add random shift to training

config.yaml

@@ -4,7 +4,7 @@ general:
   checkpoint_dir: 'checkpoint'
   train_dir: './data/train'
   batch_size: 1 # data loader is only implemented for case == 1!
-  select: 'model3'
+  select: 'model1'
   test_file: './data/test/example1.mkv'
 
 model1:
@@ -17,8 +17,8 @@ model1:
   convlstm_hidden_dim: 64
   seq_len: 8
   lr: 0.0001
-  lr_milestones: [7, 12, 15]
-  epochs: 25
+  lr_milestones: [4, 7, 9]
+  epochs: 10
 
 model2:
   name: 'funpos2'

utils/dataset.py

@@ -86,13 +86,23 @@ def load_next_video_frames(self):
             self.read_next_frame()
 
 
+    def get_uniform_random_int(self, lower, upper):
+        return torch.randint(lower, upper+1, (1,)).numpy()[0]
+
+
     def open_next_video(self):
         self.inc_video_idx()
         with open("".join(self.videos[self.video_idx][:-4]) + '.param', "r") as f:
             self.param = json.load(f)
         with open("".join(self.videos[self.video_idx][:-4]) + '.labels', "r") as f:
             l = json.load(f)
             self.labels = {int(k):l[k] for k in l.keys()}
+        self.param['zoom'] = [
+            self.param['zoom'][0] + self.get_uniform_random_int(-10, 10),
+            self.param['zoom'][1] + self.get_uniform_random_int(-10, 10),
+            self.param['zoom'][2] + self.get_uniform_random_int(-10, 10),
+            self.param['zoom'][3] + self.get_uniform_random_int(-10, 10)
+        ]
         self.param['resize'] = (self.img_width, self.img_height)
         if self.stream is not None:
             self.stream.stop()