app.py

from fastai.vision.all import *
import gradio as gr
from colorthief import ColorThief
from PIL import Image
import matplotlib.colors as mcolors
import io
import colorsys

# Load the model
learn = load_learner('outfit_recommender_resnet18.pkl')

# Define class names
class_names = ['jeans', 'top']

# Helper function to get the name of the closest color
def closest_color(requested_color):
    min_colors = {}
    for key, name in mcolors.CSS4_COLORS.items():
        r_c, g_c, b_c = mcolors.hex2color(name)
        if isinstance(requested_color, str):
            r_r, g_r, b_r = mcolors.hex2color(requested_color)
        else:
            r_r, g_r, b_r = requested_color
        rd = (r_c - r_r) ** 2
        gd = (g_c - g_r) ** 2
        bd = (b_c - b_r) ** 2
        min_colors[(rd + gd + bd)] = key
    return min_colors[min(min_colors.keys())]

def get_dominant_color(image):
    img_byte_arr = io.BytesIO()
    image.save(img_byte_arr, format='PNG')
    img_byte_arr = img_byte_arr.getvalue()
    
    color_thief = ColorThief(io.BytesIO(img_byte_arr))
    dominant_color = color_thief.get_color(quality=1)
    dominant_color = tuple(c / 255 for c in dominant_color)  # Normalize RGB values to [0, 1]
    return dominant_color

def get_monochromatic_palette(color_name, num_colors=5):
    rgb_color = mcolors.to_rgb(mcolors.CSS4_COLORS[color_name])
    h, s, v = colorsys.rgb_to_hsv(*rgb_color)
    
    palette = []
    for i in range(num_colors):
        # Vary both saturation and value
        new_s = max(0, min(1, s + (i - num_colors // 2) * 0.1))
        new_v = max(0, min(1, v + (i - num_colors // 2) * 0.1))
        
        new_rgb = colorsys.hsv_to_rgb(h, new_s, new_v)
        palette.append(closest_color(new_rgb))
    
    # Remove duplicates while preserving order
    return list(dict.fromkeys(palette))

def get_complementary_color(rgb_color):
    h, s, v = colorsys.rgb_to_hsv(*rgb_color)
    complementary_h = (h + 0.5) % 1.0
    r, g, b = colorsys.hsv_to_rgb(complementary_h, s, v)
    complementary_color = closest_color((r, g, b))
    
    # Get monochromatic palette of the complementary color
    complementary_palette = get_monochromatic_palette(complementary_color)
    
    # Ensure the main complementary color is first in the list
    if complementary_color in complementary_palette:
        complementary_palette.remove(complementary_color)
    complementary_palette.insert(0, complementary_color)
    
    return complementary_color, complementary_palette

def get_outfit_recommendation(pred_class):
    if pred_class == 'top':
        return 'Jeans'
    elif pred_class == 'jeans':
        return 'Top'
    else:
        return 'Item'

def predict(image):
    pred_class, pred_idx, outputs = learn.predict(image)
    dominant_color = get_dominant_color(image)
    complementary_color, complementary_palette = get_complementary_color(dominant_color)
    garment_recommendation = get_outfit_recommendation(pred_class)
    
    # Construct output string
    output = f"For your {pred_class}, consider pairing it with a {garment_recommendation.lower()} in {', '.join(complementary_palette)}"
    
    return output  # Return the formatted output

def gradio_predict(image):
    if isinstance(image, np.ndarray):
        image = Image.fromarray(image.astype('uint8'), 'RGB')
    return predict(image)

interface = gr.Interface(
    fn=gradio_predict,
    inputs=gr.Image(),
    outputs="text",
    title="Outfit Recommender(Jeans/Tops)",
    description="Upload an image of jeans or a top to get a recommendation for a complementary outfit based on fashion theory."
)

interface.launch(share=True)