app.py

import gradio as gr
import joblib
import cv2
import numpy as np
from PIL import Image
from tensorflow.keras.models import load_model

# Define paths to models and load the scaler
model_paths = {
  "Regressor_decision_tree": "multioutput_regressor_decision_tree.joblib",
  "Regressor_ridge": "regressor_ridge.joblib",
  "Regressor_elastic_net": "elastic_net_model.joblib",
  "NN_6_Layers": "NN_Layers_6.keras",
  "CNN": "cnn_model_bigger.keras",
  "CNN_with_reductions": "cnn_model_bigger_with_reductions.keras"
}
scaler = joblib.load("scaler.joblib")

# Function to load models based on file extension
def load_model_by_type(path):
  if path.endswith('.joblib'):
    return joblib.load(path)
  elif path.endswith('.keras'):
    return load_model(path)
  else:
    raise ValueError(f"Unsupported file extension for file {path}")

# Load models with appropriate method
models = {name: load_model_by_type(path) for name, path in model_paths.items()}


def detect_objects(image, model_name):
  model = models[model_name]
  
  # Assuming Gradio passes image as a numpy array and checking if conversion is needed
  if image.ndim == 3 and image.shape[2] == 3:  # If the image is RGB
    image_gray = cv2.cvtColor(image, cv2.COLOR_RGB2GRAY)  # Convert to grayscale using OpenCV
  else:
    image_gray = image  # Use the image as is if already grayscale
  
  # Check if the model requires CNN specific preprocessing
  if model_name in ["CNN", "CNN_with_reductions"]:
    image_processed = np.array(image_gray)
    image_processed = image_processed.reshape(1, image_gray.shape[0], image_gray.shape[1], 1)
    image_processed = image_processed.astype('float32')
    image_processed /= 255 # Normalize pixel values
  else:
    # Assuming other models might expect flattened, scaled input
    image_processed = image_gray.flatten().reshape(1, -1)
    image_processed = scaler.transform(image_processed)

  # Make prediction
  predictions = model.predict(image_processed)
  x, y, width, height = predictions[0]

  # Draw bounding box on a copy of the original image (converted back to RGB for color drawing)
  original_image_rgb = cv2.cvtColor(image, cv2.COLOR_RGB2BGR)  # Ensure image is in RGB
  cv2.rectangle(original_image_rgb, (int(x), int(y)), (int(x + width), int(y + height)), (0, 255, 0), 2)

  return Image.fromarray(original_image_rgb)

# Gradio interface setup
iface = gr.Interface(
  fn=detect_objects,
  inputs=[gr.components.Image(), gr.components.Dropdown(list(model_paths.keys()))],
  outputs=gr.components.Image(),
  title="Object Detection",
  description="Select a model and upload an image to detect objects."
)

iface.launch(show_error=True, share=True, debug=True)