mlx-examples/gan/playground.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Import Library"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 23,
   "metadata": {},
   "outputs": [],
   "source": [
    "import mnist"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 24,
   "metadata": {},
   "outputs": [],
   "source": [
    "import mlx.core as mx\n",
    "import mlx.nn as nn\n",
    "import mlx.optimizers as optim\n",
    "\n",
    "from tqdm import tqdm\n",
    "import numpy as np\n",
    "import matplotlib.pyplot as plt"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# GAN Architecture"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Generator 👨🏻‍🎨"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 25,
   "metadata": {},
   "outputs": [],
   "source": [
    "def GenBlock(in_dim:int,out_dim:int):\n",
    "   \n",
    "    return nn.Sequential(\n",
    "        nn.Linear(in_dim,out_dim),\n",
    "        nn.BatchNorm(out_dim),\n",
    "        nn.ReLU()\n",
    "    )"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 26,
   "metadata": {},
   "outputs": [],
   "source": [
    "class Generator(nn.Module):\n",
    "\n",
    "    def __init__(self, z_dim:int = 10, im_dim:int = 784, hidden_dim: int =128):\n",
    "        super(Generator, self).__init__()\n",
    "        # Build the neural network\n",
    "        self.gen = nn.Sequential(\n",
    "            GenBlock(z_dim, hidden_dim),\n",
    "            GenBlock(hidden_dim, hidden_dim * 2),\n",
    "            GenBlock(hidden_dim * 2, hidden_dim * 4),\n",
    "            GenBlock(hidden_dim * 4, hidden_dim * 8),\n",
    "\n",
    "\n",
    "            nn.Linear(hidden_dim * 8,im_dim),\n",
    "            nn.Sigmoid()\n",
    "        )\n",
    "        \n",
    "    def __call__(self, noise):\n",
    "\n",
    "        return self.gen(noise)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 27,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "Generator(\n",
       "  (gen): Sequential(\n",
       "    (layers.0): Sequential(\n",
       "      (layers.0): Linear(input_dims=100, output_dims=128, bias=True)\n",
       "      (layers.1): BatchNorm(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n",
       "      (layers.2): ReLU()\n",
       "    )\n",
       "    (layers.1): Sequential(\n",
       "      (layers.0): Linear(input_dims=128, output_dims=256, bias=True)\n",
       "      (layers.1): BatchNorm(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n",
       "      (layers.2): ReLU()\n",
       "    )\n",
       "    (layers.2): Sequential(\n",
       "      (layers.0): Linear(input_dims=256, output_dims=512, bias=True)\n",
       "      (layers.1): BatchNorm(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n",
       "      (layers.2): ReLU()\n",
       "    )\n",
       "    (layers.3): Sequential(\n",
       "      (layers.0): Linear(input_dims=512, output_dims=1024, bias=True)\n",
       "      (layers.1): BatchNorm(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n",
       "      (layers.2): ReLU()\n",
       "    )\n",
       "    (layers.4): Linear(input_dims=1024, output_dims=784, bias=True)\n",
       "    (layers.5): Sigmoid()\n",
       "  )\n",
       ")"
      ]
     },
     "execution_count": 27,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "gen = Generator(100)\n",
    "gen"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 28,
   "metadata": {},
   "outputs": [],
   "source": [
    "def get_noise(n_samples, z_dim):\n",
    "    return np.random.randn(n_samples,z_dim)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Discriminator 🕵🏻‍♂️"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 29,
   "metadata": {},
   "outputs": [],
   "source": [
    "def DisBlock(in_dim:int,out_dim:int):\n",
    "    return nn.Sequential(\n",
    "        nn.Linear(in_dim,out_dim),\n",
    "        nn.LeakyReLU(negative_slope=0.2)\n",
    "    )"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 30,
   "metadata": {},
   "outputs": [],
   "source": [
    "class Discriminator(nn.Module):\n",
    "\n",
    "    def __init__(self,im_dim:int = 784, hidden_dim:int = 128):\n",
    "        super(Discriminator, self).__init__()\n",
    "\n",
    "        self.disc = nn.Sequential(\n",
    "            DisBlock(im_dim, hidden_dim * 4),\n",
    "            DisBlock(hidden_dim * 4, hidden_dim * 2),\n",
    "            DisBlock(hidden_dim * 2, hidden_dim),\n",
    "\n",
    "            nn.Linear(hidden_dim,1),\n",
    "        )\n",
    "        \n",
    "    def __call__(self, noise):\n",
    "\n",
    "        return self.disc(noise)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 31,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "Discriminator(\n",
       "  (disc): Sequential(\n",
       "    (layers.0): Sequential(\n",
       "      (layers.0): Linear(input_dims=784, output_dims=512, bias=True)\n",
       "      (layers.1): LeakyReLU()\n",
       "    )\n",
       "    (layers.1): Sequential(\n",
       "      (layers.0): Linear(input_dims=512, output_dims=256, bias=True)\n",
       "      (layers.1): LeakyReLU()\n",
       "    )\n",
       "    (layers.2): Sequential(\n",
       "      (layers.0): Linear(input_dims=256, output_dims=128, bias=True)\n",
       "      (layers.1): LeakyReLU()\n",
       "    )\n",
       "    (layers.3): Linear(input_dims=128, output_dims=1, bias=True)\n",
       "  )\n",
       ")"
      ]
     },
     "execution_count": 31,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "disc = Discriminator()\n",
    "disc"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Model Training 🏋🏻‍♂️"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 32,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Set your parameters\n",
    "criterion = nn.losses.binary_cross_entropy\n",
    "n_epochs = 200\n",
    "z_dim = 64\n",
    "display_step = 500\n",
    "batch_size = 128\n",
    "lr = 0.00001"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 33,
   "metadata": {},
   "outputs": [],
   "source": [
    "gen = Generator(z_dim)\n",
    "mx.eval(gen.parameters())\n",
    "gen_opt = optim.Adam(learning_rate=lr)\n",
    "\n",
    "disc = Discriminator()\n",
    "mx.eval(disc.parameters())\n",
    "disc_opt = optim.Adam(learning_rate=lr)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Losses"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 34,
   "metadata": {},
   "outputs": [],
   "source": [
    "def disc_loss(gen, disc, real, num_images, z_dim):\n",
    "    noise =  mx.array(get_noise(num_images, z_dim))\n",
    "    fake_images = gen(noise)\n",
    "        \n",
    "    fake_disc = disc(fake_images)\n",
    "    \n",
    "    fake_labels = mx.zeros((fake_images.shape[0],1))\n",
    "    fake_loss = nn.losses.binary_cross_entropy(fake_disc,fake_labels,with_logits=True)\n",
    "    \n",
    "    real_disc = disc(real)\n",
    "    real_labels = mx.ones((real.shape[0],1))\n",
    "\n",
    "    real_loss = nn.losses.binary_cross_entropy(real_disc,real_labels,with_logits=True)\n",
    "\n",
    "    disc_loss = (fake_loss + real_loss) / 2\n",
    "\n",
    "    return disc_loss"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 35,
   "metadata": {},
   "outputs": [],
   "source": [
    "def gen_loss(gen, disc, num_images, z_dim):\n",
    "\n",
    "    noise = mx.array(get_noise(num_images, z_dim))\n",
    "    fake_images = gen(noise)\n",
    "    fake_disc = disc(fake_images)\n",
    "\n",
    "    fake_labels = mx.ones((fake_images.shape[0],1))\n",
    "    \n",
    "    gen_loss = nn.losses.binary_cross_entropy(fake_disc,fake_labels,with_logits=True)\n",
    "\n",
    "    return gen_loss"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 36,
   "metadata": {},
   "outputs": [],
   "source": [
    "train_images, _, test_images, _ = map(\n",
    "    mx.array, getattr(mnist, 'mnist')()\n",
    ")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 37,
   "metadata": {},
   "outputs": [],
   "source": [
    "def batch_iterate(batch_size:int, ipt:list):\n",
    "    perm = mx.array(np.random.permutation(len(ipt)))\n",
    "    for s in range(0, ipt.size, batch_size):\n",
    "        ids = perm[s : s + batch_size]\n",
    "        yield ipt[ids]"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### show batch of images"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 38,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAUkAAAFICAYAAADd1gwNAAAAOXRFWHRTb2Z0d2FyZQBNYXRwbG90bGliIHZlcnNpb24zLjcuMSwgaHR0cHM6Ly9tYXRwbG90bGliLm9yZy/bCgiHAAAACXBIWXMAAA9hAAAPYQGoP6dpAACQiklEQVR4nOy9V49l2ZXf+bvee3/D+4zISFuZVZXFYpHFIkU2DdjdYKuBgQYcYCToC+gLCHrUsx4ECIIwwmB6ultSCxTZ7C4y2SzDrGKl95HhzfXe+3vnIWfvupEm0oc9PyDAYmbEzXN2nLP22sv8l6rX6/VQUFBQUHgi6r2+AAUFBYX9jGIkFRQUFHZAMZIKCgoKO6AYSQUFBYUdUIykgoKCwg4oRlJBQUFhBxQjqaCgoLADipFUUFBQ2AHFSCooKCjsgPZ5v1GlUr3J6zhQvGyTkrKGX6Os4aujrOGr8zxrqHiSCgoKCjugGEkFBQWFHVCMpIKCgsIOPHdMUkFBQQHAZDJht9sxmUyMjY3hdDrpdDq0Wi1yuRy3bt2iUqns9WW+NhQjqaCg8EI4nU6OHTvG4OAgP//5zzl+/DiVSoVKpcLVq1f5t//23ypGUkFB4eghsuJms5mBgQEGBgYIhUIEg0HK5TLlchmn04lGo9njK329KEZSQUHhuVCr1ajVaqanp/n5z3+O3+9nYGCAXq+37euwceCM5PPUeB3GX5SCwl6jUqlQq9W43W5OnTqFx+MBONQGEvaxkRS/EK1Wi91uR6/X4/F48Hg8aLVajEYjWq0Wm82GyWSiWq1SKpWoVCosLi5SLBap1WrUarW9vhUFBYUDzL42ksIYBgIB7HY7MzMzzM7OYjQacblcmEwmhoaGcLvdpNNpNjc3icfj/OIXv2BjY4N0Oq0YSQUFhVdiXxhJlUqFRqNBo9Fgt9sxm82YTCbcbjcGg4FwOIzFYmFsbIzBwUEMBgMOhwO9Xo/D4cBsNmO32/H5fKhUKsbHxzEYDAAUCgW63S7dbneP71LhoKPVauVx0+12y9BPp9MhlUpRKpXodDp0Op09vtLdR61Wo9Fo0Ol0WCwWLBYLjUaDdru915f2yuwLI6nRaOSx+Z133mFqaorR0VHOnj2L0WjEZDKh1WrR6/UYjUb5CxH0ej1MJhOBQIBarUYgECCXy/G3f/u3JJNJms0m1Wr10MZMFN48YgM3Go1873vf46OPPpJGslqt8nd/93dcu3ZNhn2O2rOm1WoxmUw4nU5GR0fpdrtEo1Gy2exeX9ors6dGUsQd9Xo9NpsNq9VKMBhkZGSEiYkJZmdnMRgMqNVqVCoV3W5XBoiFd9hoNGi1Wuh0OgwGAxqNhnA4jN1ux+PxYDQaAajVakfuwX0W4iXXaDSoVCq5roDieT+CSqXCYDBgNpsJBAJMTU2hVj9sWCuXy/j9fqxWK91ul3K5fGifNfGMtNttOp2OdFZE5lun02E2m7FarWi1+8IHe2X29C7cbjcej4dwOMy3vvUtvF4vU1NThMNhHA4HRqORTqfD1tYWlUqFXC5HLpejVquRTqdpNBpkMhkqlQqjo6OcOnUKq9XKwMAANpuN6elpzp49SyKR4Pbt29Tr9b283X2FVqvFbDZjMBhkXDefz5NKpWg0GmSzWZrN5l5f5r7BYDAwMTGBz+djZmaGkZERaSQbjQbf//73mZiY4MqVK/z617+m0Wjs8RW/fsTGWSwWWVpaolgsylCYRqORm4jX66VSqRCLxfb4il8Pe2okrVYr4XCYmZkZvv/97xMOh3G73VitVvk9zWaTVColEzObm5sUi0VWVlaoVCpEo1FyuRznz59Hr9cTDAYZHR3F7XYzODjI5OQkWq2WhYUFxUj2odFosFgsmM1mxsfHGRkZIRqNAg89o1KppBjJPnQ6HaFQiJGREYaGhggEAtJIttttzp07x+joKLVajd/85jeH0kiKU5x477rdLm63G4vFIvMKIl/gdDrR6XR7fcmvhV03kiqVCp1Oh1arZWhoiDNnzjA8PIzL5cJsNtNsNsnn8xSLRRKJBMVikevXr5NOp0mn02QyGarVKslkknq9TqlUotFoUC6XyefzmM1mWq0WgPyl6XS6A6uhp1KpsNlsmM1mnE4ng4OD2+KxkUiEWCxGs9mkUqk885is0+nQ6XT4fD7Onj2L0+lkdnaWYDCIz+fDYrHIdT5MrWUvi0ajQa/Xy1a86elpgsHgtudJrVZjs9mAhxv/QX3Wnodut0u9XieVSqHT6Q7lZvAou24k1Wo1FosFk8nEqVOn+LM/+zNcLheDg4Po9XoSiQT5fJ4HDx7w2WefkU6n+eqrr0ilUrTbbVqtFr1eT2YQO50OvV6PfD5PNBpFo9FII2kwGLBarRiNxgP74KrVakKhEOFwmPn5ef7kT/4Es9kMPNzZf/Ob3/DrX/+aQqHA5ubmMx9aIU4wOzvLz3/+cwYGBggGgzgcDqLRKIuLi6yurnLv3j1SqdRu3OK+RlRQhMNhPvzwQ86fP4/BYHjMSPr9frxeL16vV3qYhw0RZxUnuWazycmTJ/f4qt48e+JJisCu0+nE5XJhs9lkkLdUKpFKpUgkEsRiMbLZLLlcTpbyPC0g3mq1qNVqNBqNbd6USqU6sAYSHl6/y+ViaGiIcDhMIBDAbDbLjcLpdGK1Wmk0Gk99OUXNqShfCYfDDA4O4vP5cLvd6PV6er0ezWaTcrlMtVo9kmUsT0LEbs1mMxaLZVso6FFEIlKUsx3WNRSnPYvF8lhIRq1WYzQaMZvNSuLmZTEYDBw7doxwOCyVRDQaDe12m3q9zldffcXVq1dZX1/n2rVr1Gq1ZxpIeJi9TiQST/zFHWT0ej3f/va3+bM/+zPsdruMhXU6HdrtNoODg4yPj2MymdjY2Hhi8bxWqyUYDGKxWPiTP/kTfvCDH+ByuRgbG0Oj0bC8vEwymeTq1av85je/IZfLEY/H9+Bu9x9Wq1XGIU0m0xO/R8Tp6vU67XYbh8MBPNzwD0Od4KMkk0k+//xzRkZG+P73v7/t76xWK3NzczidTu7cubNHV/h62XUjqdVqcbvdhEIhvF4vNpuNXq9HtVql2WySTCZZWVlha2uLWCwmj87PQhjZer1+qHZwtVrN0NAQp0+flqVQwutrt9tYLBbsdjvFYvGpnqQIcTidTqanp3nvvffQarVotVrq9Tr5fJ7NzU2Wl5e5ceMG1Wr1udf9sCOO2zabDY1GQ6/Xe+xk0uv1aLVaNBoNer0eRqORer1+aGO61WqVSCSCXq9/zCHR6XR4PB46nY4svzvo7Mlx22g0YrVa0ev1ALJW0mw2Mz09TavV4tq1aywsLDz3y2qz2RgeHmZgYEDu+J1OR9ZRHuS6NXGM6y9evnz5MvF4nMuXL3P58mXy+fxTs/cGg4HJyUkGBwcJhUJotVqazSaxWIxCocAf/vAHrl27xsbGhtxkDvJ6vU5arRalUolcLsfKygpqtVrGHgUqlQqTyYRGo8HtdjMwMIDBYKBcLh+qU83zIDQWer2ebBARuYSDyp4kbkwm0zYjKQp1NRoNs7OzuN1uyuXyC8U0bDYbo6Oj8gGFh96lMJIHGWEkBdVqlc8//5ybN29y//597t+/v6NhMxqNTE1NMTs7y8DAAFqtllKpxPr6OolEgn/6p3/ik08+2VZMrvCQVqtFoVBAr9ezvLxMq9Xi2LFjeDweuWkJI2kymfB6vXKNt7a2KJfLe3wHu4tOp8Pr9aLRaHA6nZhMJtmeeFA33l03kiLh0Gq1nvhCNptNGdt5nkU1GAxotVqcTid+vx+PxyPrs2q1Gvl8nnK5fKhe/m63S61Wo1wuP9PzE10QLpcLr9crM+P1ep1oNEo0GpU9xwqPI5JZWq2WpaUlSqWSNIaiU6x/A2u321Kl+yiuqTgVGo1G7HY7brebUql0oDvedt1IiratJx0PRSnP1tYWuVzumYZNZGvtdjtTU1OcO3dOZst7vR6pVIp79+6RzWYPvDfZT6fTkcmVneJewkBaLBampqY4efIkTqcTlUpFJpPhd7/7Hevr67KIXOFxyuUya2traDQaVlZWMBgMspDa5/M
      "text/plain": [
       "<Figure size 400x400 with 16 Axes>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "for X in batch_iterate(16, train_images):\n",
    "    fig,axes = plt.subplots(4, 4, figsize=(4, 4))\n",
    "\n",
    "    for i, ax in enumerate(axes.flat):\n",
    "        img = mx.array(X[i]).reshape(28,28)\n",
    "        ax.imshow(img,cmap='gray')\n",
    "        ax.axis('off')\n",
    "    break"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 39,
   "metadata": {},
   "outputs": [],
   "source": [
    "def show_images(imgs:list[int],num_imgs:int = 25):\n",
    "    fig,axes = plt.subplots(5, 5, figsize=(4, 4))\n",
    "    \n",
    "    for i, ax in enumerate(axes.flat):\n",
    "        img = mx.array(imgs[i]).reshape(28,28)\n",
    "        ax.imshow(img,cmap='gray')\n",
    "        ax.axis('off')\n",
    "    plt.show()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 43,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "array(0.675341, dtype=float32)"
      ]
     },
     "execution_count": 43,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "z_dim = 64\n",
    "gen = Generator(z_dim)\n",
    "mx.eval(gen.parameters())\n",
    "gen_opt = optim.Adam(learning_rate=lr)\n",
    "\n",
    "disc = Discriminator()\n",
    "mx.eval(disc.parameters())\n",
    "disc_opt = optim.Adam(learning_rate=lr)\n",
    "\n",
    "g_loss = gen_loss(gen, disc, 8, z_dim)\n",
    "g_loss\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 44,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "60000"
      ]
     },
     "execution_count": 44,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "len(train_images)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 45,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "  0%|          | 0/200 [00:00<?, ?it/s]"
     ]
    }
   ],
   "source": [
    "# Train the GAN for only 1000 images\n",
    "batch_size = 16\n",
    "display_step = 20\n",
    "cur_step = 0\n",
    "mean_generator_loss = 0\n",
    "mean_discriminator_loss = 0\n",
    "\n",
    "D_loss_grad = nn.value_and_grad(disc, disc_loss)\n",
    "G_loss_grad = nn.value_and_grad(gen, gen_loss)\n",
    "\n",
    "\n",
    "for epoch in tqdm(range(200)):\n",
    "\n",
    "    for real in batch_iterate(batch_size, train_images[:500]):\n",
    "                \n",
    "        D_loss,D_grads = D_loss_grad(gen, disc, real, batch_size, z_dim)\n",
    "\n",
    "        # Update optimizer\n",
    "        disc_opt.update(disc, D_grads)\n",
    "        \n",
    "        # Update gradients\n",
    "        mx.eval(disc.parameters(), disc_opt.state)\n",
    "\n",
    "        G_loss,G_grads = G_loss_grad(gen, disc, batch_size, z_dim)\n",
    "        \n",
    "        # Update optimizer\n",
    "        gen_opt.update(gen, G_grads)\n",
    "        \n",
    "        # Update gradients\n",
    "        mx.eval(gen.parameters(), gen_opt.state)\n",
    "        \n",
    "    if cur_step % display_step == 0 and cur_step > 0:\n",
    "        print(f\"Step {epoch}: Generator loss: {G_loss}, discriminator loss: {D_loss}\")\n",
    "        fake_noise = mx.array(get_noise(batch_size, z_dim))\n",
    "        fake = gen(fake_noise)\n",
    "        show_images(fake)\n",
    "        show_images(real)\n",
    "    cur_step += 1"
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "base",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.10.10"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}