JOIN
Get Time
forums   
Search | Watch Thread  |  My Post History  |  My Watches  |  User Settings
View: Flat (newest first)  | Threaded  | Tree
Previous Thread  |  Next Thread
need clarification on file structure for software submission | Reply
We are preparing our software for delivery, and we have the following questions:
1. The Final Testing Guide does not mention the val subdirectory of data. It only specifies the train directory. May we assume the val directory will be present in the same structure it is in fmow-rgb and fmow-full?
2. The Final Testing Guide says nothing about the false_detection data that you provided. May we assume that it too will be accessible?
3. The Final Testing Guide does not specify how the locations of the train (and val) data for the two datasets, fmow-rgb and fmow-full, will be specified. They will most likely be in separate locations.
4. During training, we will need to store intermediate files such as cropped images and model weights. How will this storage location be specified to the train.sh script?

May we suggest the following sample call to the training script, with the understanding that each of the two directories, /data/fmow-rgb and /data/fmow-full, will have three subdirectories named train, val, and false_detection?
./train.sh /data/fmow-rgb /data/fmow-full /temporary/storage/location

So, the train images for fmow-rgb will be in /data/fmow-rgb/train, the val images for fmow-full will be in /data/fmow-full/val, the false_detection images for fmow-rgb will be in /data/fmow-rgb/false_detection, and so on. The structure of the subdirectories and files should be exactly the same as they are in the original deliveries you gave us, and we recommend that you define this structure clearly in an addendum to the Final Testing Guide. (You defined this structure for train. It just needs to be extended to val and false_detection for each of the two buckets.)

Thanks,

Mark Pritt
Team YellowSubmarine
Re: need clarification on file structure for software submission (response to post by prittm) | Reply
Thank you for the constructive feedback. Before responding I'd like to confirm that you know that you will have 10 days to create your dockerized version after the end of the submission phase (31st December), so it's not necessary to do it now.

To your points:
1. The Final Testing Guide does not mention the val subdirectory of data. It only specifies the train directory. May we assume the val directory will be present in the same structure it is in fmow-rgb and fmow-full?
Yes.

2. The Final Testing Guide says nothing about the false_detection data that you provided. May we assume that it too will be accessible?
Yes. (The false detection data was not yet available when I wrote the testing guide, but you are correct, it should be mentioned.) It will be placed in a folder called false_detection_metadata on the same level where the train/, val/, and test/ folders are.

3. The Final Testing Guide does not specify how the locations of the train (and val) data for the two datasets, fmow-rgb and fmow-full, will be specified. They will most likely be in separate locations.
My assumption was that people will use either the rgb or the full data set, but we can have both if this is really needed. Note that this also applies to the test data: if someone uses both versions for training then most likely he will also need both during inference. Question to the contestants: is there anyone who will need both data versions for final testing?

4. During training, we will need to store intermediate files such as cropped images and model weights. How will this storage location be specified to the train.sh script?
It will not be specified, you have full control over the virtual machine you create with your docker build, you can store data whereever you wish.

I will modify the testing guide based on the result of this discussion.
Re: need clarification on file structure for software submission (response to post by walrus71) | Reply
I received no answers to this question I asked a couple of days ago: is there anyone who will need both data versions for final testing? I'd like to ask especially the top ranked contestants (who will likely participate in the final testing) to give an answer. It's totally fine to answer privately at walrus71@copilots.topcoder.com if you don't want to reveal this piece of information to others. Thank you for your cooperation.
Re: need clarification on file structure for software submission (response to post by walrus71) | Reply
Will we have write-access to the data directory (with train, val etc.) mounted to our docker container?
Re: need clarification on file structure for software submission (response to post by gonzalb) | Reply
No, the /data directory will be mounted with read only permissions. I'll add this to the testing guide, thanks for asking.
Re: need clarification on file structure for software submission (response to post by walrus71) | Reply
I do not think our team will need access to both versions of data.
RSS