What is the requirements for the data folder about digits HOT 13 CLOSED

So, there's a little tooltip next to the input field that says this:

Indicate a folder which holds subfolders full of images. Each subfolder should be named according to the desired label for the images that it holds. Can also be a URL for an apache/nginx auto-indexed folder.

That's admittedly a little cryptic. I used to have more detailed instructions on there, but they were cluttering up the page. Here is the folder layout that DIGITS is looking for:

/path/to/images/
├── cat
│   ├── 1.jpg
│   └── 2.JPEG
└── dog
    ├── 01.png
    └── 02.jpeg

So, you would enter /path/to/images in the "training images" field, and DIGITS would find two subfolders, "cat" and "dog," and use those as category names. Each category would have two images in it. Obviously, you'll want more than two images per category 😉

from digits.

@ducminhnguyen, were you able to build your dataset?

from digits.

I have try to layout the folder as you said, digit detect the folder and the number of image right but when it starts building the dataset file the same error occurs.

from digits.

Ok, then there might be something else going wrong. I'm filing this as a bug for poor error-catching. I have plans for improving the error catching in this area, so that you get a better idea of what's going wrong.

In the meantime, what is your Pillow version? Run this: pip list | grep pillow -i. You should see that Pillow is at least 2.3.0. I've seen someone with an outdated version run into an issue like yours.

from digits.

This is the bug I was getting.
I am running Pillow (2.3.0)

from digits.

I have also try to add data from the Mnist example in the web installer but it get the same error

from digits.

I finnally manage to import my files. The problem with it is that I use the original caffe for digits when I switch to NVIDIA/caffe it works perfectly.
Caffe is changing alot, do you have any plan to flow up on the changes or just keep the NVIDIA/caffe as it is now

from digits.

Is it possible to create dataset from .bmp images?
I've just tried, but failed with error: "ERROR: Did not find two valid categories".

from digits.

Is it possible to create dataset from .bmp images?

Image formats other than JPEG and PNG are not currently supported. But it is trivial to add support for more. I'm opening a new issue for this, as it is unrelated to @ducminhnguyen's problem.

from digits.

Just wondering if 16bit PNGs are also accepted? Thanks.

from digits.

I haven't tried it myself, but DIGITS should load the images just fine. However, they will be converted to 8-bits before saving to the database or passing to caffe for testing. See here:
https://github.com/NVIDIA/DIGITS/blob/v1.1.0/digits/utils/image.py#L125-L126

Is that a problem? Would you like to work with 16-bit data? Have you done that with caffe before?

from digits.

Shouldn't be a huge issue (i think). I am using it for medical images (DICOM), which tend to be 16 bit greyscale. Good to know. I'll probably stick to 8bit.

from digits.

hi, I could load mnist and train ,test that dataset. now I find medical Xray dataset, and build folder train and test and their classes and labels like mnist also it is in the path that mnist is there. but when I want to create dataset in digit, it cant create dataset and show this error message"ERROR: you do not have read access to folder "/home/homay/Xray/train". how can I solve this problem?
Xray is the folder and in xray there are train and test folders. in each of them there are 0and 1 folders that contains images. and in each test and train folder there are labels.txt and train.txt and test.txt.
please help me. :((((((

from digits.