Do not augment when validating

config/smaller.ini

@@ -0,0 +1,45 @@
+[info]
+experiment: U-net default settings
+name: unet
+
+[dataset]
+pixels: 512
+n_classes: 2
+channels: 1
+subset: None
+filenames_train: './../data/1_1_1mm_512_x_512_lung_slices/subset[0-7]/*.pkl.gz'
+filenames_validation: './../data/1_1_1mm_512_x_512_lung_slices/subset[8]/*.pkl.gz'
+
+[network]
+architecture: unet
+input_size: 512
+depth: 4
+branching_factor: 4
+
+[updates]
+optimization: nesterov
+learning_rate: 0.00001
+momentum: 0.985
+l2_lambda: 1e-4
+batch_size_train: 2
+batch_size_validation: 4
+n_epochs: 200
+
+[preprocessing]
+erode_segmentation: 3
+
+[normalization]
+zero_center: True
+mean_pixel: 0.66200809792889126
+
+[augmentation]
+augment: True
+flip: True
+zoom: 0.08 ;Not working yet
+rotation: 16
+translation: 3
+
+
+[misc]
+multiprocess_load_augmentation: False
+save_every_n_epoch: 4

src/dataset.py

@@ -13,14 +13,14 @@
 
 _EPSILON = 1e-8
 
-def get_image(filename):
+def get_image(filename, deterministic):
     with gzip.open(filename,'rb') as f:
         lung = pickle.load(f)
 
     with gzip.open(filename.replace('lung','nodule'),'rb') as f:
         truth = np.array(pickle.load(f),dtype=np.float32)
 
-    if P.AUGMENT:
+    if P.AUGMENT and not deterministic:
         lung, truth = augment([lung,truth])
 
     truth = np.array(np.round(truth),dtype=np.int64)
@@ -52,9 +52,8 @@ def get_image(filename):
 
     return lung, truth
 
-def load_images(filenames):
-    filenames = filter(lambda x: x!='.', filenames)
-    slices = map(get_image, filenames)
+def load_images(filenames, deterministic=False):
+    slices = [get_image(filename, deterministic) for filename in filenames]
     lungs, truths = zip(*slices)
 
     l = np.concatenate(lungs,axis=0)
@@ -64,7 +63,7 @@ def load_images(filenames):
     # get set to 0 (the background is -1)
     w = loss_weighting.weight_by_class_balance(t, classes=[0,1])
 
-    #Set -1 labels back to label 0
+    #Set -10 labels back to label 0
     t = np.clip(t, 0, 100000)
 
     return l, t, w

src/learn.py

@@ -103,7 +103,7 @@
         val_batches = 0
 
         np.random.shuffle(filenames_val)
-        val_gen = ParallelBatchIterator(load_images, filenames_val, ordered=False,
+        val_gen = ParallelBatchIterator(partial(load_images,deterministic=True), filenames_val, ordered=False,
                                             batch_size=P.BATCH_SIZE_VALIDATION,
                                             multiprocess=P.MULTIPROCESS_LOAD_AUGMENTATION)