avg_cost = fluid.layers.scale(x=avg_cost, scale=int(os.environ['CPU_NUM'])) 任务链接http://10.182.12.37:8910/fileview.html?type=logsdir&path=/&instance=0.app-user-20200413214256-39482--zhaoyang29_online_novelbsq_dice_paddlecloud