Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

关于encoder 和forecaster之间的数据流 #3

Closed
chencodeX opened this issue Apr 2, 2018 · 10 comments

Comments

Projects
None yet
3 participants
@chencodeX
Copy link

commented Apr 2, 2018

image
在红框处,encoder底层的最后一个RNNCell 的输出 Ht ,是forecaster顶层的第一个RNNCell的 data,还是它的Ht?

我画了一个我理解的图,您可以看看
image

@sxjscience

This comment has been minimized.

Copy link
Owner

commented Apr 2, 2018

是H_t,forecaster的第一个RNN的data是全0

@chencodeX

This comment has been minimized.

Copy link
Author

commented Apr 10, 2018

您在训练的时候有没有发现,预测的结果中,越往后的时刻,整体图像的亮度在不断变暗。

@sxjscience

This comment has been minimized.

Copy link
Owner

commented Apr 10, 2018

@chencodeX

This comment has been minimized.

Copy link
Author

commented Apr 11, 2018

您是怎么解决这个问题的呢?

@chencodeX

This comment has been minimized.

Copy link
Author

commented Apr 12, 2018

在您的论文中,有一部分是关于dBZ和降雨等级之间的转换的。
image
请问您这个log是以十为底还是以自然对数为底呢?

@sxjscience

This comment has been minimized.

Copy link
Owner

commented Apr 12, 2018

@chencodeX

This comment has been minimized.

Copy link
Author

commented Apr 12, 2018

以10为底的时候,dBZ为70的情况下,R为2260. 这正常吗?

@chencodeX chencodeX closed this Apr 12, 2018

@wcwoo

This comment has been minimized.

Copy link
Collaborator

commented Apr 28, 2018

雷達dBZ正常不超過65,其實在60很可能已在落冰雹了。
關於採用的a,b值,是根據三年實測數據以TMM和綫性回歸計算出來的。

@wcwoo wcwoo reopened this Apr 28, 2018

@wcwoo

This comment has been minimized.

Copy link
Collaborator

commented May 1, 2018

有關整體圖像不斷變暗的現象,在我們業務測試中也有留意到。我猜想是因為神經網絡明白了外推法的限制,所以當預報時限足夠長時寧願什麼都不預測。這也涉及評分準則,因為是定時定點計算,而且加重了大雨的權重,對長時間預報任何預測都幾乎成為空報,同時又未能減少漏報,所以最優化的策略就是不報了。將來可以考慮改用三維(時空)FSS,或讓評分有點容錯度,以改善這個情況。

@chencodeX

This comment has been minimized.

Copy link
Author

commented May 2, 2018

非常感谢您的解答

@wcwoo wcwoo closed this May 2, 2018

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
You can’t perform that action at this time.