import numpy as np
import matplotlib.pyplot as plt
import gradio as gr

def plot_secant(h):
    x = np.linspace(-1, 2, 400)
    y = x**2
    m = (h**2) / h
    fig, axs = plt.subplots(1, 2, figsize=(8, 4))
    for ax in axs:
        ax.set_xlim(-1, 2)
        ax.set_ylim(-1, 4)
        ax.set_xticks(np.arange(-1, 3, 1))
        ax.set_yticks(np.arange(-1, 5, 1))
        ax.grid(True, linestyle='--', linewidth=0.5, color='lightgray')
        ax.spines['top'].set_visible(False)
        ax.spines['right'].set_visible(False)
    axs[0].plot(x, y, color='black')
    axs[0].plot([0, h], [0, h**2], color='red', linewidth=2)
    axs[0].scatter([0, h], [0, h**2], color='red', zorder=5)
    axs[1].plot(x, m * x, color='red', linewidth=2)
    plt.tight_layout()
    return fig


def plot_tangent(x0):
    x = np.linspace(-1, 2, 400)
    y = x**2
    m = 2 * x0
    y0 = x0**2
    fig, axs = plt.subplots(1, 2, figsize=(8, 4))
    for ax in axs:
        ax.set_xlim(-1, 2)
        ax.set_ylim(-1, 4)
        ax.set_xticks(np.arange(-1, 3, 1))
        ax.set_yticks(np.arange(-1, 5, 1))
        ax.grid(True, linestyle='--', linewidth=0.5, color='lightgray')
        ax.spines['top'].set_visible(False)
        ax.spines['right'].set_visible(False)
    axs[0].plot(x, y, color='black')
    axs[0].plot(x, m * (x - x0) + y0, color='red', linewidth=2)
    axs[0].scatter([x0], [y0], color='red', zorder=5)
    axs[1].plot(x, 2 * x, color='black')
    axs[1].scatter([x0], [m], color='red', zorder=5)
    plt.tight_layout()
    return fig


def plot_gradient_descent(lr, init_x, steps):
    n = int(steps)
    path = [init_x]
    for _ in range(n):
        path.append(path[-1] - lr * 2 * path[-1])
    xv = np.array(path)
    fig, axs = plt.subplots(1, 2, figsize=(8, 4))
    x_plot = np.linspace(-2, 2, 400)
    axs[0].plot(x_plot, x_plot**2, color='black')
    axs[0].plot(xv, xv**2, marker='o', color='red', linewidth=2)
    for i in range(n):
        axs[0].annotate('', xy=(xv[i+1], xv[i+1]**2), xytext=(xv[i], xv[i]**2), arrowprops=dict(arrowstyle='->', color='red'))
    axs[0].set_xlim(-2, 2)
    axs[0].set_ylim(-0.5, 5)
    axs[0].set_title('Gradient Descent Path')
    axs[0].grid(True, linestyle='--', linewidth=0.5, color='lightgray')
    axs[1].plot(range(n+1), xv, marker='o', color='red', linewidth=2)
    for i in range(n):
        axs[1].annotate('', xy=(i+1, xv[i+1]), xytext=(i, xv[i]), arrowprops=dict(arrowstyle='->', color='red'))
    axs[1].set_xlim(0, n)
    axs[1].set_ylim(xv.min() - 0.5, xv.max() + 0.5)
    axs[1].set_xticks(range(0, n+1, max(1, n//5)))
    axs[1].set_xlabel('Iteration')
    axs[1].set_title('x over Iterations')
    axs[1].grid(True, linestyle='--', linewidth=0.5, color='lightgray')
    plt.tight_layout()
    return fig


def plot_chain_network(x):
    y = 2 * x
    z = 3 * y
    L = 4 * z
    fig, ax = plt.subplots(figsize=(6, 2))
    ax.axis('off')
    pos = {'x': 0.1, 'y': 0.3, 'z': 0.5, 'L': 0.7}
    for name in pos:
        ax.add_patch(plt.Circle((pos[name], 0.5), 0.05, fill=False))
        ax.text(pos[name], 0.5, name, ha='center', va='center')
    for src, dst, lbl in [
        ('x', 'y', r'$\partial y/\partial x=2$'),
        ('y', 'z', r'$\partial z/\partial y=3$'),
        ('z', 'L', r'$\partial L/\partial z=4$')
    ]:
        sx, dx = pos[src], pos[dst]
        ax.annotate('', xy=(dx, 0.5), xytext=(sx, 0.5), arrowprops=dict(arrowstyle='->'))
        ax.text((sx + dx)/2, 0.6, lbl, ha='center', va='center')
    ax.text(0.02, 0.15, r'$\frac{\partial L}{\partial x}=4\times3\times2$', transform=ax.transAxes, ha='left')
    for name, val in [('x', x), ('y', y), ('z', z), ('L', L)]:
        ax.text(pos[name], 0.3, f"{name}={val:.2f}", ha='center')
    plt.tight_layout()
    return fig


def plot_backprop_dnn(x, w1, w2, t):
    a = w1 * x
    y = w2 * a
    L = 0.5 * (y - t)**2
    fig, ax = plt.subplots(figsize=(6, 2))
    ax.axis('off')
    pos = {'x': 0.1, 'a': 0.3, 'y': 0.5, 'L': 0.7}
    for name in pos:
        ax.add_patch(plt.Circle((pos[name], 0.5), 0.05, fill=False))
        ax.text(pos[name], 0.5, name, ha='center', va='center')
    for src, dst, lbl in [
        ('x', 'a', '∂a/∂x = w₁'),
        ('a', 'y', '∂y/∂a = w₂'),
        ('y', 'L', '∂L/∂y = (y - t)')
    ]:
        sx, dx = pos[src], pos[dst]
        ax.annotate('', xy=(dx, 0.5), xytext=(sx, 0.5), arrowprops=dict(arrowstyle='->'))
        ax.text((sx + dx)/2, 0.6, lbl, ha='center', va='center')
    ax.text(0.02, 0.15, '∂L/∂w₂ = (y - t) · a', transform=ax.transAxes, ha='left')
    ax.text(0.02, 0.02, '∂L/∂w₁ = (y - t) · w₂ · x', transform=ax.transAxes, ha='left')
    plt.tight_layout()
    return fig

def update_secant(h): return plot_secant(h), f'**Δx=h={h:.4f}**, slope={(h**2)/h:.4f}'

def update_tangent(x0): return plot_tangent(x0), f'**x={x0:.2f}**, dy/dx={2*x0:.2f}'

def update_gd(lr, init_x, steps): return plot_gradient_descent(lr, init_x, steps), f'lr={lr:.2f}, init={init_x:.2f}, steps={int(steps)}'

def update_chain(x): return plot_chain_network(x), f"**Values:** y={2*x:.2f}, z={3*(2*x):.2f}, L={4*(3*(2*x)):.2f}\n**dL/dx=24**"

def update_bp(x, w1, w2, t): return plot_backprop_dnn(x, w1, w2, t), ''

def load_secant(): return plot_secant(0.01), '**Hint:** try moving the slider!'

def load_tangent(): return plot_tangent(0.0), '**Hint:** try moving the slider!'

def load_gd(): return plot_gradient_descent(0.1, 1.0, 10), '**Hint:** try moving the sliders!'

def load_chain(): return plot_chain_network(1.0), '**Hint:** try moving the slider!'

def load_bp(): return plot_backprop_dnn(0.5, 1.0, 1.0, 0.0), '**Hint:** try moving the sliders!'

def reset_all(): return (
    gr.update(value=0.01), gr.update(value=0.0), gr.update(value=0.1),
    gr.update(value=1.0), gr.update(value=10),
    gr.update(value=1.0), gr.update(value=0.5), gr.update(value=1.0),
    gr.update(value=1.0), gr.update(value=0.0)
)


demo = gr.Blocks()
with demo:
    gr.HTML('<div style="border:1px solid lightgray; padding:10px; border-radius:5px">')
    with gr.Tabs():
        with gr.TabItem('Secant Approximation'):
            gr.Markdown('''
            **Feature:** Explore secant line approximations via Δx/h
            - Draws a chord between (x, x²) and (x+h, (x+h)²)
            - As h → 0, the chord converges to the true tangent
            - Goal: See how (f(x+h)-f(x)) / h approaches the instantaneous derivative
            ''')
            h = gr.Slider(0.001, 1.0, value=0.01, step=0.001, label='h')
            p1 = gr.Plot()
            m1 = gr.Markdown()
            h.change(update_secant, [h], [p1, m1])
        with gr.TabItem('Tangent Visualization'):
            gr.Markdown('''
            **Feature:** Visualize tangent line and instantaneous slope
            - Draws the exact tangent at (x, x²)
            - Slide x to update both the red line and its numeric slope
            - Goal: Grasp the derivative as the “best‐fit” rate of change at one point
            ''')
            x0 = gr.Slider(-1.0, 2.0, value=0.0, step=0.1, label='x')
            p2 = gr.Plot()
            m2 = gr.Markdown()
            x0.change(update_tangent, [x0], [p2, m2])
        with gr.TabItem('Gradient Descent'):
            gr.Markdown('''
            **Feature:** Observe gradient descent on y=x²
            - Treats the curve as a valley; you take downhill steps
            - Left: path on the curve with arrows; Right: x vs. iteration
            - Goal: Feel how step size (learning rate) affects speed and stability toward the minimum
            ''')
            lr = gr.Slider(0.01, 0.5, value=0.1, step=0.01, label='Learning Rate')
            init = gr.Slider(-2.0, 2.0, value=1.0, step=0.1, label='Initial x')
            st = gr.Slider(1, 50, value=10, step=1, label='Iterations')
            pg = gr.Plot()
            mg = gr.Markdown()
            for inp in [lr, init, st]: inp.change(update_gd, [lr, init, st], [pg, mg])
        with gr.TabItem('Chain Rule'):
            gr.Markdown('''
            **Feature:** Demonstrate the chain rule via a computation graph
            - Nodes: x → y=2x → z=3y → L=4z with arrows showing ∂→
            - Multiply local slopes 2 × 3 × 4 = 24 for overall dL/dx
            - Goal: Visualize why and how partial derivatives combine to yield the total derivative
            ''')
            xs = gr.Slider(0.0, 2.0, value=1.0, step=0.1, label='x')
            cp = gr.Plot()
            cm = gr.Markdown()
            xs.change(update_chain, [xs], [cp, cm])
        with gr.TabItem('Backpropagation'):
            gr.Markdown('''
            **Feature:** Visualize backpropagation derivatives in a simple DNN
            - Network: x → a=w₁x → y=w₂a → L=½(y–t)² with local ∂ annotations
            - Tweak x, w₁, w₂, or t and watch error gradients flow backward
            - Goal: Demystify how output error propagates to compute each weight’s gradient
            ''')
            xb = gr.Slider(-2.0, 2.0, value=0.5, step=0.1, label='x')
            w1b = gr.Slider(-2.0, 2.0, value=1.0, step=0.1, label='w1')
            w2b = gr.Slider(-2.0, 2.0, value=1.0, step=0.1, label='w2')
            tb = gr.Slider(-2.0, 2.0, value=0.0, step=0.1, label='t')
            pb = gr.Plot()
            mb = gr.Markdown()
            for inp in [xb, w1b, w2b, tb]: inp.change(update_bp, [xb, w1b, w2b, tb], [pb, mb])
  
    demo.load(load_secant, [], [p1, m1])
    demo.load(load_tangent, [], [p2, m2])
    demo.load(load_gd, [], [pg, mg])
    demo.load(load_chain, [], [cp, cm])
    demo.load(load_bp, [], [pb, mb])
 
    with gr.Row():
        reset_btn = gr.Button('Reset to default settings')
        gr.HTML("<span style='cursor:help' title='Reset all sliders to defaults.'></span>")
        reset_btn.click(reset_all, [], [h, x0, lr, init, st, xs, xb, w1b, w2b, tb])
    gr.HTML('</div>')

demo.launch()