sorry to trouble

[xuebingbing121]

No description provided.

[yzou2]

1. 505 is for train set and 500 is for val set of cityscapes. 2) We also observe this phenomenon. My guess is that the pseudo-labels do not change to much, and leading to similar model performance with a small learning rate.

[yzou2]

It's the number of images...

…

On Mon, Jul 13, 2020 at 12:11 PM xuebingbing121 ***@***.***> wrote: Yes, i know&nbsp;505 is for train set and 500 is for val set of cityscapes, i want to ask why is the number 505 and 500? Is it iteration or epoch? Thanks! 发自我的iPhone ------------------ Original ------------------ From: yzou2 ***@***.***&gt; Date: Tue,Jul 14,2020 0:05 AM To: yzou2/CRST ***@***.***&gt; Cc: xuebingbing121 ***@***.***&gt;, Author < ***@***.***&gt; Subject: Re: [yzou2/CRST] sorry to trouble (#27) 505 is for train set and 500 is for val set of cityscapes. 2) We also observe this phenomenon. My guess is that the pseudo-labels do not change to much, and leading to similar model performance with a small learning rate. — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or unsubscribe. — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#27 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AEY3Z5724R4XLUQ3VT27BZ3R3MW33ANCNFSM4OYOFSLA> .

[xuebingbing121]

OK i see😂It's a little embarrassed. So where can i check the number of iteration and epoch？ &nbsp; 发自我的iPhone

…

------------------ Original ------------------ From: yzou2 <[email protected]&gt; Date: Tue,Jul 14,2020 0:12 AM To: yzou2/CRST <[email protected]&gt; Cc: xuebingbing121 <[email protected]&gt;, Author <[email protected]&gt; Subject: Re: [yzou2/CRST] sorry to trouble (#27) It's the number of images... On Mon, Jul 13, 2020 at 12:11 PM xuebingbing121 <[email protected]&gt; wrote: &gt; Yes, i know&amp;nbsp;505 is for train set and 500 is for val set of &gt; cityscapes, i want to ask why is the number 505 and 500? Is it iteration or &gt; epoch? Thanks! &gt; &gt; &gt; &gt; 发自我的iPhone &gt; &gt; &gt; ------------------ Original ------------------ &gt; From: yzou2 <[email protected]&amp;gt; &gt; Date: Tue,Jul 14,2020 0:05 AM &gt; To: yzou2/CRST <[email protected]&amp;gt; &gt; Cc: xuebingbing121 <[email protected]&amp;gt;, Author < &gt; [email protected]&amp;gt; &gt; Subject: Re: [yzou2/CRST] sorry to trouble (#27) &gt; &gt; &gt; &gt; &gt; &gt; &gt; 505 is for train set and 500 is for val set of cityscapes. 2) We also &gt; observe this phenomenon. My guess is that the pseudo-labels do not change &gt; to much, and leading to similar model performance with a small learning &gt; rate. &gt; &gt; — &gt; You are receiving this because you authored the thread. &gt; Reply to this email directly, view it on GitHub, or unsubscribe. &gt; &gt; — &gt; You are receiving this because you commented. &gt; Reply to this email directly, view it on GitHub &gt; <#27 (comment)&gt;, or &gt; unsubscribe &gt; <https://github.com/notifications/unsubscribe-auth/AEY3Z5724R4XLUQ3VT27BZ3R3MW33ANCNFSM4OYOFSLA&gt; &gt; . &gt; — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or unsubscribe.

[yzou2]

The iterations and epochs should be in the training log part, not the validation/testing part. On Mon, Jul 13, 2020 at 12:17 PM xuebingbing121 <[email protected]> wrote:

…

OK i see😂It's a little embarrassed. So where can i check the number of iteration and epoch？ &nbsp; 发自我的iPhone ------------------ Original ------------------ From: yzou2 ***@***.***&gt; Date: Tue,Jul 14,2020 0:12 AM To: yzou2/CRST ***@***.***&gt; Cc: xuebingbing121 ***@***.***&gt;, Author < ***@***.***&gt; Subject: Re: [yzou2/CRST] sorry to trouble (#27) It's the number of images... On Mon, Jul 13, 2020 at 12:11 PM xuebingbing121 ***@***.***&gt; wrote: &gt; Yes, i know&amp;nbsp;505 is for train set and 500 is for val set of &gt; cityscapes, i want to ask why is the number 505 and 500? Is it iteration or &gt; epoch? Thanks! &gt; &gt; &gt; &gt; 发自我的iPhone &gt; &gt; &gt; ------------------ Original ------------------ &gt; From: yzou2 ***@***.***&amp;gt; &gt; Date: Tue,Jul 14,2020 0:05 AM &gt; To: yzou2/CRST ***@***.***&amp;gt; &gt; Cc: xuebingbing121 ***@***.***&amp;gt;, Author < &gt; ***@***.***&amp;gt; &gt; Subject: Re: [yzou2/CRST] sorry to trouble (#27) &gt; &gt; &gt; &gt; &gt; &gt; &gt; 505 is for train set and 500 is for val set of cityscapes. 2) We also &gt; observe this phenomenon. My guess is that the pseudo-labels do not change &gt; to much, and leading to similar model performance with a small learning &gt; rate. &gt; &gt; — &gt; You are receiving this because you authored the thread. &gt; Reply to this email directly, view it on GitHub, or unsubscribe. &gt; &gt; — &gt; You are receiving this because you commented. &gt; Reply to this email directly, view it on GitHub &gt; <#27 (comment)&gt;, or &gt; unsubscribe &gt; < https://github.com/notifications/unsubscribe-auth/AEY3Z5724R4XLUQ3VT27BZ3R3MW33ANCNFSM4OYOFSLA&gt; &gt; . &gt; — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or unsubscribe. — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#27 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AEY3Z5ZYYG6S6YATFSB3UMDR3MXRNANCNFSM4OYOFSLA> .

[yzou2]

It's in the log file....

[xuebingbing121]

Oh! I commented it out because this sentence will show errors before.

[yzou2]

You may want to debug it out. For the training data, if you want to keep the GTA2Cityscapes benchmark, the val set cannot be changed. But the training set of my code the data list is one part of the entire training set of cityscapes, you can modify the list accordingly.

…

On Mon, Jul 13, 2020 at 10:54 PM xuebingbing121 ***@***.***> wrote: Oh! I commented it out because this sentence will show errors before. — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#27 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AEY3Z53S5AARSH3Z3WAMGWTR3PCHFANCNFSM4OYOFSLA> .

[xuebingbing121]

Is the iou in the paper also a single picture of iou?

[xuebingbing121]

Hello, in the paper, is the miou displayed in the result a certain picture with the highest miou value? Or is it something else?

[yzou2]

Usually segmentation papers use miou for the whole dataset/images already been seen for evaluation and comparison...

[xuebingbing121]

2020-07-02 23:35:10,328 Host
[85.42 54.93 74.01 27.39 13.44 37.51 31.59 16.92 84.69 38.55 79.67 57.88
20.52 78.75 24.23 40.44 0.00 28.06 39.41]
2020-07-02 23:35:11,655 Host Done 490/500 with speed: 0.75/s
2020-07-02 23:35:11,655 Host mean iou: 43.87%
2020-07-02 23:35:11,655 Host
[85.39 54.92 74.01 27.42 13.44 37.53 31.59 16.92 84.72 38.63 79.66 57.88
20.52 78.80 24.23 40.44 0.00 28.06 39.41]
2020-07-02 23:35:12,977 Host Done 491/500 with speed: 0.75/s
2020-07-02 23:35:12,978 Host mean iou: 43.86%
2020-07-02 23:35:12,978 Host
[85.38 54.85 73.97 27.47 13.43 37.50 31.59 16.91 84.70 39.01 79.66 57.88
20.52 78.65 24.21 40.18 0.00 28.06 39.41]
2020-07-02 23:35:14,307 Host Done 492/500 with speed: 0.75/s
2020-07-02 23:35:14,307 Host mean iou: 43.86%
2020-07-02 23:35:14,307 Host
[85.36 54.82 73.96 27.45 13.43 37.51 31.58 16.87 84.75 39.23 79.65 57.88
20.51 78.55 24.21 40.17 0.00 28.06 39.41]
2020-07-02 23:35:15,636 Host Done 493/500 with speed: 0.75/s
2020-07-02 23:35:15,636 Host mean iou: 43.88%
2020-07-02 23:35:15,636 Host
[85.38 54.83 73.96 27.44 13.43 37.53 31.63 17.06 84.77 39.23 79.73 57.86
20.47 78.50 24.18 40.18 0.00 28.06 39.41]
2020-07-02 23:35:16,965 Host Done 494/500 with speed: 0.75/s
2020-07-02 23:35:16,966 Host mean iou: 43.94%
2020-07-02 23:35:16,966 Host
[85.39 54.81 73.96 27.42 13.43 37.55 32.09 17.06 84.78 39.19 79.80 57.69
20.84 78.48 24.18 40.18 0.00 27.82 40.22]
2020-07-02 23:35:18,294 Host Done 495/500 with speed: 0.75/s
2020-07-02 23:35:18,294 Host mean iou: 43.96%
2020-07-02 23:35:18,295 Host
[85.42 54.82 73.95 27.40 13.51 37.59 32.08 17.20 84.78 39.20 79.94 57.69
20.84 78.53 24.18 40.10 0.00 27.82 40.22]
2020-07-02 23:35:19,621 Host Done 496/500 with speed: 0.75/s
2020-07-02 23:35:19,621 Host mean iou: 43.97%
2020-07-02 23:35:19,621 Host
[85.44 54.83 73.96 27.39 13.51 37.58 32.25 17.11 84.78 39.19 80.06 57.68
20.83 78.52 24.16 40.10 0.00 27.82 40.20]
2020-07-02 23:35:20,939 Host Done 497/500 with speed: 0.75/s
2020-07-02 23:35:20,939 Host mean iou: 43.97%
2020-07-02 23:35:20,939 Host
[85.47 54.83 73.95 27.37 13.51 37.55 32.23 17.43 84.78 38.96 80.11 57.68
20.83 78.52 24.16 40.10 0.00 27.82 40.20]
2020-07-02 23:35:22,262 Host Done 498/500 with speed: 0.75/s
2020-07-02 23:35:22,262 Host mean iou: 43.98%
2020-07-02 23:35:22,263 Host
[85.48 54.82 73.94 27.29 13.42 37.54 32.23 17.43 84.82 39.16 80.17 57.67
20.81 78.52 24.19 40.17 0.00 27.82 40.19]
2020-07-02 23:35:23,582 Host Done 499/500 with speed: 0.75/s
2020-07-02 23:35:23,582 Host mean iou: 43.99%
2020-07-02 23:35:23,583 Host
[85.50 54.81 73.95 27.29 13.42 37.54 32.23 17.42 84.84 39.17 80.20 57.67
20.81 78.60 24.19 40.17 0.00 27.82 40.19]
2020-07-02 23:35:24,900 Host Done 500/500 with speed: 0.75/s
2020-07-02 23:35:24,900 Host mean iou: 43.99%
2020-07-02 23:35:24,900 Host
[85.50 54.80 73.98 27.26 13.41 37.54 32.23 17.42 84.85 39.17 80.30 57.67
20.81 78.57 24.19 40.14 0.00 27.82 40.19]
2020-07-02 23:35:24,915 Host ###### Finish evaluating in target domain test set in round 3! Time cost: 663.12 seconds. ######

Do you mean to average out all 500 results?
I want to know how do you get the result 46.47|89.91 | 53.84 | 79.73 | 30.29 | 19.21..., etc.？

[yzou2]

Line 757 the class ScoreUpdater in CRST_seg.py defines how to evaluate the results.

[xuebingbing121]

OK, Thanks!
I noticed that in train.py, BATCH_SIZE = 4, the number of iterations is 1000, but there are 24966 images in the training set. In this case, one epoch cannot pass all the samples in the training set. What should I do?

Also, how should I increase the number of epochs? Because I want to confirm whether the model converges after I add the dropout layer.Thanks!

[xuebingbing121]

This is very important to me and I look forward to your reply！

[yzou2]

In original DeepLab, iterations rather than epochs are used. As long as the network see enough images, it does not matter if a full epoch in the very end is went through. And in my original code, num_steps is set to be 100,000.....

[xuebingbing121]

Ok , Thanks!
What is the difference between the model trained using train.py and the model provided on github? Why is there a big gap between the results of the two models?

[yzou2]

What are your specific numbers? And I think you change your network structures. And a trick here is that not training model for to long since the model will overfit to gta5 And lose generalization ability. You may want to evaluate every 5000 iterations to see which model is best for transfer learning.

On Wed, Jul 29, 2020 at 11:58 PM bbzfighting ***@***.***> wrote: Ok , Thanks! What is the difference between the model trained using train.py and the model provided on github? Why is there a big gap between the results of the two models? [image: 2020-07-30 11-52-59 的屏幕截图] <https://user-images.githubusercontent.com/48796790/88879011-d673ed80-d25b-11ea-8782-2b528f15c220.png> — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#27 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AEY3Z5Y5R3YDJNAZBIO3FF3R6DVVTANCNFSM4OYOFSLA> .

-- null

[xuebingbing121]

I didn’t change the network structure. I trained according to your default parameters. I used 5000 and 100000 models respectively, but their effects were not good.

[xuebingbing121]

Oh , I commented out these :

[xuebingbing121]

Is the source domain training model you published on github trained with the train .py file?
Do you mean I don 't need to train the entire data set, and it 's better to evaluate it on the target domain every 5000 iterations? Is it okay even if the loss of the model does not converge?
When I evaluate, do I mean to look at the result of evaluate.py on the target domain, or do I need to evaluate both the source domain and the target domain?
Thank you for your patience to answer, I'm sorry to have lost your time！

[yzou2]

Sometimes you can use the target labels to find a source model that has a good transfer learning performance, although this may not be true in practice.

[xuebingbing121]

Hi！The final result of your paper Road iou=89.91
Is this iou the average of 500 pictures or the iou with the highest 500 pictures?
If it is the average of 500 pictures, why didn't I see the calculation of the average in the code?
Because the output of your code is the iou of a single picture.
Hope you can tell me！Thanks a lot！

[yzou2]

Do you know what is iou, confusion matrix, how to calculate confusion matrix for multiple classes and how to calculate iou from confusion matrices? If you do know, you can understand from the code the iou is also calculated from confusion matrix and confusion matrix is cumulated from every image by "self._confs += conf" in the function of "def update(self, pred_label, label, i, computed=True)".

[xuebingbing121]

Ok, I seem to see what you mean. You're saying that the last &nbsp;output iou is a total IOU after 500 pictures have been updated, and it's not the iou for the last picture, right? Sorry, &nbsp;as for the clear definition and calculation of the iou, confusion matrix, I'm going to learn and study it carefully. 发自我的iPhone

…

------------------ Original ------------------ From: yzou2 <[email protected]&gt; Date: Tue,Aug 4,2020 10:41 PM To: yzou2/CRST <[email protected]&gt; Cc: bbzfighting <[email protected]&gt;, Author <[email protected]&gt; Subject: Re: [yzou2/CRST] sorry to trouble (#27) Do you know what is iou, confusion matrix, how to calculate confusion matrix for multiple classes and how to calculate iou from confusion matrices? If you do know, you can understand from the code the iou is also calculated from confusion matrix and confusion matrix is cumulated from every image by "self._confs += conf" in the function of "def update(self, pred_label, label, i, computed=True)". — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or unsubscribe.

[yzou2]

Yes. You get the point.