You cannot select more than 25 topics
Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
28 lines
1.3 KiB
Plaintext
28 lines
1.3 KiB
Plaintext
how to train:
|
|
just run: python train.py with config_params.json
|
|
|
|
|
|
format of ground truth:
|
|
|
|
Lables for each pixel is identified by a number . So if you have a binary case n_classes should be set to 2 and
|
|
labels should be 0 and 1 for each class and pixel.
|
|
In the case of multiclass just set n_classes to the number of classes you have and the try to produce the labels
|
|
by pixels set from 0 , 1 ,2 .., n_classes-1.
|
|
The labels format should be png.
|
|
|
|
If you have an image label for binary case it should look like this:
|
|
|
|
Label: [ [[1 0 0 1], [1 0 0 1] ,[1 0 0 1]], [[1 0 0 1], [1 0 0 1] ,[1 0 0 1]] ,[[1 0 0 1], [1 0 0 1] ,[1 0 0 1]] ]
|
|
this means that you have an image by 3*4*3 and pixel[0,0] belongs to class 1 and pixel[0,1] to class 0.
|
|
|
|
training , evaluation and output:
|
|
train and evaluation folder should have subfolder of images and labels.
|
|
And output folder should be empty folder which the output model will be written there.
|
|
|
|
patches:
|
|
|
|
if you want to train your model with patches, the height and width of patches should be defined and also number of
|
|
batchs (how many patches should be seen by model by each iteration).
|
|
In the case that model should see the image once, like page extraction, the patches should be set to false.
|
|
|