search

weiliu89 / caffe

Caffe: a fast open framework for deep learning.
http://caffe.berkeleyvision.org/
Other
4.77k stars 1.67k forks source link

Applying SSD to Spacenet data (repo available to suggest improvements) #210

Open aurotripathy opened 8 years ago

aurotripathy commented 8 years ago

Hi, This project (link below) applies the Caffe-based Single-Shot Detector (SSD) algorithm to the Spacenet dataset. Early results (demonstrating feasibility) are shown below.

https://github.com/aurotripathy/ssd-spacenet

Looks promising, but mAP is still quite low (0.3) after 17,000 iterations and an (effective) learning rate of 0.0001. In the training script, the base_lr is 0.000004.

If you have questions, I'll gladly answer them.

Suggestions for applying image prepossessing, image data augmentation, and hyperparameter selection are very welcome.

weiliu89 commented 8 years ago

Why is the learning rate so low? Are you fine-tuning from one of the provided models?

aurotripathy commented 8 years ago

@weiliu89 Thanks for taking a look. Yes, I'm fine-tuning from exactly the same model as the PASCAL VOC dataset, VGG_ILSVRC_16_layers_fc_reduced.caffemodel.

Any guidance on whether it is appropriate for satellite imagery datasets?

weiliu89 commented 8 years ago

Why not use same base_lr (0.001)? If the loss goes to nan after some iterations, you could use lower lr to train for a few hundred iterations, cancel the job after the loss is stable, and resume training with 0.001.

0.0001 seems too low unless you are fine-tuning from a trained VOC or COCO model.

aurotripathy commented 8 years ago

What you suggest is interesting and I will try it that way. I do recall the loss going to NaN right away (after just a few iterations)