app.py

from flask import Flask, request, jsonify
from PIL import Image
import base64
from io import BytesIO
from transformers import CLIPSegProcessor, CLIPSegForImageSegmentation
import torch
import numpy as np
import matplotlib.pyplot as plt
import cv2

app = Flask(__name__)

processor = CLIPSegProcessor.from_pretrained("CIDAS/clipseg-rd64-refined")
model = CLIPSegForImageSegmentation.from_pretrained("CIDAS/clipseg-rd64-refined")

def process_image(image, prompt, threshold, alpha_value, draw_rectangles):
    inputs = processor(
        text=prompt, images=image, padding="max_length", return_tensors="pt"
    )

    # predict
    with torch.no_grad():
        outputs = model(**inputs)
        preds = outputs.logits

    pred = torch.sigmoid(preds)
    
    if len(pred.shape) == 4:  # Check if the shape is (batch_size, channels, height, width)
        mat = pred[0, 0].cpu().numpy()  # Access the first channel of the first batch
    else:
        mat = pred[0].cpu().numpy()  # If the shape is (channels, height, width)

    mask = Image.fromarray(np.uint8(mat * 255), "L")  # Convert to PIL Image

    # Convert the mask to a NumPy array for calculation
    mask_array = np.array(mask)

    # normalize the mask
    mask_min = mask_array.min()
    mask_max = mask_array.max()
    mask_array = (mask_array - mask_min) / (mask_max - mask_min)

    # threshold the mask
    bmask = mask_array > threshold
    # zero out values below the threshold
    mask_array[mask_array < threshold] = 0

    bmask = Image.fromarray((bmask * 255).astype(np.uint8), "L")

    return bmask




@app.route('/')
def index():
    return "Hello, World! clipseg2"

@app.route('/api/mask_image', methods=['POST'])
def mask_image_api():
    data = request.get_json()

    base64_image = data.get('base64_image', '')
    prompt = data.get('prompt', '')
    threshold = data.get('threshold', 0.4)
    alpha_value = data.get('alpha_value', 0.5)
    draw_rectangles = data.get('draw_rectangles', False)

    # Decode base64 image
    image_data = base64.b64decode(base64_image.split(',')[1])
    image = Image.open(BytesIO(image_data))

    # Process the image
    output_mask = process_image(image, prompt, threshold, alpha_value, draw_rectangles)

    # Convert the output mask to base64
    buffered_mask = BytesIO()
    output_mask.save(buffered_mask, format="PNG")
    result_mask = base64.b64encode(buffered_mask.getvalue()).decode('utf-8')

    return jsonify({'result_mask': result_mask})

if __name__ == '__main__':
    app.run(host='0.0.0.0', port=7860, debug=False)