Fix ignoring short videos in VideoDataset

pytorch/nvvl/dataset.py

@@ -186,7 +186,6 @@ class VideoDataset(torch.utils.data.Dataset):
     def __init__(self, filenames, sequence_length, device_id=0,
                  get_label=None, processing=None, log_level="warn"):
         self.ffi = lib._ffi
-        self.filenames = filenames
         self.sequence_length = sequence_length
         self.device_id = device_id
         self.get_label = get_label if get_label is not None else lambda x,y,z: None
@@ -203,9 +202,6 @@ def __init__(self, filenames, sequence_length, device_id=0,
             print("Invalid log level", log_level, "using warn.", file=sys.stderr)
             log_level = lib.LogLevel_Warn
 
-        if not filenames:
-            raise ValueError("Empty filenames list given to VideoDataset")
-
         if sequence_length < 1:
             raise ValueError("Sequence length must be at least 1")
 
@@ -214,17 +210,22 @@ def __init__(self, filenames, sequence_length, device_id=0,
         self.total_frames = 0
         self.frame_counts = []
         self.start_index = []
+        self.filenames = []
         for f in filenames:
             count = lib.nvvl_frame_count(self.loader, str.encode(f));
             if count < self.sequence_length:
                 print("NVVL WARNING: Ignoring", f, "because it only has", count,
                       "frames and the sequence length is", self.sequence_length)
                 continue
+            self.filenames.append(f)
             count = count - self.sequence_length + 1
             self.frame_counts.append(count)
             self.total_frames += count
             self.start_index.append(self.total_frames) # purposefully off by one for bisect to work
 
+        if not filenames:
+            raise ValueError("Empty filenames list given to VideoDataset")
+
         size = lib.nvvl_video_size(self.loader)
         self.width = size.width
         self.height = size.height