Open TianyuLee opened 5 years ago
Hi @piergiaj
Thanks for sharing your code! I have also a similar question on pre-trained I3D classification results on Charades dataset.
I tried to test predictions by adding a prediction layer (Sigmoid) after logits (averaged) on Charades dataset. The rgb_charades.pt
model checkpoint seems to give decent (correct) predictions. However, the flow_charades.pt
model always gives approximately the same (some random) value.
Could you please confirm if the uploaded model checkpoint flow_charades.pt
is the correct one and has the expected performance?
Please let me know if you need more info or if I am missing anything!
I also think that a wrong checkpoint is uploaded as flow_charades.pt
. If anyone is still looking, feel free to use the following checkpoint (finetuned for 30k iterations and have ~14 mAP val) for the flow model.
flow_temp_030000.pt
Hello, Does someone has benchmarked this implementation? I could not find results anywhere and I don't know what performance I could expect from it.
Dear author: Have you tested the pytorch model in Kinetics or Charades dataset and if the results are close to the official tensorflow version? Best regards!