Possible label corruption problem

[wishforgood]

What I have done: I have downloaded your data from https://drive.google.com/file/d/1_ObEsWHHOrzAwLf5OvdY-64r3Ne4nTcR/view?usp=sharing. Then I run the code train_graph_VG and test_graph_VG to get the predictions. I followed the order in file "torchdata/pruned_graph_cocosplit.t7" to parse binary labels to be words. What the problem is: I found there seems to be hundreds of images with wrong labels. The wrong labels always looks like ['person', 'tree', 'building', 'sky', 'grass', 'head', 'hand', 'hair', 'people', 'clouds', 'snow', 'hat', 'dog', 'boat', 'bird', 'cow', 'field', 'sheep', 'rock', 'coat', 'book', 'sports ball', 'dining table', 'cell phone', 'refrigerator'] or something similar. These things can hardly appear in a single image. And I have checked many examples that the labels are not consistent with the fact.

If this is exactly the dataset as described in the paper, then it's OK, I'll ignored the questionable images. If not, could you please repair it? If I have parsed the labels in a wrong way, could you please provide some instructions on transforming the 316-dim vectors to meaningful words?