Basically, this module is for python3
$ pip install ictrainer
Also you can install manually.
clone repo
$ git https://github.com/koji/icTrainer.git
$ cd icTrainer/ictrainer
$ python setup.py install
In this gude, we will create a dog/cat image classifier.
https://icrawler.readthedocs.io/en/latest/
$ ictrainer --mode collect --keyword dog -n 250
$ ictrainer --mode collect --keyword cat -n 250
You'll have dogs & cats images under dataset
folder.
In this step, we will change all images size for training. The current input size must be 320 x 180
(required).
This step may be mess up images you collected, so you need to check all images manually. In the furture, there will be a function that save your time.
$ ictrainer --mode resize --target dog
$ ictrainer --mode resize --target cat
For people want to use resize mode for other thing, you can use reize images with the following command. The folder structure should be the same the above.
$ ictrainer --mode resize --target cat --image_width 480 --image_height 320
This step, we'll need to create folders and distribute images to train
& validation
folder.
Create a couple of folders under dataset. This step will be automated in the future.
dataset
train
cat
dog
val
cat
dog
Move images we got via image collect mode
. In this case, probably we have 250 images for each other.
We will put 225 images for train and 25 images for validation so that train/dog
has 225 images and validation/dog
has 25 images. The cats should be the same.
There are some options we need to put. The most important one is --classes
which will be labels. In this case, we have dog & cat, so we need to put them as classes.
--batch
: batch size default 16
--epoch
: epoch default 30
--mname
: output model name
--lr
: learning rate default 1e-3
momentum
: mementum default 0.9
We will use default settings.
$ ictrainer --mode train --classes "cat" "dog" --mname "dogAndcat_"
From 0.2.0 ictrainer allows you to use face detection. The command is following. This function is using OpenCV
Cascade filter to detect front faces. When you successfully run this command, ictrainer creates output
folder and there are faces and something which means still you need check all image by yourself.
--mode face
--target target folder
$ ictrainer --mode face --target dataset/celeb
image collecting mode
https://www.youtube.com/watch?v=k5r_xrW_cxE
smart device
https://github.com/koji/icTrainer/blob/master/model/smartdevice_epoch30.h5
classes = ['echo', 'echoplus', 'echoshow', 'googlehome', 'googlehomemini', 'nest']