# -*- coding: utf-8 -*-
"""HandWrittenDigitRecognition.ipynb

Automatically generated by Colaboratory.

Original file is located at
    https://colab.research.google.com/drive/1_tPuKWVLLeOAiFbtZCBrp7fIsPNE2DTT
"""


"""# Imports"""

import numpy as np
import keras
from keras.models import Sequential
from keras.layers import Dense, Conv2D
from keras.datasets import mnist

# Gradio to build web app for presenting the ML model
import gradio as gr

"""# Loading pre-processed MNIST dataset"""

# mnist.load_data() returns list of tuples
(x_train, y_train), (x_test, y_test) = mnist.load_data()

# print(f"X_train Shape: {x_train.shape} | Type - {type(x_train)}")
# print(f"y_train Shape: {y_train.shape} | Type - {type(y_train)}")
# print(f"X_test Shape: {x_test.shape} | Type - {type(x_test)}")
# print(f"y_test Shape: {y_test.shape} | Type - {type(y_test)}")

"""-> Make modifications to the train and test datasets to suit the **ML model**




"""

# flatten the shapes of x_train and x_test
x_train = x_train.reshape(60000, 28*28)
x_test = x_test.reshape(10000, 28*28)

# TODO: Try without resizing the images and there'll be **significant difference** in accuracy when we compare both the models
x_train = x_train/255.0
x_test = x_test/255.0

"""# ML Model"""

# Simple Neural Network with 3 hidden layers

model = Sequential()

model.add(Dense(512, activation='relu', input_dim=784))
model.add(Dense(256, activation='relu'))
model.add(Dense(128, activation='relu'))
model.add(Dense(10, activation='sigmoid'))

model.compile(optimizer='adam', loss='sparse_categorical_crossentropy', metrics=['accuracy'])

model.fit(x_train, y_train, epochs=20, validation_split=0.3)

"""# Evaluation of the Model"""

results = model.evaluate(x_test, y_test)
# print(f"Accuracy: {100*(results[1])}")

# print(model.predict(x_test[0:1])) # returns list of lists

"""# Visualization using Gradio"""

def predict_digit(img):
  # img is a numpy array of shape 28*28 (it is fetched by gradio by default)
  img=img.reshape(1, 784)
  img = img/255.0
  results_dict = {}
  prediction=model.predict(img)
  predictions = prediction[0]
  for i in range(len(predictions)):
    results_dict[str(i)] = float(predictions[i])
  return results_dict

# Sketchpad - Input - Image of size 28*28 (default)
# Note: gr.outputs.Label(num_top_classes=3) - to filter out top 3 classes 

ui = gr.Interface(predict_digit, inputs="sketchpad", outputs=gr.outputs.Label())
ui.launch(debug='True')