experiment.ipynb

gistfile1.txt

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [],
   "source": [
    "from pathlib import Path\n",
    "import numpy as np\n",
    "import os\n",
    "import re\n",
    "import math\n",
    "import tensorflow as tf\n",
    "import copy\n",
    "from typing import Tuple, Optional, Callable, List"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {},
   "outputs": [],
   "source": [
    "from toai.image import ImageAugmentor, ImageDataset, ImageParser, ImageResizer"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 36,
   "metadata": {},
   "outputs": [],
   "source": [
    "class ImageDatasetConfig:\n",
    "\n",
    "    def __init__(\n",
    "            self,\n",
    "            img_dims: Tuple[int, int, int],\n",
    "            parallel_calls: int = 1,\n",
    "            prefetch: int = 1,\n",
    "            preprocess_pipeline: List[Callable] = [],\n",
    "            batch_size: int = 8,\n",
    "            shuffle: bool = False,\n",
    "            regexp: str = None\n",
    "            ):\n",
    "        self.parallel_calls = parallel_calls\n",
    "        self.img_dims = img_dims\n",
    "        self.prefetch = prefetch\n",
    "        self.preprocess_pipeline = preprocess_pipeline\n",
    "        self.batch_size = batch_size\n",
    "        self.shuffle = shuffle\n",
    "        self.regexp = regexp\n",
    "\n",
    "    def copy(self, preprocess_pipeline: List[Callable], shuffle: bool = False) -> 'ImageDatasetConfig':\n",
    "        new = copy.deepcopy(self)\n",
    "        new.shuffle = shuffle\n",
    "        new.preprocess_pipeline = preprocess_pipeline\n",
    "        return new\n",
    "\n",
    "\n",
    "class ImageDataset:\n",
    "    data: tf.data.Dataset\n",
    "    x: np.ndarray\n",
    "    y: np.ndarray\n",
    "    length: int\n",
    "    steps: int\n",
    "    classes: np.ndarray\n",
    "    n_classes: int\n",
    "\n",
    "    def __init__(self, config: ImageDatasetConfig):\n",
    "        self.config = config\n",
    "\n",
    "    def build_from_df(self) -> 'ImageDataset':\n",
    "        pass\n",
    "\n",
    "    def build_from_path(self, path: Path, default_label: Optional[str] = None) -> 'ImageDataset':\n",
    "        paths = []\n",
    "        labels = []\n",
    "\n",
    "        if self.config.regexp:\n",
    "            for value in os.listdir(str(path)):\n",
    "                match = re.match(self.config.regexp, value)\n",
    "                if match:\n",
    "                    labels.append(match.group(1))\n",
    "                elif default_label:\n",
    "                    labels.append(default_label)\n",
    "                else:\n",
    "                    raise ValueError(f\"No match found and no default value provided for value: {value}\")\n",
    "\n",
    "                paths.append(f\"{path}/{value}\")\n",
    "        else:\n",
    "            raise ValueError(\"Unexpected configuration\")\n",
    "\n",
    "        return self.__build(np.asarray(paths), np.asarray(labels))\n",
    "\n",
    "    def __build(self, x: np.ndarray, y: np.ndarray) -> 'ImageDataset':\n",
    "        self.x = x\n",
    "        self.y = y\n",
    "        self.length = len(y)\n",
    "        self.classes = np.unique(y)\n",
    "        self.n_classes = len(self.classes)\n",
    "        self.steps = math.ceil(self.length / self.config.batch_size)\n",
    "\n",
    "        image_ds = tf.data.Dataset.from_tensor_slices(x)\n",
    "\n",
    "        for fun in self.config.preprocess_pipeline:\n",
    "            image_ds = image_ds.map(fun, num_parallel_calls=self.config.parallel_calls)\n",
    "\n",
    "        label_ds = tf.data.Dataset.from_tensor_slices(y.astype(float))\n",
    "        dataset = tf.data.Dataset.zip((image_ds, label_ds))\n",
    "\n",
    "        if self.config.shuffle:\n",
    "            dataset = dataset.shuffle(self.config.batch_size)\n",
    "\n",
    "        self.data = dataset.batch(self.config.batch_size).repeat().prefetch(self.config.prefetch)\n",
    "\n",
    "        return self"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 37,
   "metadata": {},
   "outputs": [],
   "source": [
    "DATA_DIR = Path(\"food\")\n",
    "IMG_DIMS = (224, 224, 3)\n",
    "\n",
    "base_config = ImageDatasetConfig(\n",
    "    img_dims=IMG_DIMS,\n",
    "    parallel_calls=1,\n",
    "    batch_size=8,\n",
    "    prefetch=1,\n",
    "    regexp=r\"(\\d+)\"\n",
    ")\n",
    "\n",
    "train = ImageDataset(\n",
    "    base_config\n",
    "        .copy(\n",
    "            preprocess_pipeline = [\n",
    "                ImageParser(),\n",
    "                ImageResizer(img_dims=IMG_DIMS, resize=\"random_crop\", crop_adjustment=1.6),\n",
    "                ImageAugmentor(level=3, flips=\"both\"),\n",
    "            ],\n",
    "            shuffle = True\n",
    "        )\n",
    ").build_from_path(DATA_DIR/'training')\n",
    "\n",
    "validation = ImageDataset(\n",
    "     base_config\n",
    "        .copy(\n",
    "            preprocess_pipeline = [\n",
    "                ImageParser(),\n",
    "                ImageResizer(img_dims=IMG_DIMS, resize=\"crop\", crop_adjustment=1.0),\n",
    "            ]\n",
    "        )\n",
    ").build_from_path(DATA_DIR/'validation')"
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3",
   "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": "3.7.3"
  },
  "toc": {
   "base_numbering": 1,
   "nav_menu": {},
   "number_sections": true,
   "sideBar": true,
   "skip_h1_title": false,
   "title_cell": "Table of Contents",
   "title_sidebar": "Contents",
   "toc_cell": false,
   "toc_position": {},
   "toc_section_display": true,
   "toc_window_display": false
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}