 Stephen
|
116af0efce
This version makes the correctly predicted white pixels get a red mask image saved to file. Good results start at 3 epochs in terms of redness.
|
2 years ago |
|
Stephen
|
ad4a3d2608
I made a hack to fix the boundary box sizes. The PILLOW function image.paste(im, sizewhatever) seems to be bugged? Or maybe my pictures are strange in some way. To be clear, occasionally image.getbbox() returns something other than (0, 0, width, height), (width, height) come from image.size.
|
2 years ago |
|
Stephen
|
a63b55f1ef
Currently getting a weird error where the image.getbbox() function returns a different size than image.size; BUT ONLY FOR INVERTED IMAGES!
|
2 years ago |
|
Stephen
|
551469d2d8
Ok, the Autoinit thing was messing things up. Working again.
|
2 years ago |
|
Stephen
|
e87dda9966
Something is very wrong. The program no longer creates good outputs.
|
2 years ago |
|
Stephen
|
06e0d9860f
Commented out autoinit for now. About to start going ham on this code.
|
2 years ago |
|
Stephen
|
a52787b597
Added note to remind myself to change the background setting after I fix the foreground/background issue.
|
2 years ago |
|
Stephen
|
c33ab62fac
Added the AutoInit thingy. Tested to see it work. Must do tests to see what effect it has.
|
2 years ago |
|
Stephen
|
007429a02e
Implemented a function to create a mask of the prediction with the truth-- Areas in both are shown. Also, I discovered that my model is technically finding the background instead of the foreground. It shouldn't matter much technically, but it explains some of the weird issues I've been having.
|
2 years ago |
|
Stephen
|
80e5619d56
Remembered to use the clever way of multiplication elementwise.
|
2 years ago |
|
Stephen
|
dabe7d428f
Created mask function to output a mask of the whole predicted images.
|
2 years ago |
|
Stephen
|
2952a1b013
Added jaccard and dice accuracy score calculations for the predicted full sized images.
|
2 years ago |
|
Stephen
|
2398085d44
Found the error. There was a high in the wide loop. line 595. lol. This version works to produce binarized images of any size!
|
2 years ago |
|
Stephen
|
1a9fe5c301
Ok, this one is making the right looking output but maybe in the wrong order. testing.
|
2 years ago |
|
Stephen
|
cbdf311d6c
Found a way that is outputing the right-ish size using slicing. Almost got it. Testing this change on 596 now.
|
2 years ago |
|
Stephen
|
aa36014a2b
I found that on line 234, I accidentally put NUM_SQUARES instead of the size of the squares. So all my previous methods to image output may have worked fine...
|
2 years ago |
|
Stephen
|
2c9a151757
My outputs are long images containing all the squares. They are not put into rows correctly; just one big column.
|
2 years ago |
|
Stephen
|
652421dd32
Looks like the images are not actually stitched back together yet. I see that the first one is shape 16,64,64,1 so it must still be in squares.
|
2 years ago |
|
Stephen
|
43aae200fb
As of now this "works". It no longer crashes with a ton of weird errors. But the output of the whole image predict is just the original image.
|
2 years ago |
|
Stephen
|
bb39871212
Added making a copy of the conda environment. Added another graph output. Removed many print statements and old comments.
|
2 years ago |
|
Stephen
|
505880dc79
Changed to use same number of test squares as training squares. I had previously used that global variable for the size of the validation set before I found out you could do valdation splitting automatically in keras. Also made the number of test squares max out at the actual number of squares. Also, reduced number of example pictures to save to a small number randomly picked from the prediction outputs. Need to remove prints and get rid of all global variables. They are becoming a headache, I thought there would only be one or two.
|
2 years ago |
|
Stephen
|
9a2b6317b0
Working version of jaccard and dice output. Text file output needs to be cleaned up.
|
2 years ago |
|
Stephen
|
57cc83e02b
Added a test of the jaccard and dice functions.
|
2 years ago |
|
Stephen
|
50ae25bbdf
Added smoothing coefficient.
|
2 years ago |
|
Stephen
|
e8e9fcfb25
Added jaccard and dice functions.
|
2 years ago |
|
Stephen
|
f7518c14c8
Bade 4 image saves, before binarization out, binarized out, test truth, test binarized. Also added randseeds.
|
2 years ago |
|
Stephen
|
39e850a7b5
Added dropout back in. I compared with and without, and dropout increased the accuracy of the outputs by visual inspection AND by accuracy scores.
|
2 years ago |
|
Stephen
|
71e14afb4a
Added conda environment. Reverted some recent changes. They were making th model output garbage during testing sized runs. They might be useful later but not now.
|
2 years ago |
|
Stephen
|
0235b44eff
Added making a copy of the source code to the output folder.
|
2 years ago |
|
Stephen
|
adb3d20f68
Fixed bug where entering different square size would crash.
|
2 years ago |
