app.py

import gradio as gr
from PIL import Image
import numpy as np
from tensorflow.keras.preprocessing import image as keras_image
from tensorflow.keras.applications.resnet50 import preprocess_input as resnet_preprocess_input
from tensorflow.keras.models import load_model

# Klassenlabels (ersetzen Sie durch Ihre spezifischen Klassen)
class_labels = ['bench_press', 'deadlift', 'hip_thrust', 'lat_pulldown', 'pull_up', 'squat', 'tricep_dips']

# Load your trained models
resnet_model = load_model('exercise_classification_model_resnet50.h5')
cnn_model = load_model('exercise_classification_model.h5')

def predict_exercise(img, model_type):
    try:
        # Convert the image to RGB format
        img = Image.fromarray(img.astype('uint8'), 'RGB')
        # Resize the image to the required input size of the model
        img = img.resize((150, 150))
        # Convert the image to an array
        img_array = keras_image.img_to_array(img)
        # Expand dimensions to match the model's input shape
        img_array = np.expand_dims(img_array, axis=0)
        
        if model_type == 'ResNet50':
            # Preprocess the input as expected by ResNet50
            img_array = resnet_preprocess_input(img_array)
            # Predict using the ResNet50 model
            prediction = resnet_model.predict(img_array)
        elif model_type == 'CNN':
            # Preprocess the input as expected by the CNN
            img_array /= 255.0  # Normalisierung
            # Predict using the CNN model
            prediction = cnn_model.predict(img_array)
        else:
            return {"error": "Invalid model type selected"}
        
        # Debugging: print the prediction shape
        print(f"Prediction shape: {prediction.shape}")
        
        # Check prediction shape
        if prediction.shape[1] == len(class_labels):  # Ensure the prediction matches the number of classes
            # Return the prediction as a dictionary with class probabilities
            return {class_labels[i]: float(prediction[0][i]) for i in range(len(class_labels))}
        else:
            return {"error": f"Unexpected prediction shape: {prediction.shape}"}
    except Exception as e:
        return {"error": str(e)}

# Custom CSS for Dark Mode and Elegant Design
custom_css = """
body {background-color: #121212; color: #e0e0e0; font-family: 'Arial', sans-serif;}
h1 {color: #ff5722;}
label {color: #ff9800;}
input[type=radio] {accent-color: #ff5722;}
button:hover {background-color: #e64a19;}
.footer {display: none !important;}
"""

# Define the Gradio interface
interface = gr.Interface(
    fn=predict_exercise,
    inputs=[
        gr.Image(type="numpy", label="Upload an image of the exercise"),
        gr.Radio(['ResNet50', 'CNN'], label="Choose Model")
    ],
    outputs=gr.Label(num_top_classes=len(class_labels), label="Prediction"),  # Show all classes
    title="Exercise Classifier",
    description="Upload an image of an exercise and the classifier will predict the exercise.",
    css=custom_css  # Apply the custom CSS
)

# Launch the interface
interface.launch()