Artificial Intelligence Forum
cancel
Showing results for 
Search instead for 
Did you mean: 
Read only

Training Service for Customizable Image Object Detection - long runtime, no result

kmeersman
Discoverer
0 Likes
1,840

Dear Experts,

we are using this training service for fine tuning the default model. However, submitted jobs run for a very long time(days). Afterwards, the "jobs" folder and sub-folders remain mostly empty. There is 1 pipeline_xxx.log file but it has only 1 line: INFO:__main__:Using default checkpoint for training.

Training data:

Job submit body:

{ "jobName": "boxes_od2", "dataset": "boxes_dataset", "modelName": "boxes_od2", "iterations": "1000" }

Please advise.

Thanks & Regards,

Klaas.

1 ACCEPTED SOLUTION
Read only

AndreasForster
Product and Topic Expert
Product and Topic Expert
0 Likes
1,708

Hello Klaas,

Thank you very much for the feedback on how you made it work.

I am sure this will be helpful for others and I will pass it on to our product group for consideration.

Many Greetings

Andreas

View solution in original post

Dear Experts,

we are using this training service for fine tuning the default model. However, submitted jobs run for a very long time(days). Afterwards, the "jobs" folder and sub-folders remain mostly empty. There is 1 pipeline_xxx.log file but it has only 1 line: INFO:__main__:Using default checkpoint for training.

Training data:

Job submit body:

{ "jobName": "boxes_od2", "dataset": "boxes_dataset", "modelName": "boxes_od2", "iterations": "1000" }

Please advise.

Thanks & Regards,

Klaas.

3 REPLIES 3
Read only

kmeersman
Discoverer
0 Likes
1,708

Update:

Turns out the annotation filename was wrong. It should be "annotation" and not "annotations".

I'm surprised the job will continue to run for days even if no annotation file can be found. Also nothing in the log about a missing annotation file.

Now we are facing the next problem. For every annotation file entry we see this log entry: "row x removed as object size is too small"

We have double checked, height and width are correct and the objects are larger than 1/200th of the image size.

This is our annotation file:

IMG_0835;1008;756;Box;78;239;753;499
IMG_0836;1008;756;Box;219;242;528;358

And this is the resulting log:

INFO:__main__:Using default checkpoint for training.

INFO:check_annotations:#####################################

INFO:check_annotations:Checking Annotation File

INFO:check_annotations:#####################################

INFO:check_annotations:

INFO:check_annotations:Row 0 removed as object size is too small

INFO:check_annotations:Row 1 removed as object size is too small

INFO:check_annotations:Annotation file has passed the validity test

INFO:check_annotations:#####################################

INFO:check_annotations:Annotation Cleared

INFO:check_annotations:#####################################

INFO:check_annotations:

INFO:create_tfrecord:#####################################

INFO:create_tfrecord:Creating TFRecord

INFO:create_tfrecord:#####################################

INFO:create_tfrecord:

Read only

kmeersman
Discoverer
0 Likes
1,708

Update2:

Finally got the service working: the annotation file should be a ',' separated file and not a ';' separated file.

So if you are using Excel to create the file, make sure to use comma delimited!

Maybe this information will be helpful to someone.

Regards,

Klaas.

Read only

AndreasForster
Product and Topic Expert
Product and Topic Expert
0 Likes
1,709

Hello Klaas,

Thank you very much for the feedback on how you made it work.

I am sure this will be helpful for others and I will pass it on to our product group for consideration.

Many Greetings

Andreas