ARC-stuff / arc-vit /export_task_to_image.py
Alignment-Lab-AI's picture
Upload folder using huggingface_hub
d5bfab8 verified
import arc_json_model as ajm
import numpy as np
from simple_image import image_new, set_pixel, draw_rect, draw_box
# ARC images are maximum 30x30. With a 1px border around, then it's the IMAGE_SIZE 32x32.
IMAGE_SIZE = 32
# The color codes [0..9] are the ARC colors
# Color used when expanding the input/output images to size 30x30
COLOR_OUTSIDE = 10
# Color used for a 1px wide box around the input/output images
COLOR_PADDING = 11
# Color that indicates the single pixel that is to be predicted
# Color that indicates the box with the prediction areas
COLOR_HIGHLIGHT = 12
class ExportTaskToImage:
def __init__(self, task: ajm.Task):
self.task = task
self.pixels = ExportTaskToImage.image_from_task(task)
@classmethod
def image_from_task(cls, task: ajm.Task) -> np.ndarray:
# each ARC pair is IMAGE_SIZE wide
# in total it's all the pairs * IMAGE_SIZE
image_width = IMAGE_SIZE * len(task.pairs)
# input images in the top row
# output images in the bottom row
# in total there are 2 rows, thus the image_height is IMAGE_SIZE * 2
image_height = IMAGE_SIZE * 2
pixels = image_new(image_width, image_height, COLOR_OUTSIDE)
# copy input images
for pair_index, pair in enumerate(task.pairs):
for row_index, rows in enumerate(pair.input.pixels):
for column_index, pixel in enumerate(rows):
x = IMAGE_SIZE * pair_index + 1 + column_index
y = IMAGE_SIZE * 0 + 1 + row_index
set_pixel(pixels, x, y, pixel)
# copy output images
for pair_index, pair in enumerate(task.pairs):
for row_index, rows in enumerate(pair.output.pixels):
for column_index, pixel in enumerate(rows):
x = IMAGE_SIZE * pair_index + 1 + column_index
y = IMAGE_SIZE * 1 + 1 + row_index
set_pixel(pixels, x, y, pixel)
# draw boxes around images
for pair_index, pair in enumerate(task.pairs):
x = IMAGE_SIZE * pair_index
draw_box(pixels, x, 0, IMAGE_SIZE, IMAGE_SIZE, COLOR_PADDING)
draw_box(pixels, x, IMAGE_SIZE, IMAGE_SIZE, IMAGE_SIZE, COLOR_PADDING)
# mask out the test output areas
# and draw a highlighted box around the test output areas
for pair_index, pair in enumerate(task.pairs):
if pair.pair_type != ajm.PairType.TEST:
continue
x = IMAGE_SIZE * pair_index
width = pair.output.pixels.shape[1]
height = pair.output.pixels.shape[0]
draw_rect(pixels, x, IMAGE_SIZE, width + 2, height + 2, COLOR_OUTSIDE)
draw_box(pixels, x, IMAGE_SIZE, width + 2, height + 2, COLOR_HIGHLIGHT)
return pixels
def image_with_mark(self, test_index: int, x: int, y: int) -> np.ndarray:
pixels = self.pixels.copy()
for pair_index, pair in enumerate(self.task.pairs):
if pair.pair_type != ajm.PairType.TEST:
continue
if pair.pair_index != test_index:
continue
set_x = IMAGE_SIZE * pair_index + 1 + x
set_y = IMAGE_SIZE + 1 + y
set_pixel(pixels, set_x, set_y, COLOR_HIGHLIGHT)
return pixels
if __name__ == '__main__':
filename = 'testdata/af902bf9.json'
task = ajm.Task.load(filename)
#print(task)
exporter = ExportTaskToImage(task)
#print(exporter)
pixels = exporter.image_with_mark(0, 3, 3)
print("pixels.shape", pixels.shape)