question about using pre-trained caffe model

[amy-hyunji]

Hi I realized that you load pre-trained parameters of caffe.

I was wondering if using the pre-trained parameters of caffe is necessary because I want to try training with other datasets rather than BSD500 and I assume that if pre-trained parameters are used, my new datasets won't be able to give an effect.

Thank you :)

[xwjabc]

In training procedure: python hed.py --vgg16_caffe ./data/5stage-vgg.py36pickle There are 2 steps:

1. Loaded pre-trained VGG-16 weights, which can provide better convolutional features.
2. Fine-tune the model on BSDS500 dataset. If you want to use your own dataset, you only need to change the step 2 with your dataset. It will be effective when you fine-tune your model on your dataset.

[herleeyandi]

Hi @xwjabc thank you for your works! I don't install matlab in my computer, I am still curious, what is this command for? (echo "data_dir = '../output-mypretrain/test'"; cat eval_edge.m)|matlab -nodisplay -nodesktop -nosplash The .png files inside the folder is your pretrained model result right?, then why you need to display using matlab?

[xwjabc]

@herleeyandi The command line you mentioned is used to evaluate the OIS and ODS metrics. If you only need to obtain the generated images, you can ignore this line.

[long123524]

Can I train the HED network directly without loading the pre-training weights?

[xwjabc]

@long123524 I would suggest training the HED with VGG-16 pre-trained weights, since the VGG features can provide visual cues for the edge detection (esp BSDS is a relatively small dataset). However, you can still try to train HED from scratch.