Epoch size about u-2-net

[xuebinqin]

Thanks for your interests. Instead of counting the epoch number, we use iteration number and trained around 600K iterations (it takes around 120 hours on single 1080 ti GPU) with batch size of 12 on horizontally flipped DUTS-TR dataset. What do you mean by "tar loss of 0.01"?

[ds-jpg]

Hi Nathan
Sorry, my bad, what I meant to say was when I start training your pre-trained model with duts dataset, it gives tar: 0.014615 and train loss: 0.165076. What I wanted to know is that suppose if I train this model from scratch only with my custom dataset which is nearly 20K images ) then how can I detect if the model is overfitting. And do I need to train the model until the tar reaches: 0.01 and train loss: 0.16
Thanks a lot

[xuebinqin]

No you don't have to care about the tar and train loss. Because the losses for different datasets may differ. You may have to observe the training loss curve (or log(loss) curve). To avoid "overfitting", you can also split your data into train 80% and val 20%. We didn't use validation set here because I trained many times (in the whole process of my algorithm development) and I know that the training process won't overfit in the training set. In addition, We have to evaluate the performance on six different datasets, which are very different. We were actually using MSRA 2500 as the validation set but that dosn't give us the best model (because the distribution of MSRA 2500 and other datasets are very different, validation loss and accuracy on this dataset are not able to reflect the model performance). Later, we give the validation process up and pick several models based on the training loss and evaluate them on different datasets and chose the best model weights. I don't know what kind of dataset are you using, but I think you dataset is good in terms of the size 20K. I guess the probability of overfitting should be low.

…

On Sun, May 10, 2020 at 11:53 PM Deep190 ***@***.***> wrote: Hi Nathan Sorry, my bad, what I meant to say was when I start training your pre-trained model with duts dataset, it gives tar: 0.014615 and train loss: 0.165076. What I wanted to know is that suppose if I train this model from scratch only with my custom dataset which is nearly 20K images ) then how can I detect if the model is overfitting. And do I need to train the model until the tar reaches: 0.01 and train loss: 0.16 Thanks a lot — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#4 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ADSGORNQMNLAD4SHDN6QYFLRQ6HF7ANCNFSM4M5GNLUQ> .

-- Xuebin Qin PhD Department of Computing Science University of Alberta, Edmonton, AB, Canada Homepage:https://webdocs.cs.ualberta.ca/~xuebin/

[ds-jpg]

Hi Nathan,
Thank you so much for the quick reply.

I have another query that I want to ask you. After training the model from scratch with our custom dataset which is around 20K suppose we generate another 10K new dataset, in that case -

a) Can we do transfer learning on the last saved weight with the newly collected 10K dataset

Or

b) Should we do incremental learning (by mixing our old 20K dataset with the new 10K dataset and then train on the last saved model?)

Thanks in advance and looking forward to hearing from you.

Best regards,
Deep

[xuebinqin]

a) Can we do *transfer learning* on the last saved weight with the newly collected 10K dataset b) Should we do *incremental learning* (by mixing our old 20K dataset with the new 10K dataset and then train on the last saved model?) Yes, you can do both. I prefer the second way. Because that avoids the overfitting on the new 10K dataset. Honestly, It uis hard to see which would be better. It really depends on the data distribution and your last saved model (it is usually in a local minimua of the previous dataset). But I suggest you also retrain the network from scratch by mixing all the data if you have enough time and resources.

…

On Mon, May 11, 2020 at 7:12 AM Deep190 ***@***.***> wrote: Hi Nathan, Thank you so much for the quick reply. I have another query that I want to ask you. After training the model from scratch with our custom dataset which is around 20K suppose we generate another 10K new dataset, in that case - a) Can we do *transfer learning* on the last saved weight with the newly collected 10K dataset Or b) Should we do *incremental learning* (by mixing our old 20K dataset with the new 10K dataset and then train on the last saved model?) Thanks in advance and looking forward to hearing from you. Best regards, Deep — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#4 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ADSGORI75IUVAFGDPOMJAOLRQ72VBANCNFSM4M5GNLUQ> .

-- Xuebin Qin PhD Department of Computing Science University of Alberta, Edmonton, AB, Canada Homepage:https://webdocs.cs.ualberta.ca/~xuebin/

[ds-jpg]

Hi Nathan,

I hope you are well.

Can you please tell me how can we generate a 4 channel mask, currently the mask that is being generated from the model happens to be a 2 channel mask but we want to generate a 4 channel mask.

looking forward to hearing from you.

Thanks a lot
Deep

[xuebinqin]

Sorry Deep, I couldn't fully understand your question. Do you mean the outputs of our network are 2 channels ? I think the both our dataloader and network will only generate 1 channel training labels and prediction probability maps. In order to handle multiple classes, you have to change the output channel of the network and also change the sigmoid to softmax. Besides, the dataloader has also to be changed for loading multiple classes segmentation mask (which is usually prepared as a single mask with pixels of different classes labeled as 0 (background), 1, 2, ..., num_of_classes (it can be converted to a four channels one-hot mask by pytorch functions)).

…

On Fri, May 15, 2020 at 2:23 AM Deep190 ***@***.***> wrote: Hi Nathan, Hope you are well. Can you please help tell me how can we generate 4 channel mask, currently the mash that is generated happens to be a 2 channel mask but we want to generate a 4 channel mask. looking forward to hearing from you. Thanks a lot Deep — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#4 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ADSGORK76XIOTSDIDJIHWDTRRT3WXANCNFSM4M5GNLUQ> .

-- Xuebin Qin PhD Department of Computing Science University of Alberta, Edmonton, AB, Canada Homepage:https://webdocs.cs.ualberta.ca/~xuebin/

[ds-jpg]

Hi Nathan,
Thanks for the quick reply.

Also, Nathan, It will be really helpful if you can kindly answer the following things

Situation 1 :
When I am training your model on my custom dataset with an 80K image + mask, then in that case do I need to Horizontally flip the images. Or will the images get horizontally flipped from the model itself?

Situation 2 :
What should we do, when training the model with a large dataset of let's say 100K or more, and in that case if the changes are required then where and what changes should we do in the model?

Situation 3 :
We have already trained your model with our custom dataset of around 40K containing a lot of sharp edges, but the performance of the model is still poor on images with sharp edges and especially near the edges of the image, In order to get the best performance do we need to add more dataset like that

Thanks a lot in advance :)

Looking forward to hearing from you.

Best Regards
Deep

[xuebinqin]

Hi, sorry for the late reply. Situation 1 : When I am training your model on my custom dataset with an 80K image + mask, then in that case do I need to Horizontally flip the images. Or will the images get horizontally flipped from the model itself? *RES: The horizontal flipping augment the data usually does improve the performance (we suggest to do that but not mandantory). We did it offline so that it is not included in the code. But the horizontal flipping is implemented in the RandomCrop function in data_loader.py.* Situation 2 : What should we do, when training the model with a large dataset of let's say 100K or more, and in that case if the changes are required then where and what changes should we do in the model? *RES: You can change the middle and output filter numbers of each RSU block and you can also add more more layers into each stage. Besides, you can also try to replace the standard re_bn_cov element by a residual block. In summary, try to increase the network capacity by increase the model width and depth if you have no limitations on the computation resources. * Situation 3 : We have already trained your model with our custom dataset of around 40K containing a lot of sharp edges, but the performance of the model is still poor on images with sharp edges and especially near the edges of the image, In order to get the best performance do we need to add more dataset like that *RES: There are two ways to get the sharp edges: (1) try to train the model with more iterations, we train our model over 600K iterations with batch size of 12 (around 20K images), (2) replace the bce by our previous bsi losses (used in BASNet), it will give you sharp edges with less iteration. You can try both and pick the best one.*

…

On Thu, May 21, 2020 at 10:58 PM Deep190 ***@***.***> wrote: Hi Nathan, Thanks for the quick reply. Also, Nathan, It will be really helpful if you can kindly answer the following things Situation 1 : When I am training your model on my custom dataset with an 80K image + mask, then in that case do I need to Horizontally flip the images. Or will the images get horizontally flipped from the model itself? Situation 2 : What should we do, when training the model with a large dataset of let's say 100K or more, and in that case if the changes are required then where and what changes should we do in the model? Situation 3 : We have already trained your model with our custom dataset of around 40K containing a lot of sharp edges, but the performance of the model is still poor on images with sharp edges and especially near the edges of the image, In order to get the best performance do we need to add more dataset like that Thanks a lot in advance :) Looking forward to hearing from you. Best Regards Deep — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#4 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ADSGORJ5HD2B5N5Y475XPXTRSYA7ZANCNFSM4M5GNLUQ> .

-- Xuebin Qin PhD Department of Computing Science University of Alberta, Edmonton, AB, Canada Homepage:https://webdocs.cs.ualberta.ca/~xuebin/

[yaju1234]

Hi Nathan,
I humbly request you to provide the following information.

"You can change the middle and output filter numbers of each RSU block
and you can also add more more layers into each stage. Besides, you can
also try to replace the standard re_bn_cov element by a residual block. In
summary, try to increase the network capacity by increase the model width
and depth if you have no limitations on the computation resources. *"

We want to train our model with 100K(50k normal + 50k horizontal flip ) dataset from scratch.

1. " you can also add more more layers into each stage"
   Can you please guide me on how to add more layers. Do we need to create RSU8, RSU9, etc like that?
2. "try to replace the standard re_bn_cov element by a residual block"
   Can you tell me where exactly I need to make changes in your code? Also please guide what is the 're_bn_cov element'?
3. " change the middle and output filter numbers of each RSU block"
   How much we need to increase the middle and output filter numbers. Now in your model, it is
   mid_ch=12, out_ch=3
4. Do we need to change each image size within 320*320?

Looking forward to your reply.
Thanks