comments for digits.py sample

diff --git a/samples/python2/common.py b/samples/python2/common.py
index 54e23d6..0eb49b8 100644 (file)
--- a/samples/python2/common.py
+++ b/samples/python2/common.py
@@ -2,6 +2,7 @@ import numpy as np
 import cv2\r
 import os\r
 from contextlib import contextmanager\r
+import itertools as it\r
 \r
 image_extensions = ['.bmp', '.jpg', '.jpeg', '.png', '.tif', '.tiff', '.pbm', '.pgm', '.ppm']\r
 \r
@@ -170,3 +171,22 @@ class RectSelector:
             return\r
         x0, y0, x1, y1 = self.drag_rect\r
         cv2.rectangle(vis, (x0, y0), (x1, y1), (0, 255, 0), 2)\r
+\r
+\r
+def grouper(n, iterable, fillvalue=None):\r
+    '''grouper(3, 'ABCDEFG', 'x') --> ABC DEF Gxx'''\r
+    args = [iter(iterable)] * n\r
+    return it.izip_longest(fillvalue=fillvalue, *args)\r
+\r
+def mosaic(w, imgs):\r
+    '''Make a grid from images. \r
+\r
+    w    -- number of grid columns\r
+    imgs -- images (must have same size and format)\r
+    '''\r
+    imgs = iter(imgs)\r
+    img0 = imgs.next()\r
+    pad = np.zeros_like(img0)\r
+    imgs = it.chain([img0], imgs)\r
+    rows = grouper(w, imgs, pad)\r
+    return np.vstack(map(np.hstack, rows))\r

diff --git a/samples/python2/digits.py b/samples/python2/digits.py
index c3494c0..f4bb0a5 100644 (file)
--- a/samples/python2/digits.py
+++ b/samples/python2/digits.py
@@ -1,89 +1,78 @@
-import numpy as np\r
-import cv2\r
-import itertools as it\r
-\r
 '''\r
-from scipy.io import loadmat\r
+Neural network digit recognition sample.\r
+Usage:\r
+   digits.py\r
 \r
-m = loadmat('ex4data1.mat')\r
-X = m['X'].reshape(-1, 20, 20)\r
-X = np.transpose(X, (0, 2, 1))\r
-img = np.vstack(map(np.hstack, X.reshape(-1, 100, 20, 20)))\r
-img = np.uint8(np.clip(img, 0, 1)*255)\r
-cv2.imwrite('digits.png', img)\r
+   Sample loads a dataset of handwritten digits from 'digits.png'.\r
+   Then it trains a neural network classifier on it and evaluates\r
+   its classification accuracy.\r
 '''\r
 \r
+import numpy as np\r
+import cv2\r
+from common import mosaic\r
+\r
 def unroll_responses(responses, class_n):\r
+    '''[1, 0, 2, ...] -> [[0, 1, 0], [1, 0, 0], [0, 0, 1], ...]'''\r
     sample_n = len(responses)\r
     new_responses = np.zeros((sample_n, class_n), np.float32)\r
     new_responses[np.arange(sample_n), responses] = 1\r
     return new_responses\r
     \r
 \r
-SZ = 20\r
+SZ = 20 # size of each digit is SZ x SZ\r
+CLASS_N = 10\r
 digits_img = cv2.imread('digits.png', 0)\r
 \r
+# prepare dataset\r
 h, w = digits_img.shape\r
 digits = [np.hsplit(row, w/SZ) for row in np.vsplit(digits_img, h/SZ)]\r
 digits = np.float32(digits).reshape(-1, SZ*SZ)\r
 N = len(digits)\r
-labels = np.repeat(np.arange(10), N/10)\r
+labels = np.repeat(np.arange(CLASS_N), N/CLASS_N)\r
 \r
+# split it onto train and test subsets\r
 shuffle = np.random.permutation(N)\r
 train_n = int(0.9*N)\r
-\r
 digits_train, digits_test = np.split(digits[shuffle], [train_n])\r
 labels_train, labels_test = np.split(labels[shuffle], [train_n])\r
 \r
-labels_train_unrolled = unroll_responses(labels_train, 10)\r
-\r
+# train model\r
 model = cv2.ANN_MLP()\r
-layer_sizes = np.int32([SZ*SZ, 25, 10])\r
+layer_sizes = np.int32([SZ*SZ, 25, CLASS_N])\r
 model.create(layer_sizes)\r
-        \r
-# CvANN_MLP_TrainParams::BACKPROP,0.001\r
-params = dict( term_crit = (cv2.TERM_CRITERIA_COUNT, 300, 0.01),\r
+params = dict( term_crit = (cv2.TERM_CRITERIA_COUNT, 100, 0.01),\r
                train_method = cv2.ANN_MLP_TRAIN_PARAMS_BACKPROP,\r
                bp_dw_scale = 0.001,\r
                bp_moment_scale = 0.0 )\r
 print 'training...'\r
+labels_train_unrolled = unroll_responses(labels_train, CLASS_N)\r
 model.train(digits_train, labels_train_unrolled, None, params=params)\r
 model.save('dig_nn.dat')\r
 model.load('dig_nn.dat')\r
 \r
-ret, resp = model.predict(digits_test)\r
-resp = resp.argmax(-1)\r
-error_mask = (resp == labels_test)\r
-print error_mask.mean()\r
-\r
-def grouper(n, iterable, fillvalue=None):\r
-    "grouper(3, 'ABCDEFG', 'x') --> ABC DEF Gxx"\r
-    args = [iter(iterable)] * n\r
-    return it.izip_longest(fillvalue=fillvalue, *args)\r
-\r
-def mosaic(w, imgs):\r
-    imgs = iter(imgs)\r
-    img0 = imgs.next()\r
-    pad = np.zeros_like(img0)\r
-    imgs = it.chain([img0], imgs)\r
-    rows = grouper(w, imgs, pad)\r
-    return np.vstack(map(np.hstack, rows))\r
-\r
-test_img = np.uint8(digits_test).reshape(-1, SZ, SZ)\r
-\r
-def vis_resp(img, flag):\r
+def evaluate(model, samples, labels):\r
+    '''Evaluates classifier preformance on a given labeled samples set.'''\r
+    ret, resp = model.predict(samples)\r
+    resp = resp.argmax(-1)\r
+    error_mask = (resp == labels)\r
+    accuracy = error_mask.mean()\r
+    return accuracy, error_mask\r
+\r
+# evaluate model\r
+train_accuracy, _ = evaluate(model, digits_train, labels_train)\r
+print 'train accuracy: ', train_accuracy\r
+test_accuracy, test_error_mask = evaluate(model, digits_test, labels_test)\r
+print 'test accuracy: ', test_accuracy\r
+\r
+# visualize test results\r
+vis = []\r
+for img, flag in zip(digits_test, test_error_mask):\r
+    img = np.uint8(img).reshape(SZ, SZ)\r
     img = cv2.cvtColor(img, cv2.COLOR_GRAY2BGR)\r
     if not flag:\r
         img[...,:2] = 0\r
-    return img\r
-\r
-test_img = mosaic(25, it.starmap(vis_resp, it.izip(test_img, error_mask)))\r
-cv2.imshow('test', test_img)\r
+    vis.append(img)\r
+vis = mosaic(25, vis)\r
+cv2.imshow('test', vis)\r
 cv2.waitKey()\r
-\r
-\r
-\r
-\r
-\r
-\r
-\r