How max_windows affect dataset generation?

[ammarsdc]

Hi @AdeelH @lewfish.

In our pipeline, we tried few number for max_windows. As of now, the max we obtained is 32. If we set it lower, training and validation dataset windows generated is following that number. But not the case if we set higher than that.

My questions are:

1. How to know the maximum windows we could use for each scene?
2. Why RV set max to 32?
3. If we have multiple datasets/scenes to train later on, what will be the value for this? Is it will randomly picked 32 windows from all those datasets?

Thanks!

[lewfish]

The max_windows argument is used when method is set to random as described in these docs. This controls the maximum number of randomly sampled windows per scene. This sets the maximum number of windows as opposed to an exact number of windows since a random sampler is used to get each window with a maximum number of tries (set by max_sample_attempts), and the sampler may fail to find a window. This may be surprising, but if the AOI is very small and/or has a complex shape, it may be difficult to randomly sample windows that fit within that AOI. The default value is 10,000. I'm not sure where you are getting the value of 32 from. Could you point me to the source?

[lewfish]

The train.png and valid.png files just show a sample of the windows for debugging purposes. I think the number of samples is equal to the batch_sz. The 313 is the number of batches per epoch which is roughly equal to the total number of samples divided by the batch size. (An epoch encompasses a single run through all the samples.)

[ammarsdc]

That's a great crystal clear explanation@lewfish! I'll see my configuration shortly to understand better on this.

[ammarsdc]

Hi @lewfish. I just reviewed the config.

I think the number of samples is equal to the batch_sz.

For now, our batch_sz was remain as the default, which is 32. If number of samples was the same means there is 32 samples. What would be the calculation for total number of samples then to be divided with the batch size?

@AdeelH if you would like to add in something, I would appreciate it so much.

Appreciate all your help on this.
Once again, thanks!

[lewfish]

Actually, I wasn't totally clear :) I meant that the number of windows shown in train.png and valid.png is equal to the batch_sz. The total number of samples in the dataset is 10,000. I know that because of the logs you showed above:

2022-11-01 06:25:15:rastervision.pytorch_learner.learner: INFO - train_ds: 10000 items
2022-11-01 06:25:15:rastervision.pytorch_learner.learner: INFO - valid_ds: 10000 items

So, the (ceiling of the) number of samples divided by the batch size is the number of steps (ie. batches processed) per epoch, which in your case is 313.

[ammarsdc]

Okay @lewfish.
That sounds really make sense for me 👍

Thanks for this explanation.
I'll try to figure it out from here.