lambdaji

http://labs.criteo.com/2014/02/kaggle-display-advertising-challenge-dataset/

wide&deep直接调的TensorFlow Pre-Made Estimator，没走libsvm

确实，浪费一个feat_id的内存

http://labs.criteo.com/2014/02/kaggle-display-advertising-challenge-dataset/

worker-1是'task': {'type': 'evaluator', 'index': 0}}，ckpt多久保存一次？

参考run_dist.sh

需要自己划分，然后代码里根据index决定读取哪一份 ps：把现在代码里glob那一行改了

是的，没做

http://labs.criteo.com/2014/02/download-kaggle-display-advertising-challenge-dataset/

python2.7 tensorflow1.4