search

loolzaaa / faster-rcnn-pytorch

A PyTorch implementation of Faster R-CNN
MIT License
17 stars 5 forks source link

Multi scale training #7

Open ghost opened 4 years ago

ghost commented 4 years ago

How do I change the config to start training at multiple scales? The min image size parameter in the config does not accept multiple values and throws error when training starts.

loolzaaa commented 4 years ago

What you mean "multiple scales"? This parameters are responsible for the actual size of the image AFTER resizing. Other words, min image size parameter guarantee size of the minimal side of the image.

ghost commented 4 years ago

Sorry for any confusion, I am a beginner, I want to train the FRCNN for multiple values of min image size (1000 px, 1200px, and 1600px) chosen randomly for different batches. Other repos like this have scale parameter as given in the config here which accepts multiple scales. Do I need to change the collate function or is there any similar parameter in your config as well?