nsymbolize: A PureBasic repository from AndreyCorelli

Set up the project

This project is build in Python 3.8.10 virtual environment. Check that your video card drivers are actual and CUDA is installed. CUDA should not conflict with Pytorch.

Most probably you'll have to install specific version of Pytorch for you CUDA version with a command like

pip3 install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu116

The project requires the library backgroundremover==0.1.9. Most probably the library will require a different pytorch version and there'll be no chance to make this package run inside the project's venv. The solution is to:

open / create another project
install backgroundremover==0.1.9 with all the requirements
copy & paste the file cmd/remove_background.sh in this project, change the paths respectively and run the scrip in the new project's venv.

Vectorize an image

The model used here in example code includes one large file: src/vectorizer/models/model_01-09-2022/variables/variables.data-00000-of-00001. The file is compressed. To uncompress the file run a command

cd cmd && ./unpack_model.sh

from preprocessing.encode_ascii import ASCIIEncoder
from vectorizer.image_vectorizer import ImageVectorizer

# image should have 128x128 size
imv = ImageVectorizer()
imv.load_model("<model folder>")
vector = imv.vectorize_image("<source image path>")
text = ASCIIEncoder().decode_vector(vector)
text_line = "\n".join(["".join(l) for l in text])
with open("<output text file path>", 'w') as f:
    f.write(text_line)

Train and save a model

from vectorizer.image_vectorizer import ImageVectorizer

imv = ImageVectorizer()
imv.build_model()
imv.train("<source image folder, all images are 128x128>", 
          "<folder with text files, matching with images by names>")
imv.save_model("<folder to store the model>")

Prepare the data

Source data images may have any format that cv2.imread() can read. Colorful images are expected. Source image folder is data/src_images.

Collect images from Google

To collect images from Google do the following:

open Google images, scroll to the bottom to show more images on the page
open the Google Chrome console, copy & paste there the script from src/data_collection/google_js.txt
there's a new text file urls.txt with the images' hyperlinks.
repeat steps 1 - 3 to collect several files urls.txt, urls(1).txt, .... Combine the files' content into the single file src/data_collection/urls.txt.
run the script python src/data_collection/download_google_images.py

Make text files out of images

First, remove background from original images: cmd/remove_background.sh. The script will read the source files from data/src_images and store the output data/cln_images. The script will skip all the files that are already processed. Processing 2000 files takes ~ 3 hours.

Second, convert the files from the data/cln_images folder with the cmd/symbolize.sh command. The command will also skip the files that are already converted. The output (Y for training) will be the files in the data/ascii_files folder. Each file has 160 columns and 80 lines as it was set up in the script.