Apply new pixel means

lib/core/config.py

@@ -513,10 +513,8 @@ __C.USE_NCCL = True
 # Hosts for Inter-Machine communication
 # Hosts for Inter-Machine communication
 __C.HOSTS = []
 __C.HOSTS = []
 
 
-# Pixel mean values (BGR order) as a (1, 1, 3) array
-# We use the same pixel mean for all networks even though it's not exactly what
-# they were trained with
-__C.PIXEL_MEANS = np.array([[[102.9801, 115.9465, 122.7717]]])
+# Pixel mean values (BGR order)
+__C.PIXEL_MEANS = [102., 115., 122.]
 
 
 # Default weights on (dx, dy, dw, dh) for normalizing bbox regression targets
 # Default weights on (dx, dy, dw, dh) for normalizing bbox regression targets
 # These are empirically chosen to approximately lead to unit variance targets
 # These are empirically chosen to approximately lead to unit variance targets

lib/modeling/base.py

@@ -35,7 +35,7 @@ class Bootstarp(torch.nn.Module):
             'arguments': {
             'arguments': {
                 'dtype': self.dtype,
                 'dtype': self.dtype,
                 'data_format': 'NCHW',
                 'data_format': 'NCHW',
-                'mean_values': [102.9801, 115.9465, 122.7717],
+                'mean_values': cfg.PIXEL_MEANS,
             }
             }
         }
         }
 
 

lib/modeling/detector.py

@@ -109,8 +109,8 @@ class Detector(torch.nn.Module):
         #   1) NHWC => NCHW
         #   1) NHWC => NCHW
         #   2) Uint8 => Float32 or Float16
         #   2) Uint8 => Float32 or Float16
         #   3) Mean subtraction
         #   3) Mean subtraction
-        processed_data = self.bootstarp(inputs['data'])
-        features = self.body(processed_data)
+        image_data = self.bootstarp(inputs['data'])
+        features = self.body(image_data)
 
 
         # 2. Apply the FPN to enhance features if necessary
         # 2. Apply the FPN to enhance features if necessary
         if hasattr(self, 'fpn'):
         if hasattr(self, 'fpn'):

lib/modeling/retinanet.py

@@ -43,7 +43,7 @@ class RetinaNet(torch.nn.Module):
         self.cls_score = conv3x3(dim_in, self.C * A, bias=True)
         self.cls_score = conv3x3(dim_in, self.C * A, bias=True)
         self.bbox_pred = conv3x3(dim_in, 4 * A, bias=True)
         self.bbox_pred = conv3x3(dim_in, 4 * A, bias=True)
         self.cls_prob = torch.nn.Sigmoid(inplace=True)
         self.cls_prob = torch.nn.Sigmoid(inplace=True)
-        self.relu = torch.nn.ELU(inplace=True)
+        self.relu = torch.nn.ReLU(inplace=True)
         self.decoder = RetinaNetDecoder()
         self.decoder = RetinaNetDecoder()
 
 
         ########################################
         ########################################

lib/ssd/data/preprocessing/expand.py

@@ -44,7 +44,7 @@ class Expander(object):
 
 
         new_im = np.empty((expand_h, expand_w, 3), dtype=np.uint8)
         new_im = np.empty((expand_h, expand_w, 3), dtype=np.uint8)
         new_im[:] = cfg.PIXEL_MEANS
         new_im[:] = cfg.PIXEL_MEANS
-        new_im[h_off: h_off + im_h, w_off: w_off + im_w, :] = im
+        new_im[h_off : h_off + im_h, w_off : w_off + im_w, :] = im
 
 
         if gt_boxes is not None:
         if gt_boxes is not None:
             ex_gt_boxes = gt_boxes.astype(gt_boxes.dtype, copy=True)
             ex_gt_boxes = gt_boxes.astype(gt_boxes.dtype, copy=True)

lib/utils/blob.py

@@ -41,8 +41,8 @@ def im_list_to_blob(ims):
     blob = np.empty(unify_shape, dtype=np.uint8)
     blob = np.empty(unify_shape, dtype=np.uint8)
     blob[:] = cfg.PIXEL_MEANS
     blob[:] = cfg.PIXEL_MEANS
 
 
-    for idx, im in enumerate(ims):
-        blob[idx, 0:im.shape[0], 0:im.shape[1], :] = im
+    for i, im in enumerate(ims):
+        blob[i, :im.shape[0], :im.shape[1], :] = im
 
 
     return blob
     return blob