test.py

import os
import subprocess
import torch
import requests
from PIL import Image
from io import BytesIO

is_production = True

os.chdir("/repository" if is_production else ".")
os.environ['AM_I_DOCKER'] = 'False'
os.environ['BUILD_WITH_CUDA'] = 'True'
os.environ['CUDA_HOME'] = '/usr/local/cuda-11.7/' if is_production else '/usr/local/cuda-12.1/'

# Install Segment Anything
subprocess.run(["python", "-m", "pip", "install", "-e", "segment_anything"])

# Install Grounding DINO
subprocess.run(["python", "-m", "pip", "install", "-e", "GroundingDINO"])

subprocess.run("wget https://huggingface.co/Uminosachi/sam-hq/resolve/main/sam_hq_vit_h.pth -O ./sam_hq_vit_h.pth", shell=True)

# Install diffusers
subprocess.run(["pip", "install", "--upgrade", "diffusers[torch]"])

# Install osx
subprocess.run(["git", "submodule", "update", "--init", "--recursive"])
subprocess.run(["bash", "grounded-sam-osx/install.sh"], cwd="grounded-sam-osx")

# Install RAM & Tag2Text
subprocess.run(["git", "clone", "https://github.com/xinyu1205/recognize-anything.git"])
subprocess.run(["pip", "install", "-r", "./recognize-anything/requirements.txt"])
subprocess.run(["pip", "install", "-e", "./recognize-anything/"])

from grounded_sam_demo import grounded_sam_demo
import numpy as np
from PIL import Image
from scipy.ndimage import convolve
from scipy.ndimage import binary_dilation


def get_sd_mask(color_mask_pil, target=(72, 4, 84), tolerance=50):
    image_array = np.array(color_mask_pil)

    # Update target based on the number of color channels in the image array
    target = np.array(list(target) + [255] *
                      (image_array.shape[-1] - len(target)))

    mask = np.abs(image_array - target) <= tolerance
    mask = np.all(mask, axis=-1)

    new_image_array = np.ones_like(image_array) * 255  # Start with white
    # Apply black where condition met
    new_image_array[mask] = [0] * image_array.shape[-1]

    return Image.fromarray(new_image_array)


def expand_white_pixels(input_pil, expand_by=1):
    img_array = np.array(input_pil)
    is_white = np.all(img_array == 255, axis=-1)

    kernel = np.ones((2*expand_by+1, 2*expand_by+1), bool)
    expanded_white = binary_dilation(is_white, structure=kernel)

    expanded_array = np.where(expanded_white[..., None], 255, img_array)

    expanded_pil = Image.fromarray(expanded_array.astype('uint8'))
    return expanded_pil


config_file = "GroundingDINO/groundingdino/config/GroundingDINO_SwinT_OGC.py"
grounded_checkpoint = "groundingdino_swint_ogc.pth"
sam_checkpoint = "sam_hq_vit_h.pth"


def just_get_sd_mask(input_pil, text_prompt, padding):
    print("Doing sam")

    colored_mask_pil = grounded_sam_demo(
        input_pil, config_file, grounded_checkpoint, sam_checkpoint, text_prompt)

    print("doing to white")

    sd_mask_pil = get_sd_mask(colored_mask_pil)

    print("expanding white pixels")

    sd_mask_withpadding_pil = expand_white_pixels(sd_mask_pil, padding)

    return sd_mask_withpadding_pil