xiaohangzhan / mix-and-match Goto Github PK

For using the original pretrained models (jigsaw, colorize), they have to be modified (because of different BatchNorm/BN layer). Can you share your modification code? Or at least make some clues about how to do it?

"...resize them to a fixed size of 128×128. Then we extract “pool5” feature of these patches from the CNN for later usage. We assign the patches’ unique labels as the central pixel labels using the corresponding label maps. Then we perform the iterative strategy to construct the graph as discussed in the methodology section."

Hi @XiaohangZhan, I'm confused at the part where patches are labeled. After getting a feature at pool5 of size 4x4 from input 128x128, how does one "assign the patches’ unique labels as the central pixel labels using the corresponding label maps" from a ground truth label maps of 4x4?

Thank you.

Hi, when I tried to train the model (sh run_graph.sh), I have a problem: can anyone tell me how to fix it?

I0319 21:39:59.797385 7363 solver.cpp:372] Solving AlexNet
I0319 21:39:59.797394 7363 solver.cpp:373] Learning Rate Policy: step
*** Aborted at 1521466800 (unix time) try "date -d @1521466800" if you are using GNU date ***
PC: @ 0x7fc32322f38d cv::resize()
*** SIGSEGV (@0x1010000) received by PID 7363 (TID 0x7fc32b54aa00) from PID 16842752; stack trace: ***
@ 0x7fc3215e6330 (unknown)
@ 0x7fc32322f38d cv::resize()
@ 0x7fc2cc09911d pyopencv_cv_resize()
@ 0x7fc322604b94 (unknown)
@ 0x7fc322604b19 (unknown)
@ 0x7fc322606c3d (unknown)
@ 0x7fc322606dd0 (unknown)
@ 0x7fc3225db0a3 (unknown)
@ 0x7fc32257fd6d (unknown)
@ 0x7fc3225db0a3 (unknown)
@ 0x7fc32266e5f7 (unknown)
@ 0x7fc3225895d7 (unknown)
@ 0x7fc2c9dbfaef caffe::PythonLayer<>::Forward_cpu()
@ 0x7fc32ad27389 caffe::Net<>::ForwardFromTo()
@ 0x7fc32ad277b7 caffe::Net<>::ForwardPrefilled()
@ 0x7fc32ad12705 caffe::Solver<>::Step()
@ 0x7fc32ad1302f caffe::Solver<>::Solve()
@ 0x407910 train()
@ 0x405db6 main
@ 0x7fc32122ef45 (unknown)
@ 0x40639b (unknown)
@ 0x0 (unknown)
Segmentation fault (core dumped)

Hi, thanks for the interesting work.

Will M&M further boost the performance of supervisied pretrained model?
Or
I wonder what will happen if I intert the M&M training between a supervised classification pretraining phrase and a supervised segmentation fine-tuning phase?

Thanks a lot!