Few Questions regarding training on custom data

[rono221]

Hi, I am trying to train OneFormer on the custom dataset and I was able to start the training. But, I have a few questions regarding choosing the right settings. Currently I resued ADE20k config file after editing the number of classes, iterations, and batch size.

1. What does DETECTIONS_PER_IMAGE do and how to choose the right value?
2. How to choose the right crop size? and will it impact the training or prediction time?
3. I have 20k labeled images and I am training on 4 NVIDIA A100 40GB GPUs with batch size 4, what is the minimum number of iterations required to get good results?

[rono221]

@praeclarumjj3 could you please help

[praeclarumjj3]

Hi @rono221, thanks for your interest in our work. To answer your questions:

1. DETECTIONS_PER_IMAGE is the top k queries considered from the total number of queries during the inference stage for the final instance segmentation predictions. Usually, it's fine to set it equal to NUM_QUERIES.
   

   OneFormer/oneformer/oneformer_model.py

   Line 214 in 7611899

   "test_topk_per_image": cfg.TEST.DETECTIONS_PER_IMAGE,

   
   

   OneFormer/oneformer/oneformer_model.py

   Line 445 in 7611899

   scores_per_image, topk_indices = scores.flatten(0, 1).topk(self.test_topk_per_image, sorted=False)

2. The right crop size depends on your use case. If, during the inference, you want to input high-resolution or low-resolution images, it's beneficial to train with a comparable resolution. The crop size is only used during training, so training with a larger resolution will take more time.

3. It depends on the number of classes in your dataset as well, but since the size is similar to ADE20K, I would recommend training for 160k iterations to establish a baseline.

[praeclarumjj3]

Closing this, feel free to re-open.