Merge remote-tracking branch 'origin/master'

deep learning/CPPN/CPPN.ipynb

@@ -0,0 +1,368 @@
+{
+ "cells": [
+  {
+   "cell_type": "markdown",
+   "metadata": {
+    "toc": true
+   },
+   "source": [
+    "<h1>Table of Contents<span class=\"tocSkip\"></span></h1>\n",
+    "<div class=\"toc\" style=\"margin-top: 1em;\"><ul class=\"toc-item\"><li><span><a href=\"#Instantiate-CPPN\" data-toc-modified-id=\"Instantiate-CPPN-1\"><span class=\"toc-item-num\">1&nbsp;&nbsp;</span>Instantiate CPPN</a></span></li><li><span><a href=\"#Generate-Image\" data-toc-modified-id=\"Generate-Image-2\"><span class=\"toc-item-num\">2&nbsp;&nbsp;</span>Generate Image</a></span></li><li><span><a href=\"#Animation\" data-toc-modified-id=\"Animation-3\"><span class=\"toc-item-num\">3&nbsp;&nbsp;</span>Animation</a></span></li><li><span><a href=\"#Parameters-Grid-Search\" data-toc-modified-id=\"Parameters-Grid-Search-4\"><span class=\"toc-item-num\">4&nbsp;&nbsp;</span>Parameters Grid Search</a></span></li><li><span><a href=\"#Considerations\" data-toc-modified-id=\"Considerations-5\"><span class=\"toc-item-num\">5&nbsp;&nbsp;</span>Considerations</a></span></li></ul></div>"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 1,
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stderr",
+     "output_type": "stream",
+     "text": [
+      "Using TensorFlow backend.\n"
+     ]
+    }
+   ],
+   "source": [
+    "import numpy as np\n",
+    "import yaml\n",
+    "import os\n",
+    "import cv2\n",
+    "import pandas as pd\n",
+    "import seaborn as sns\n",
+    "import matplotlib.pyplot as plt\n",
+    "from matplotlib import animation\n",
+    "from datetime import datetime\n",
+    "\n",
+    "from pathlib import Path\n",
+    "\n",
+    "plt.rcParams['animation.ffmpeg_path'] = str(Path.home() / \"anaconda3/envs/image-processing/bin/ffmpeg\")\n",
+    "\n",
+    "%matplotlib notebook\n",
+    "\n",
+    "%load_ext autoreload\n",
+    "%autoreload 2\n",
+    "\n",
+    "from CPPN import CPPN"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "# Instantiate CPPN"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 2,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "res_path = Path.home() / 'Documents' / 'datasets' / 'cppn'"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 3,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "extra_funs= {\n",
+    "    'base': lambda x,y: x*y,\n",
+    "    'cos_sin': lambda x,y: np.cos(x)*np.sin(y),\n",
+    "    'cube': lambda x,y: x**3 + 3*y - y**3 -3*x,\n",
+    "    'rand': lambda x,y: np.sqrt(x*x+y*y) + (x*x) + np.tan(y),\n",
+    "}"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 4,
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/plain": [
+       "{'hidden_dim': 16,\n",
+       " 'inner_activation': 'tanh',\n",
+       " 'inner_architecture_key': 'base',\n",
+       " 'kernel_init_mean': 0.0,\n",
+       " 'kernel_init_stddev': 1.0,\n",
+       " 'nb_channels': 3,\n",
+       " 'nb_hidden_layers': 5,\n",
+       " 'scale_factor': 1.0,\n",
+       " 'z_dim': 16}"
+      ]
+     },
+     "execution_count": 4,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "# load config file\n",
+    "with open('cppn_config.yaml', 'r') as f:\n",
+    "    model_config = yaml.load(f)\n",
+    "model_config = model_config['rgb_test']\n",
+    "model_config"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# customize model config\n",
+    "batch_size = 1\n",
+    "\n",
+    "img_width = 256\n",
+    "img_height = 256\n",
+    "img_size = (img_width, img_height)\n",
+    "\n",
+    "model_config['nb_hidden_layers'] = 8\n",
+    "#model_config['kernel_init_stddev'] = 1.\n",
+    "model_config['kernel_init_mean'] = 0.\n",
+    "model_config['nb_channels'] = 1\n",
+    "\n",
+    "#model_config['inner_architecture_key'] = 'residual'\n",
+    "\n",
+    "# init model\n",
+    "cppn = CPPN(batch_size=batch_size, img_width=img_width, img_height=img_height, \n",
+    "            **model_config)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "cppn.model.summary()"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "# Generate Image"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "x, y, r, e = cppn.get_data(extra_fun=extra_funs['cos_sin'])"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "z = cppn.get_z()"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "plt.imshow(cppn.generate_imgs(x, y, r, e, z)[0], cmap='gray')"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "# Animation"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "save_anim = False\n",
+    "animate_data = True\n",
+    "FRAMES = 300\n",
+    "batch_size = 1\n",
+    "img_width = 1080\n",
+    "img_height = 720\n",
+    "img_size = (img_width, img_height)\n",
+    "\n",
+    "\n",
+    "dpi = 100\n",
+    "if save_anim:\n",
+    "    fig, ax = plt.subplots(dpi=dpi, figsize=(img_width/dpi, img_height/dpi))\n",
+    "else:\n",
+    "    fig, ax = plt.subplots(dpi=100, figsize=(5, 5))\n",
+    "plt.axis('off')\n",
+    "\n",
+    "# Init model and data\n",
+    "with open('cppn_config.yaml', 'r') as f:\n",
+    "    model_config = yaml.load(f)\n",
+    "data_config = model_config['test_config']\n",
+    "model_config = model_config['rgb_test']\n",
+    "cppn = CPPN(batch_size=batch_size, img_width=img_width, img_height=img_height, \n",
+    "            **model_config)\n",
+    "x, y, r, e = cppn.get_data(scale=data_config['scale'], \n",
+    "                           translation=data_config['translation'],\n",
+    "                           rotation=data_config['rotation'], \n",
+    "                           extra_fun=extra_funs[data_config['extra_fun']])\n",
+    "z = cppn.get_z()\n",
+    "\n",
+    "z_max_val = 1\n",
+    "#z = np.zeros((1, model_config['z_dim']))-z_max_val\n",
+    "add_val = (z_max_val*2)/FRAMES\n",
+    "add_val = 1.0\n",
+    "def animate(i, ax, cppn, x, y, r, e, z, data_config, extra_funs, animate_data):\n",
+    "    z_idx = int(i/(int(FRAMES/model_config['z_dim'])))\n",
+    "    z[0][z_idx] += add_val\n",
+    "    #z[0] += add_val\n",
+    "    \n",
+    "    if animate_data:\n",
+    "        x, y, r, e = cppn.get_data(scale=data_config['scale']+i*data_config['scale_speed'], \n",
+    "                                   translation=data_config['translation']+i*data_config['translation_speed'], \n",
+    "                                   rotation=data_config['rotation']+i*data_config['rotation_speed'], \n",
+    "                                   extra_fun=extra_funs[data_config['extra_fun']])\n",
+    "    \n",
+    "    ax.imshow(cppn.generate_imgs(x, y, r, e, z)[0], cmap='gray')\n",
+    "\n",
+    "ani = animation.FuncAnimation(fig, animate, frames=FRAMES, interval=10,\n",
+    "                              fargs=[ax, cppn, x, y, r, e, z, data_config, extra_funs, animate_data])\n",
+    "if save_anim:\n",
+    "    ani.save(str(res_path / 'tests' / 'anim_{}.mp4'.format(datetime.strftime(datetime.now(), \"%Y-%m-%d_%H-%M\"))), \n",
+    "             animation.FFMpegFileWriter(fps=30))"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "# Parameters Grid Search"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 5,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "from sklearn.model_selection import ParameterGrid"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "param_grid = {\n",
+    "                'inner_architecture_key': ['base', 'residual', 'softplus'],\n",
+    "                'kernel_init_stddev': np.linspace(0.7, 4., num=4),\n",
+    "                'scale': np.linspace(-2., 2., num=5),\n",
+    "                'translation': np.linspace(-4., 4., num=3),\n",
+    "                #'rotation': np.linspace(1, 360, num=4),\n",
+    "                'nb_hidden_layers': np.arange(3, 7, 2),\n",
+    "                'z_dim': [8, 16, 32, 64],\n",
+    "                'hidden_dim': [16, 32, 64],\n",
+    "                'extra_fun': ['base', 'cos_sin', 'cube', 'rand']\n",
+    "             }\n",
+    "grid = ParameterGrid(param_grid)\n",
+    "\n",
+    "# Init model and data\n",
+    "with open('cppn_config.yaml', 'r') as f:\n",
+    "    model_config = yaml.load(f)\n",
+    "data_config = model_config['test_config']\n",
+    "model_config = model_config['rgb_test']\n",
+    "\n",
+    "batch_size = 5\n",
+    "img_width = 1920\n",
+    "img_height = 1080\n",
+    "\n",
+    "for i, params in enumerate(grid):\n",
+    "    print(\"Params {}: {}\".format(i, params))\n",
+    "    out_dir = res_path / 'hq' / f'params_{i}'\n",
+    "    out_dir.mkdir(parents=True, exist_ok=False) \n",
+    "    current_config = model_config.copy()\n",
+    "    current_data_config = data_config.copy()\n",
+    "    current_config.update(params)\n",
+    "    current_data_config.update(params)\n",
+    "    cppn = CPPN(batch_size=batch_size, img_width=img_width, img_height=img_height, \n",
+    "                **current_config)\n",
+    "    x, y, r, e = cppn.get_data(scale=data_config['scale'], \n",
+    "                               translation=data_config['translation'], \n",
+    "                               rotation=data_config['rotation'],\n",
+    "                               extra_fun=extra_funs['cos_sin'])\n",
+    "    z = cppn.get_z()\n",
+    "    imgs = cppn.generate_imgs(x, y, r, e, z)\n",
+    "    for j, img in enumerate(imgs): \n",
+    "        plt.imsave(str(out_dir / f'sample_{j}.png'), img, cmap='gray')"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "# Considerations"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "ReLu more interesting gradients but clear cut producing uninteresting entries for higher stddev."
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "Should use a bias_initializer? Should it be the same as kernel?\n",
+    "\n",
+    "In general low kernel_init_stddev provides blurry results, still around 0.5. One can notice improvements with increased values for hidden_dim and even nb_hidden_layers, but still too much blur to be interesting. With 5.0 stddev we are instead already on the other opposite, with clear cute areas, cell-shading like.\n",
+    "\n",
+    "Didn't notice major differenced when increasing the z_dim, sometimes even had impression of more interesting results with less"
+   ]
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": "Python [conda env:image-processing]",
+   "language": "python",
+   "name": "conda-env-image-processing-py"
+  },
+  "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 3
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython3",
+   "version": "3.6.7"
+  },
+  "notify_time": "30",
+  "toc": {
+   "nav_menu": {},
+   "number_sections": true,
+   "sideBar": true,
+   "skip_h1_title": false,
+   "toc_cell": true,
+   "toc_position": {},
+   "toc_section_display": "block",
+   "toc_window_display": true
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 2
+}