### -------------------------------- ###
###            libraries             ###
### -------------------------------- ###
import gradio as gr
import numpy as np
import os
import tensorflow as tf

### -------------------------------- ###
###           model loading          ###
### -------------------------------- ###
model = tf.keras.models.load_model('model.h5')

## --------------------------------- ###
###     reading: categories.txt      ###
### -------------------------------- ###
labels = ['please upload categories.txt' for i in range(10)] # placeholder

if os.path.isfile("categories.txt"):
    # open categories.txt in read mode
    categories = open("categories.txt", "r")
    labels = categories.readline().split()

## --------------------------------- ###
###          page description        ###
### -------------------------------- ###
title = "Seefood: Hot dog or... not hot dog"
description = "A Hugging Space demo created by datasith!"
article = \
'''
#### Hot dog or not hot dog

Jìan-Yang's masterpiece from the show Silicon Valley serves as a great exercise to get 
familiar with Hugging Face spaces! All the necessary files are included for everything 
to run smoothly on HF's Spaces:

- app.py
- categories.txt
- model.h5 (AlexNet)
- requirements.txt
- README.md
- nay.jpg (Not-hot-dog example)
- yay.jpg (Hot-dog example)

The data used to train the model is available as a 
[Kaggle dataset](https://www.kaggle.com/datasets/dansbecker/hot-dog-not-hot-dog).
The step-by-step process for generating, training, and testing the Image Classification model is 
available at my [GitHub respository](https://github.com/datasith/ds-experiments-image-classification/tree/main/hotdog-not-hotdog).

If you enjoy my work feel free to follow me here on HF and/or on:

- [GitHub](https://github.com/datasith)
- [Kaggle](https://kaggle.com/datasith)
- [Twitter](https://twitter.com/datasith)
- [LinkedIn](https://linkedin.com/in/datasith)

Either way, enjoy!
'''

### -------------------------------- ###
###        interface creation        ###
### -------------------------------- ###
samples = ['yay.jpg', 'nay.jpg']

def preprocess(image):
    image = tf.image.resize(image, [256, 256])
    img_array = tf.keras.utils.img_to_array(image)
    img_array = tf.expand_dims(img_array, 0)
    # image = np.array(image) / 255
    # image = np.expand_dims(image, axis=0)
    return img_array
  
def predict_image(image):
    # pred = model.predict(preprocess(image))
    # results = {}
    # for row in pred:
    #     for idx, item in enumerate(row):
    #         results[labels[idx]] = float(item)
    predictions = model.predict(preprocess(image))
    scores = tf.nn.softmax(predictions[0])
    results = {}
    for idx, res in enumerate(scores):
        results[labels[idx]] = float(res)
    return results

# generate img input and text label output
image = gr.inputs.Image(label="Upload Your Image Here")
label = gr.outputs.Label(num_top_classes=len(labels))

# generate and launch interface
interface = gr.Interface(fn=predict_image, inputs=image, 
                         outputs=label, article=article, theme='default', 
                         title=title, allow_flagging='never', description=description, 
                         examples=samples)
interface.launch()