silburt / deepmoon Goto Github PK

Can you provide Mercury DEM images?Thanks

Chase Roberts created mltest for some simple tests of convnet integrity self-consistency. His code works for TensorFlow, and we support a version of Keras that can't auto-convert to TF, so perhaps we can rewrite his test suite for Keras use.

The way you are cropping the images randomly to generate the datasets. Does it generates data leakage? Thanks

@cczhu since you changed get_unique_craters.py there is now a bug in there somewhere. I ran a new set of extracting crater distributions and there are tons of nan entries in the long/lat coordinates, e.g.

[        nan         nan  4.65171524]
[       nan        nan  3.4377277]
[        nan         nan  7.02831142]
[        nan         nan  3.58227701]

not sure yet at the moment where the problem is, but you might better understand where it could be coming from? My guess is that it has to do with taking np.sin or np.cos of a number not in [-1,1] range, but I actually don't know at the moment. If you don't think you have time to look into this then I will.

It might ultimately be easier for me to look at it since I have the get_unique_craters.py pipeline set up and could just print everything every time a nan shows up....

I think that the train/dev/test datasets belong on zenodo (and provide a shareable link in the DeepMoon readme and paper), but I think a final CNN-predicted crater distribution (for train/dev/test) belongs in this repo (in csv format), in the same folder as our LROC and Head datasets. What do you think @cczhu?

I ran pytest test_get_unique_craters.py in my Python 2 and Python 3 virtualenvs, and in the Python 3 one it template_match_t2c returned N_match == 0 under the test_extract test. Not obvious why this is happening on a quick inspection of the function.

The loss often reaches "Nan" when I run the run_model_train, but I haven't found a solution yet. Have you ever run into this situation?

Hi, I just found there are some difference between the equation and the code in the function estimate_longlatdiamkm from get_unique_craters.py. That is , how to convert the pixel coordinate into latitude/longitude coordinate. Specifically, why there is sin and cos here.

    lat_deg = lat_central - (deg_per_pix * (lat_pix - dim[1] / 2.) *
                             (np.pi * latdiff / 180.) /
                             np.sin(np.pi * latdiff / 180.))
    # Determine longitude using determined latitude.
    long_deg = long_central + (deg_per_pix * (long_pix - dim[0] / 2.) /
                               np.cos(np.pi * lat_deg / 180.))

When I use "using zenodo data.ipynb", I meet this error

And I don't know how can i modify this.

Charles - Py 2.7/3.5
Ari - Keras & TF

Could you tell me how to use "USGS Astrogeology Cloud Processing service"? Thank you very much for your kind consideration and I am looking forward to your early reply.