how can i test my large image, about 3000*4000? should i resize it to small size? about rcf

[yun-liu]

Yes, you should resize images first. As I know, there is not such GPU that can support so large images.

[wanghuok]

Thanks for your reply.
What's the total loss and every dsn loss when you have finished the training?
i modified the solve.py, use 'rcf_pretrained_bsds.caffemodel' instead of '5stage-vgg.caffemodel',but the first iteration its total loss is 1400+, how can i know whether the loss is convergence?

I0915 09:39:08.640388 4169 net.cpp:284] Network initialization done.
I0915 09:39:08.640522 4169 solver.cpp:60] Solver scaffolding done.
I0915 09:39:09.026629 4169 solver.cpp:337] Iteration 0, Testing net (#0)
I0915 09:41:25.588088 4169 solver.cpp:228] Iteration 0, loss = 1436.18
I0915 09:41:25.588129 4169 solver.cpp:244] Train net output #0: dsn1_loss = 22.4332 (* 1 = 22.4332 loss)
I0915 09:41:25.588135 4169 solver.cpp:244] Train net output #1: dsn2_loss = 20.1456 (* 1 = 20.1456 loss)
I0915 09:41:25.588137 4169 solver.cpp:244] Train net output #2: dsn3_loss = 21.0044 (* 1 = 21.0044 loss)
I0915 09:41:25.588141 4169 solver.cpp:244] Train net output #3: dsn4_loss = 15.401 (* 1 = 15.401 loss)
I0915 09:41:25.588145 4169 solver.cpp:244] Train net output #4: dsn5_loss = 12.1914 (* 1 = 12.1914 loss)
I0915 09:41:25.588148 4169 solver.cpp:244] Train net output #5: fuse_loss = 14.5831 (* 1 = 14.5831 loss)
I0915 09:41:25.597915 4169 sgd_solver.cpp:106] Iteration 0, lr = 1e-06

[yun-liu]

The loss is different on different datasets. You can train and test the network for several times to find the secret of convergence for your dataset.