# Update

For multi-scale test work correct

Author: SnowMasaya

datasets/pascal_voc.py

@@ -235,7 +235,7 @@ def _do_python_eval(self, output_dir='output'):
             aps += [ap]
             print(('AP for {} = {:.4f}'.format(cls, ap)))
             if output_dir is not None:
-                with open(os.path.join(output_dir, cls + '_pr.pkl'), 'w') as f:
+                with open(os.path.join(output_dir, cls + '_pr.pkl'), 'wb') as f:
                     pickle.dump({'rec': rec, 'prec': prec, 'ap': ap}, f)
         print(('Mean AP = {:.4f}'.format(np.mean(aps))))
         print('~~~~~~~~')

datasets/voc_eval.py

@@ -115,11 +115,11 @@ def voc_eval(detpath,
                     i + 1, len(imagenames)))
         # save
         print('Saving cached annotations to {:s}'.format(cachefile))
-        with open(cachefile, 'w') as f:
+        with open(cachefile, 'wb') as f:
             pickle.dump(recs, f)
     else:
         # load
-        with open(cachefile, 'r') as f:
+        with open(cachefile, 'rb') as f:
             recs = pickle.load(f)
 
     # extract gt objects for this class

test.py

@@ -2,6 +2,7 @@
 import cv2
 import numpy as np
 import pickle
+import argparse
 
 from darknet import Darknet19
 import utils.yolo as yolo_utils
@@ -18,6 +19,13 @@ def preprocess(fname):
     return image, im_data
 
 
+parser = argparse.ArgumentParser(description='PyTorch Yolo')
+parser.add_argument('--image_size_index', type=int, default=0,
+                    metavar='image_size_index',
+                    help='setting images size index 0:320, 1:352, 2:384, 3:416, 4:448, 5:480, 6:512, 7:544, 8:576')
+args = parser.parse_args()
+
+
 # hyper-parameters
 # ------------
 imdb_name = cfg.imdb_test
@@ -44,10 +52,11 @@ def test_net(net, imdb, max_per_image=300, thresh=0.5, vis=False):
     # timers
     _t = {'im_detect': Timer(), 'misc': Timer()}
     det_file = os.path.join(output_dir, 'detections.pkl')
+    size_index = args.image_size_index
 
     for i in range(num_images):
 
-        batch = imdb.next_batch()
+        batch = imdb.next_batch(size_index=size_index)
         ori_im = batch['origin_im'][0]
         im_data = net_utils.np_to_variable(batch['images'], is_cuda=True,
                                            volatile=True).permute(0, 3, 1, 2)
@@ -65,7 +74,9 @@ def test_net(net, imdb, max_per_image=300, thresh=0.5, vis=False):
                                                           prob_pred,
                                                           ori_im.shape,
                                                           cfg,
-                                                          thresh)
+                                                          thresh,
+                                                          size_index
+                                                          )
         detect_time = _t['im_detect'].toc()
 
         _t['misc'].tic()
@@ -122,7 +133,7 @@ def test_net(net, imdb, max_per_image=300, thresh=0.5, vis=False):
     # data loader
     imdb = VOCDataset(imdb_name, cfg.DATA_DIR, cfg.batch_size,
                       yolo_utils.preprocess_test,
-                      processes=2, shuffle=False, dst_size=cfg.inp_size)
+                      processes=2, shuffle=False, dst_size=cfg.multi_scale_inp_size)
 
     net = Darknet19()
     net_utils.load_net(trained_model, net)

utils/yolo.py

@@ -78,9 +78,10 @@ def preprocess_train(data, size_index):
     return im, boxes, gt_classes, [], ori_im
 
 
-def preprocess_test(data):
+def preprocess_test(data, size_index):
 
     im, _, inp_size = data
+    inp_size = inp_size[size_index]
     if isinstance(im, str):
         im = cv2.imread(im)
     ori_im = np.copy(im)
@@ -94,7 +95,8 @@ def preprocess_test(data):
     return im, [], [], [], ori_im
 
 
-def postprocess(bbox_pred, iou_pred, prob_pred, im_shape, cfg, thresh=0.05):
+def postprocess(bbox_pred, iou_pred, prob_pred, im_shape, cfg, thresh=0.05,
+                size_index=0):
     """
     bbox_pred: (bsize, HxW, num_anchors, 4)
                ndarray of float (sig(tx), sig(ty), exp(tw), exp(th))
@@ -105,7 +107,7 @@ def postprocess(bbox_pred, iou_pred, prob_pred, im_shape, cfg, thresh=0.05):
     # num_classes, num_anchors = cfg.num_classes, cfg.num_anchors
     num_classes = cfg.num_classes
     anchors = cfg.anchors
-    W, H = cfg.out_size
+    W, H = cfg.multi_scale_out_size[size_index]
     assert bbox_pred.shape[0] == 1, 'postprocess only support one image per batch'  # noqa
 
     bbox_pred = yolo_to_bbox(