normalize image by per-color mean and std(#22)

theanompi/models/alex_net.py

@@ -419,9 +419,10 @@ def train_iter(self, count,recorder):
 
             else:
 
-                img_mean = self.data.rawdata[-1]
+                img_mean = self.data.rawdata[4]
+                img_std = self.data.rawdata[5]
                 import hickle as hkl
-                arr = hkl.load(img[self.current_t]) - img_mean
+                arr = (hkl.load(img[self.current_t]) - img_mean)/255./img_std
 
                 from theanompi.models.data.utils import crop_and_mirror
 
@@ -520,9 +521,10 @@ def val_iter(self, count,recorder):
             else:
 
 
-                img_mean = self.data.rawdata[-1]
+                img_mean = self.data.rawdata[4]
+                img_std = self.data.rawdata[5]
                 import hickle as hkl
-                arr = hkl.load(img[self.current_v]) - img_mean
+                arr = (hkl.load(img[self.current_v]) - img_mean)/255./img_std
 
                 from theanompi.models.data.utils import crop_and_mirror
 

theanompi/models/be_model.py

@@ -40,10 +40,11 @@
 
     # load data
 
-    img_mean=model.data.rawdata[-1]
+    img_mean=model.data.rawdata[4]
+    img_std=model.data.rawdata[5]
 
     import hickle as hkl
-    arr = hkl.load(model.data.train_img_shard[0]) - img_mean
+    arr = (hkl.load(model.data.train_img_shard[0]) - img_mean)/255./img_std
 
     from theanompi.models.data.utils import crop_and_mirror
 

theanompi/models/data/imagenet.py

@@ -17,15 +17,17 @@
 
 sc=False # if the architecture already includes subtract mean and cropping
 
+debug=False
+
 class ImageNet_data(object):
 
     def __init__(self, verbose):
 
         # data hyperparams
 
-        self.data_path  = '/scratch/ilsvrc12/'
-        self.train_folder = 'train_hkl_b256_b_128'
-        self.val_folder = 'val_hkl_b256_b_128'
+        self.data_path  = dir_head #'/scratch/ilsvrc12/'
+        self.train_folder = train_folder #'train_hkl_b256_b_128/'
+        self.val_folder = val_folder #'val_hkl_b256_b_128/'
 
         self.channels = 3
         self.width =256
@@ -89,11 +91,14 @@ def get_data(self, file_batch_size=128):
             #print 'BGR_mean %s' % image_mean #[ 122.22585297  116.20915222  103.56548309]
             import numpy as np
             image_mean = image_mean[:,np.newaxis,np.newaxis,np.newaxis]
+
+        img_std = np.array([0.229, 0.224, 0.225]).astype(np.float32)
+        img_std = img_std[:,np.newaxis,np.newaxis,np.newaxis]
 
 
 
         self.rawdata = [train_filenames,train_labels,\
-                    val_filenames,val_labels,img_mean] # 5 items
+                    val_filenames,val_labels,img_mean,img_std] # 6 items
 
 
 
@@ -286,6 +291,7 @@ def para_load_init(self, shared_x, input_width, input_height,
         config['rand_crop'] = rand_crop
         config['batch_crop_mirror'] = batch_crop_mirror
         config['img_mean'] = self.rawdata[4]
+        config['img_std'] = self.rawdata[5]
 
         import os
         _sock_data = ((sock_data + int(os.getpid())) % 64511)+1024

theanompi/models/data/proc_load_mpi.py

@@ -33,6 +33,7 @@
     rand_crop  =    config['rand_crop']        
     batch_crop_mirror  =    config['batch_crop_mirror']
     img_mean = config['img_mean']
+    img_std=config['img_std']
     import os
     if "CPULIST_train" in os.environ:
         cpulist = os.environ['CPULIST_train']
@@ -95,7 +96,7 @@
 
             arr = hkl.load(str(filename)).astype('float32')
 
-            arr = arr - img_mean
+            arr = (arr - img_mean)/255./img_std
 
             arr = crop_and_mirror(arr, mode, 
                                 rand_crop, 

theanompi/models/googlenet.py

@@ -766,9 +766,10 @@ def train_iter(self, count, recorder):
 
             else:
 
-                img_mean = self.data.rawdata[-1]
+                img_mean = self.data.rawdata[4]
+                img_std = self.data.rawdata[5]
                 import hickle as hkl
-                arr = hkl.load(img[self.current_t]) - img_mean
+                arr = (hkl.load(img[self.current_t]) - img_mean)/255./img_std
 
                 from theanompi.models.data.utils import crop_and_mirror
 
@@ -866,9 +867,10 @@ def val_iter(self, count,recorder):
             else:
 
 
-                img_mean = self.data.rawdata[-1]
+                img_mean = self.data.rawdata[4]
+                img_std = self.data.rawdata[5]
                 import hickle as hkl
-                arr = hkl.load(img[self.current_v]) - img_mean
+                arr = (hkl.load(img[self.current_v]) - img_mean)/255./img_std
 
                 from theanompi.models.data.utils import crop_and_mirror
 

theanompi/models/lasagne_model_zoo/resnet50.py

@@ -220,6 +220,8 @@ def build_model_resnet50(input_shape):
 
 batch_crop_mirror = True 
 rand_crop = True
+
+image_mean = np.array([103.939, 116.779, 123.68],dtype='float32')[:,np.newaxis,np.newaxis,np.newaxis]
 dataname = 'imagenet'
 
 monitor_grad = False
@@ -254,6 +256,7 @@ def __init__(self, config):
         # data
         from theanompi.models.data import ImageNet_data
         self.data = ImageNet_data(verbose=False)
+        self.data.rawdata[4] = image_mean
         self.channels = self.data.channels # 'c' mean(R,G,B) = (103.939, 116.779, 123.68)
         self.input_width = input_width # '0' single scale training 224
         self.input_height = input_height # '1' single scale training 224
@@ -494,9 +497,10 @@ def train_iter(self, count,recorder):
 
             else:
 
-                img_mean = self.data.rawdata[-1]
+                img_mean = self.data.rawdata[4]
+                img_std = self.data.rawdata[5]
                 import hickle as hkl
-                arr = hkl.load(img[self.current_t]) - img_mean
+                arr = (hkl.load(img[self.current_t]) - img_mean)/255./img_std
 
                 from theanompi.models.data.utils import crop_and_mirror
 
@@ -595,10 +599,11 @@ def val_iter(self, count,recorder):
             else:
 
 
-                img_mean = self.data.rawdata[-1]
+                img_mean = self.data.rawdata[4]
+                img_std = self.data.rawdata[5]
 
                 import hickle as hkl
-                arr = hkl.load(img[self.current_v]) - img_mean
+                arr = (hkl.load(img[self.current_v]) - img_mean)/255./img_std
 
                 from theanompi.models.data.utils import crop_and_mirror
 

theanompi/models/lasagne_model_zoo/vgg16.py

@@ -393,9 +393,10 @@ def train_iter(self, count,recorder):
 
             else:
 
-                img_mean = self.data.rawdata[-1]
+                img_mean = self.data.rawdata[4]
+                img_std = self.data.rawdata[5]
                 import hickle as hkl
-                arr = hkl.load(img[self.current_t]) - img_mean
+                arr = (hkl.load(img[self.current_t]) - img_mean)/255./img_std
 
                 from theanompi.models.data.utils import crop_and_mirror
 
@@ -493,9 +494,10 @@ def val_iter(self, count,recorder):
             else:
 
 
-                img_mean = self.data.rawdata[-1]
+                img_mean = self.data.rawdata[4]
+                img_std = self.data.rawdata[5]
                 import hickle as hkl
-                arr = hkl.load(img[self.current_v]) - img_mean
+                arr = (hkl.load(img[self.current_v]) - img_mean)/255./img_std
 
                 from theanompi.models.data.utils import crop_and_mirror