{ "cells": [ { "cell_type": "markdown", "id": "d94621c9-dd17-4cab-a081-09b46e2bda97", "metadata": {}, "source": [ "# MAE experiments\n", "\n", "This notebook just experiments with the impact of rounding and accuracy on the MAE.\n", "\n", "To run it you may need to customise it - in my setup all the images are in the `unzip` directory and data files in the `data` directory. My local environmet looks like this\n", "- unzip\n", " - planet-dec17\n", " - ... lots of `png` images\n", " - planet-dec18\n", " - ... lots of `png` images\n", " - planet-jun18\n", " - ... lots of `png` images\n", " - planet-jun18\n", " - ... lots of `png` images\n", " -sentinel\n", " - ... lots of `tif` images\n", "- data\n", " - auxilary_data_unique.csv\n", " - extra_train.csv\n", " - test.csv\n", " - train-unique.csv\n", "\n", "## License\n", "\n", "MG Ferreira
\n", "© Ferra Solutions (Pty) Ltd https://www.ferrasolutions.com
\n", "Note that this is *copyrighted* material.\n", "\n", "You are welcome to use it to compete, but you can not publish this notebook elsewhere as if you created it, or use it in e.g. a lecture or as part of a paper or in some other setting without written permission from me.\n", "\n", "## Disclaimer\n", "\n", "This does not imply anything and is just a toy example to play with at your own risk." ] }, { "cell_type": "code", "execution_count": null, "id": "4f6d791f-57c1-4590-92e8-983771cb40e9", "metadata": {}, "outputs": [], "source": [ "import skimage.io\n", "import pandas as pd\n", "import numpy as np\n", "\n", "# Data paths - customise for your environment\n", "data_path = 'data'\n", "img_path = 'unzip'" ] }, { "cell_type": "code", "execution_count": null, "id": "a39ffd6a-3941-4e06-9f03-58cb2ec68638", "metadata": {}, "outputs": [], "source": [ "# Some utilities, mostly taken from the starter or derived directly from info in the starter\n", "CONST_X = 10.986328125 / 2\n", "CONST_Y = 10.985731758 / 2\n", "\n", "# Load the training and auxilliary data\n", "def load_data ():\n", "\n", " train = pd.read_csv ( f'{ data_path }/train-unique.csv' )\n", " aux = pd.read_csv ( f'{ data_path }/auxilary_data-unique.csv' )\n", " extra = pd.read_csv ( f'{ data_path }/extra_train.csv' )\n", "\n", " return pd.concat ( [ train, aux, extra ] )\n", "\n", "# Get image center in image pixels\n", "def img_center_pixels ( img ):\n", "\n", " return img.shape [ 1 ] // 2, img.shape [ 0 ] // 2\n", "\n", "# Given an image and target coordinates, return the target location in image pixels\n", "def img_target_pixels ( img, tgtx, tgty ):\n", "\n", " ctr = img_center_pixels ( img )\n", "\n", " x1 = ctr [ 0 ] - np.round ( tgtx / CONST_X * img.shape [ 1 ] )\n", " y1 = ctr [ 1 ] + np.round ( tgty / CONST_Y * img.shape [ 0 ] )\n", "\n", " return int ( x1 ), int ( y1 )\n", "\n", "# Given an image, the image center and some location in pixels, return the target coordinates\n", "def img_pixels_target ( img, pixx, pixy ):\n", "\n", " ctr = img_center_pixels ( img )\n", "\n", " x = CONST_X * float ( ctr [ 0 ] - pixx ) / float ( img.shape [ 1 ] )\n", " y = CONST_Y * float ( pixy - ctr [ 1 ] ) / float ( img.shape [ 0 ] )\n", "\n", " return x, y\n", "\n", "# Supplied - load RGB image\n", "def load_RGB_images ( ID ):\n", "\n", " # e.g id_0b242e06 -> 0b242e06\n", " name = ID.split('_')[1]\n", "\n", " img_jun17 = skimage.io.imread ( f'{img_path}/planet-jun17/{name}.png' )\n", " img_dec17 = skimage.io.imread ( f'{img_path}/planet-dec17/{name}.png' )\n", " img_jun18 = skimage.io.imread ( f'{img_path}/planet-jun18/{name}.png' )\n", " img_dec18 = skimage.io.imread ( f'{img_path}/planet-dec18/{name}.png' )\n", "\n", " return img_jun17, img_dec17, img_jun18, img_dec18\n", "\n", "# Load the RGB mages and return them as a stack\n", "def load_rgb_stack ( ID ):\n", "\n", " rgbs = load_RGB_images ( ID )\n", " return np.concatenate ( [ rgbs [ 0 ], rgbs [ 1 ], rgbs [ 2 ], rgbs [ 3 ] ], axis = 2 )" ] }, { "cell_type": "code", "execution_count": null, "id": "d95b9e12-2afd-4a43-9524-6232aae0f2ae", "metadata": {}, "outputs": [], "source": [ "# Iterate through the data and report the perfectly accurate plot center\n", "# accurate to a pixel and then accurate to half a pixel\n", "\n", "data = load_data ()\n", "mae = 0.0\n", "mae_half = 0.0\n", "n = 0\n", "\n", "for irow, row in data.iterrows ():\n", "\n", " img = load_rgb_stack ( row.ID )\n", " px, py = img_target_pixels ( img, row.x, row.y )\n", " x, y = img_pixels_target ( img, px, py )\n", " xu, yu = img_pixels_target ( img, px + 0.5, py + 0.5 )\n", " xl, yl = img_pixels_target ( img, px - 0.5, py - 0.5 )\n", " mae = mae + abs ( x - row.x ) + abs ( y - row.y )\n", " mae_half = mae_half + min ( abs ( xu - row.x ), abs ( x - row.x ), abs ( xl - row.x ) ) \\\n", " + min ( abs ( yu - row.y ), abs ( y - row.y ), abs ( yl - row.y ) )\n", " n = n + 2\n", "\n", "mae = mae / n\n", "mae_half = mae_half / n\n", "\n", "print ( f\"MAE accurate to pixel { mae }\" )\n", "print ( f\"MAE accurate to half pixel { mae_half }\" )" ] }, { "cell_type": "code", "execution_count": null, "id": "4b1344ea-7dec-426d-abab-57544f573fd1", "metadata": {}, "outputs": [], "source": [] } ], "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": "3.9.5" } }, "nbformat": 4, "nbformat_minor": 5 }