BichenWuUCB
training without pre-trained model
Thank you for the code. It's great.I want to train a result now, without using pre-trained model, If it can work, what should be changed or noted?If not, why not? I am now training on Kitty data without the pre-trained model, but the accuracy rate is about 0.1%
Hello,
Are you using the same dataset provided in readme.md file for training without pre-trained models? Thanks.
hello,I have encountered the same problem as you. I deleted the input of the source code pre-training model, initialized with normal-distributed parameters, and trained hundreds of thousands of times to verify that there is still no precision.
Training an object detector from scratch is possible but not trivial, have a look at this paper: DSOD: Learning Deeply Supervised Object Detectors from Scratch
It is not possible to train squeezedet from scratch, you have to start with pretrained weights, I would also recommend looking at other more recent designs such as SSD Mobilenet, or a version of Retinanet.
