search

gustavla / autocolorize

Automatic colorizaton of grayscale images using Deep Learning
BSD 3-Clause "New" or "Revised" License
220 stars 65 forks source link

Unable to train from scratch #9

Open kiminannu opened 7 years ago

kiminannu commented 7 years ago

Hi,

Firstly, I want to congratulate you for this brilliant piece of work. I have tried your autocolorize and it works great ! Many thanks for making it publicly available.

I am trying to work on NIR images. Idea is to convert this single channel image to a color image. Since the nature of image is different from a normal grayscale image, I have to train the system from scratch / fine-tune your model. I followed your simple training instructions and I guess I am successful in copying the required files and setting up the training environment. I have setup the training without 'extended data layer'.

My input is LMDB (converted .jpg files to LMDB using create_imagenet.sh) and it contains 3 channel 640x480. When I run the training in CPU-only mode, I get the following error:

I0905 16:48:48.431213 14444 net.cpp:434] sr_x_minus_one <- sr_minus_one I0905 16:48:48.431243 14444 net.cpp:408] sr_x_minus_one -> sr_x_minus_one F0905 16:48:48.431275 14444 eltwise_layer.cpp:34] Check failed: bottom[i]->shape() == bottom[0]->shape()

I apologize if this is a trivial issue. I am not much experienced in CAFFE, thus, I request you to kindly help me to figure out the possible issue and get the training running.

Thank you in advance, Kimi

kiminannu commented 7 years ago

Can anyone please help me to solve the above issue?

alwyz commented 6 years ago

I am also having the same problem. Did you ever get it sorted? Can anyone help? I've tried everything - resizing images to every possible combination of 1/3 size and every other trick/tip ive read online. I'm convinced this script is broken, but hopefully the author can chime in and give some clarity as to whats going on.