app.py

from io import BytesIO
import gradio as gr
import tensorflow as tf
import logging
import requests
import numpy as np
from PIL import Image
from tensorflow.keras.utils import CustomObjectScope
from tensorflow.keras.layers.experimental.preprocessing import RandomHeight

logging.basicConfig(
    level=logging.WARNING,  
    format="%(asctime)s - %(levelname)s - %(message)s",  
)
with CustomObjectScope({'RandomHeight': RandomHeight}):
    model_0 = tf.keras.models.load_model('bestmodel.h5')

    
print("TensorFlow version:", tf.__version__)
print("GPU Available:", tf.config.list_physical_devices('GPU'))

# Configure TensorFlow to use memory growth
gpus = tf.config.experimental.list_physical_devices('GPU')
if gpus:
    try:
        for gpu in gpus:
            tf.config.experimental.set_memory_growth(gpu, True)
    except RuntimeError as e:
        print(e)

def clean_url(input_string):
    if "http" in input_string:
        return input_string[input_string.find("http"):]
    return input_string


def classify_image(inp):
    # Convert to PIL Image if we have a numpy array
    image = None
    logging.warning("entree dans ckassify_image")
    logging.warning(inp)
    if isinstance(inp, np.ndarray):
        image = Image.fromarray(inp)
        logging.warning("1")

    if isinstance(inp, str) :
        logging.warning("2")

        inp = clean_url(inp)
        response = requests.get(inp)
        response.raise_for_status()
        image = Image.open(BytesIO(response.content))
    if isinstance(inp, str) and (inp.startswith("http://") or inp.startswith("https://")):
        logging.warning("3")
        response = requests.get(inp)
        response.raise_for_status()
        image = Image.open(BytesIO(response.content))
       
    else:
              logging.warning("4")

              image = inp
    if image.mode != "RGB":
        image = image.convert("RGB")
    # Resize image to 224x224
    image = image.resize((224, 224), Image.Resampling.LANCZOS)
    
    # Convert to numpy array and ensure correct shape
    image_array = np.array(image)
    
    # Handle grayscale images
    if len(image_array.shape) == 2:
        image_array = np.stack([image_array] * 3, axis=-1)
    
    # Add batch dimension and ensure correct shape
    inp = image_array.reshape((-1, 224, 224, 3))
    # inp = inp.reshape((-1, 224, 224, 3))
    prediction = model_0.predict(inp)
    output  = ""
    if prediction[0][prediction.argmax()] < 0.84:
      output = "bonne image"
    elif prediction.argmax() == 0:
      output = "Rifle violence"
    elif prediction.argmax() == 1:
      output = "guns violence"
    elif prediction.argmax() == 2:
      output = "knife violence"
    elif prediction.argmax() == 3:
      output = "image porno"
    elif prediction.argmax() == 4:
      output = "personne habillée" 
    else:
      output = "tank violence" 
    return output


gr.Interface(
    fn=classify_image, inputs=gr.Text(label="Url de l image"), outputs="text",live=True,title="API de détection des images violentes",
).launch(share=True)