icTrainer is a python module which allows users to train image classifier easily

Basically, this module is for python3

Install

$ pip install ictrainer

Also you can install manually.

clone repo
$ git https://github.com/koji/icTrainer.git
$ cd icTrainer/ictrainer
$ python setup.py install

How to Use

In this gude, we will create a dog/cat image classifier.

1.Collect Images

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.

2. Resize images

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

3.Create folders for classes

This step, we'll need to create folders and distribute images to train & validation folder.

3-1. create folders

Create a couple of folders under dataset.
This step will be automated in the future.

 dataset
    ├── train
    │   ├── cat
    │   └── dog
    └── val
        ├── cat
        └── dog

3-2. distribute images

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.

4.Train Images

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_"

5 Face detection

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

video

image collecting mode

https://www.youtube.com/watch?v=k5r_xrW_cxE

pre-train model

smart device

https://github.com/koji/icTrainer/blob/master/model/smartdevice_epoch30.h5

classes = ['echo', 'echoplus', 'echoshow', 'googlehome', 'googlehomemini', 'nest']