I am trying to train retinanet on my custom data, I have converted my dataset into DOTA format and now my data-direcctory is like {Dataset_name-> images, train,val,test} where Dataset_name = root_directory and contains 4 sub-folders, one contains all images and three contains labels.

I started from the default notebook provided by the author and run it on default data ssd_tiny dataset to see the working (but it is running next day I forcefully stopped it, no outcome, same behaviour is on custom data).

Then I created a config file as stated in documentation and modified the required field and started training (this time atleast I got the .json file, log file and one other file) but then it gave me an error like
ValueError: need at least one array to concatenate

I searched out this error but it seems like everybody is directing to change the datapath in dotav1.py file, I did it but no outcome.

My query is that can we just trained a model with our custom data if it is in the above format I mentioned (I dont want to use the split file and crop my images, as my images are already small).

I also used the Retinanet config file given as example but still same problem.

Tried both relative and absolute paths for the data.

I am using windows:
python 3.8
torch 1.9.0
cuda 11.1
mmdet 2.28.2
mmrotate 0.3.4
mmcv-full 1.5.3