关于DCN

[CXY573]

首先，感谢分享，原文有一句话是We propose the DCN model that enables Web-scale automatic feature learning with both sparse and dense inputs 是可以把稀疏数据作为训练数据的，请问你的分享是否有这样的接口呢。

[CXY573]

[1, 100] loss: 583.635897 metric: 0.591613 time: 15.8 s

---

[1] loss: 22.228953 metric: 0.616724 time: 19.0 s

---

---

[1] loss: 22.255722 metric: 0.614940 time: 20.2 s

---

[2, 100] loss: 88.273439 metric: 0.584751 time: 14.7 s

---

[2] loss: 29.329584 metric: 0.622159 time: 17.9 s

---

---

[2] loss: 29.419680 metric: 0.619503 time: 19.1 s

---

[3, 100] loss: 45.896620 metric: 0.574566 time: 13.4 s

---

[3] loss: 34.074569 metric: 0.587830 time: 16.7 s

---

---

[3] loss: 34.868626 metric: 0.584307 time: 17.8 s

---

[4, 100] loss: 35.304552 metric: 0.573835 time: 14.0 s

---

[4] loss: 44.318500 metric: 0.564652 time: 16.6 s

---

---

[4] loss: 44.822131 metric: 0.561964 time: 17.8 s

---

[5, 100] loss: 28.368226 metric: 0.526387 time: 13.9 s

---

[5] loss: 74.812706 metric: 0.538661 time: 16.9 s

---

---

[5] loss: 75.696912 metric: 0.537339 time: 18.1 s

---

这个loss怎么是往上升的 作者能回复一下吗。。。@nzc

[nzc]

@CXY573 可能是由于数据集太小导致的。也有可能是dcn前面几轮是很难收敛的。会出现抖动。