gather and create image dataset for machine learning.
pip install mlimages
Or clone the repository. Then you can execute examples.
If you want to do fine tuning, you can download pretrained model in examples/pretrained by git lfs.
This tool dependes on Python 3.5 that has async/await feature!
Please make python file in your project folder as below.
import mlimages.scripts.gather_command as cmd
if __name__ == "__main__":
ps = cmd.make_parser()
args = ps.parse_args()
cmd.main(args)Confirm the WordnetID on the ImageNet site
Then download it.
python your_script_file.py -p path/to/data/folder -imagenet --wnid n11531193
You can create training data from images data folder.
Please make python file in your project folder as below.
import mlimages.scripts.label_command as cmd
if __name__ == "__main__":
ps = cmd.make_parser()
args = ps.parse_args()
cmd.main(args)Then run it.
python label.py path/to/images/folder --out path/to/training_data.txt
Now, you have images and training_data.txt. But you have to do some pre-processing to train your model. For example...
- resize image
- normalize image
- sometimes change color...
😭
Don't warry. mlimages supports you!
from mlimages.model import LabelFile, ImageProperty
lf = LabelFile("path/to/training_data.txt", img_root="path_to_your_image_folder")
prop = ImageProperty(width=32, gray_scale=True)
td = lf.to_training_data(prop)
td.make_mean_image("path/to/mean_image") # store mean image to normalize
for d in td.generate():
# d is numpy array that adjusted according to ImageProperty, and normalized by mean_image!
# only you have to do is train the model by it!
print(d)And also, you can restore the image from data.
image = td.result_to_image(numpy_array, label)