Closed merceyz closed 8 years ago
Ok, so deepcl_unittests
is for running automated unit tests, to check stuff is working ok. To run training you'll want to use deepcl_train
.
Yeah i figured as much, but how do i present it with my data?
Ah, good point .Link should work now https://github.com/hughperkins/DeepCL/blob/master/doc/Loaders.md
That jpeg loader will use RGB right or do i have to specify something for that?
Oh, good question. I think you just need to put planes=3
in the manifest, and then it should work. Can you try ,and see if this works ok?
(I think the only requirements on the images are:
All requirements are met
Works ok?
"Something went wrong: can't open *file name here"
Does it really say '*file name here'? or does that have an actual filepath?
If you look at the first screenshot you can see the actual path
ah...
ok, can you try removing the leading /
, ie 0/620cc66f-f919-...etc..
Same error
Looks like the filepath is relative to the current directory, not to the datadir. That's probably not really intuitive, or correct, but I think that's how it works currently. Can you try using either absolute paths, or paths relative to the current directory? (Might need to add TestFolder/
to the start of each path?)
That fixed it and it puked out a bunch of stuffs
If i'm not mistaken something is wrong tho as "training loss" says "-1.#IND" and all values are the same on each epoch finished
ah yeah, that training loss means someting went wrong...
maybe try a smaller learning rate?
There we go, loss started at ~800 going down more and more
Accuracy on both the training and test keeps climbing so it appears to be working :+1:
It got to epoch 12 and "clblas teardown" showed up, i assume that's it's built in "limit"
Accuracy on both the training and test keeps climbing so it appears to be working
Cool :-)
It got to epoch 12 and "clblas teardown" showed up, i assume that's it's built in "limit"
By default, it trains 12 epochs :-P You can change that using numepochs=100
(or however many epochs yo uwant)
Awesome, thanks.
Suggesting a fix to this:
Looks like the filepath is relative to the current directory, not to the datadir. That's probably not really intuitive, or correct, but I think that's how it works currently. Can you try using either absolute paths, or paths relative to the current directory? (Might need to add TestFolder/ to the start of each path?)
Otherwise i have no more questions for now so you can close this issue if you like (or keep it open as a reminder of the quote above)
I'll close this, since it is more general, but you could open an issue specifically for the filepath perhaps?
(for info, I've fixed the issue with relative path, in the sourcecode. Question: are you using a prebuilt release, or building from source?)
As "requested" in issue #73 i'm moving my question part here
I'm trying to use the CMD API to have a network learn to label some 96x96x3 (RGB) images.
I have a folder structure where each folder for each label has a bunch of images
(water)
I opened cmd and ran the following commands:
This resulted in the following output![970fa81b05292c65274d5321acb3b80c](https://cloud.githubusercontent.com/assets/3842800/17055828/b374dd9e-500f-11e6-922d-042e342f50e5.png)
Now i'm kind of stuck, I was reading on this page at the "File types" section there is a line saying "See Loaders for information on available loaders" which doesn't work, the link leads to an 404 page
So my question/issue is: How do I give it data to train/test on?