From 455349a0ebe6b275867e20111762ba4da21a9378 Mon Sep 17 00:00:00 2001 From: Alexander Mordvintsev Date: Wed, 6 Jun 2012 05:52:28 +0000 Subject: [PATCH] comments for digits.py sample --- samples/python2/common.py | 20 ++++++++++ samples/python2/digits.py | 93 +++++++++++++++++++++-------------------------- 2 files changed, 61 insertions(+), 52 deletions(-) diff --git a/samples/python2/common.py b/samples/python2/common.py index 54e23d6..0eb49b8 100644 --- a/samples/python2/common.py +++ b/samples/python2/common.py @@ -2,6 +2,7 @@ import numpy as np import cv2 import os from contextlib import contextmanager +import itertools as it image_extensions = ['.bmp', '.jpg', '.jpeg', '.png', '.tif', '.tiff', '.pbm', '.pgm', '.ppm'] @@ -170,3 +171,22 @@ class RectSelector: return x0, y0, x1, y1 = self.drag_rect cv2.rectangle(vis, (x0, y0), (x1, y1), (0, 255, 0), 2) + + +def grouper(n, iterable, fillvalue=None): + '''grouper(3, 'ABCDEFG', 'x') --> ABC DEF Gxx''' + args = [iter(iterable)] * n + return it.izip_longest(fillvalue=fillvalue, *args) + +def mosaic(w, imgs): + '''Make a grid from images. + + w -- number of grid columns + imgs -- images (must have same size and format) + ''' + imgs = iter(imgs) + img0 = imgs.next() + pad = np.zeros_like(img0) + imgs = it.chain([img0], imgs) + rows = grouper(w, imgs, pad) + return np.vstack(map(np.hstack, rows)) diff --git a/samples/python2/digits.py b/samples/python2/digits.py index c3494c0..f4bb0a5 100644 --- a/samples/python2/digits.py +++ b/samples/python2/digits.py @@ -1,89 +1,78 @@ -import numpy as np -import cv2 -import itertools as it - ''' -from scipy.io import loadmat +Neural network digit recognition sample. +Usage: + digits.py -m = loadmat('ex4data1.mat') -X = m['X'].reshape(-1, 20, 20) -X = np.transpose(X, (0, 2, 1)) -img = np.vstack(map(np.hstack, X.reshape(-1, 100, 20, 20))) -img = np.uint8(np.clip(img, 0, 1)*255) -cv2.imwrite('digits.png', img) + Sample loads a dataset of handwritten digits from 'digits.png'. + Then it trains a neural network classifier on it and evaluates + its classification accuracy. ''' +import numpy as np +import cv2 +from common import mosaic + def unroll_responses(responses, class_n): + '''[1, 0, 2, ...] -> [[0, 1, 0], [1, 0, 0], [0, 0, 1], ...]''' sample_n = len(responses) new_responses = np.zeros((sample_n, class_n), np.float32) new_responses[np.arange(sample_n), responses] = 1 return new_responses -SZ = 20 +SZ = 20 # size of each digit is SZ x SZ +CLASS_N = 10 digits_img = cv2.imread('digits.png', 0) +# prepare dataset h, w = digits_img.shape digits = [np.hsplit(row, w/SZ) for row in np.vsplit(digits_img, h/SZ)] digits = np.float32(digits).reshape(-1, SZ*SZ) N = len(digits) -labels = np.repeat(np.arange(10), N/10) +labels = np.repeat(np.arange(CLASS_N), N/CLASS_N) +# split it onto train and test subsets shuffle = np.random.permutation(N) train_n = int(0.9*N) - digits_train, digits_test = np.split(digits[shuffle], [train_n]) labels_train, labels_test = np.split(labels[shuffle], [train_n]) -labels_train_unrolled = unroll_responses(labels_train, 10) - +# train model model = cv2.ANN_MLP() -layer_sizes = np.int32([SZ*SZ, 25, 10]) +layer_sizes = np.int32([SZ*SZ, 25, CLASS_N]) model.create(layer_sizes) - -# CvANN_MLP_TrainParams::BACKPROP,0.001 -params = dict( term_crit = (cv2.TERM_CRITERIA_COUNT, 300, 0.01), +params = dict( term_crit = (cv2.TERM_CRITERIA_COUNT, 100, 0.01), train_method = cv2.ANN_MLP_TRAIN_PARAMS_BACKPROP, bp_dw_scale = 0.001, bp_moment_scale = 0.0 ) print 'training...' +labels_train_unrolled = unroll_responses(labels_train, CLASS_N) model.train(digits_train, labels_train_unrolled, None, params=params) model.save('dig_nn.dat') model.load('dig_nn.dat') -ret, resp = model.predict(digits_test) -resp = resp.argmax(-1) -error_mask = (resp == labels_test) -print error_mask.mean() - -def grouper(n, iterable, fillvalue=None): - "grouper(3, 'ABCDEFG', 'x') --> ABC DEF Gxx" - args = [iter(iterable)] * n - return it.izip_longest(fillvalue=fillvalue, *args) - -def mosaic(w, imgs): - imgs = iter(imgs) - img0 = imgs.next() - pad = np.zeros_like(img0) - imgs = it.chain([img0], imgs) - rows = grouper(w, imgs, pad) - return np.vstack(map(np.hstack, rows)) - -test_img = np.uint8(digits_test).reshape(-1, SZ, SZ) - -def vis_resp(img, flag): +def evaluate(model, samples, labels): + '''Evaluates classifier preformance on a given labeled samples set.''' + ret, resp = model.predict(samples) + resp = resp.argmax(-1) + error_mask = (resp == labels) + accuracy = error_mask.mean() + return accuracy, error_mask + +# evaluate model +train_accuracy, _ = evaluate(model, digits_train, labels_train) +print 'train accuracy: ', train_accuracy +test_accuracy, test_error_mask = evaluate(model, digits_test, labels_test) +print 'test accuracy: ', test_accuracy + +# visualize test results +vis = [] +for img, flag in zip(digits_test, test_error_mask): + img = np.uint8(img).reshape(SZ, SZ) img = cv2.cvtColor(img, cv2.COLOR_GRAY2BGR) if not flag: img[...,:2] = 0 - return img - -test_img = mosaic(25, it.starmap(vis_resp, it.izip(test_img, error_mask))) -cv2.imshow('test', test_img) + vis.append(img) +vis = mosaic(25, vis) +cv2.imshow('test', vis) cv2.waitKey() - - - - - - - -- 2.7.4