changeset 108:a7cd8dd3221c

pipeline.py: placé les modules dans le bon ordre + store NIST en bytes plutôt qu'en float32 dans la RAM
author boulanni <nicolas_boulanger@hotmail.com>
date Mon, 15 Feb 2010 16:17:48 -0500
parents a9b87b68101d
children 9c45e0071b52
files transformations/pipeline.py
diffstat 1 files changed, 12 insertions(+), 8 deletions(-) [+]
line wrap: on
line diff
--- a/transformations/pipeline.py	Mon Feb 15 15:35:17 2010 -0500
+++ b/transformations/pipeline.py	Mon Feb 15 16:17:48 2010 -0500
@@ -41,17 +41,19 @@
     import pylab
     pylab.ion()
 
-#from add_background_image import AddBackground
-from affine_transform import AffineTransformation
 from PoivreSel import PoivreSel
 from thick import Thick
 from BruitGauss import BruitGauss
+from DistorsionGauss import DistorsionGauss
+from PermutPixel import PermutPixel
 from gimp_script import GIMP1
 from Rature import Rature
 from contrast import Contrast
-from Occlusion import Occlusion
 from local_elastic_distortions import LocalElasticDistorter
 from slant import Slant
+from Occlusion import Occlusion
+from add_background_image import AddBackground
+from affine_transform import AffineTransformation
 
 if DEBUG:
     from visualizer import Visualizer
@@ -60,7 +62,8 @@
     # AFTER_EACH_MODULE_HOOK list (but not both, it's redundant)
     VISUALIZER = Visualizer(to_dir=DEBUG_OUTPUT_DIR,  on_screen=False)
 
-MODULE_INSTANCES = [Thick(),Slant(),GIMP1(),AffineTransformation(),LocalElasticDistorter(),Occlusion(),Rature(),Contrast()]
+###---------------------order of transformation module
+MODULE_INSTANCES = [Slant(),Thick(),AffineTransformation(),LocalElasticDistorter(),GIMP1(),Rature(),Occlusion(), PermutPixel(),DistorsionGauss(),AddBackground(), PoivreSel(), BruitGauss(), Contrast()]
 
 # These should have a "after_transform_callback(self, image)" method
 # (called after each call to transform_image in a module)
@@ -201,11 +204,12 @@
         self.ocr_data = open(ocr_path, 'rb')
         self.ocr_labels = open(ocrlabel_path, 'rb')
 
+# cet iterator load tout en ram
 def nist_supp_iterator(nist, prob_font, prob_captcha, prob_ocr, num_img):
-    img = ft.read(nist.train_data).astype(numpy.float32)/255
+    img = ft.read(nist.train_data)
     labels = ft.read(nist.train_labels)
     if prob_ocr:
-        ocr_img = ft.read(nist.ocr_data).astype(numpy.float32)/255
+        ocr_img = ft.read(nist.ocr_data)
         ocr_labels = ft.read(nist.ocr_labels)
 
     for i in xrange(num_img):
@@ -216,10 +220,10 @@
             pass #get captcha
         elif r <= prob_font + prob_captcha + prob_ocr:
             j = numpy.random.randint(len(ocr_labels))
-            yield ocr_img[j], ocr_labels[j]
+            yield ocr_img[j].astype(numpy.float32)/255, ocr_labels[j]
         else:
             j = numpy.random.randint(len(labels))
-            yield img[j], labels[j]
+            yield img[j].astype(numpy.float32)/255, labels[j]
 
 
 # Mostly for debugging, for the moment, just to see if we can