Training on Ego4d NLQ without Naq

[TensorsSun]

Thanks for your great work.
When I try to train on Ego4d NLQ without Naq, I am consistently 2% behind the reported performance. I pasted my best result below.

+Window Pre-filtering Epoch 124-------+---------+---------+
| Rank@1 | Rank@5 | Rank@10 | Rank@20 | Rank@30 | Rank@50 |
+--------+--------+---------+---------+---------+---------+
| 35.23  | 70.26  |  87.79  |  98.19  |  99.20  |  100.00 |
+--------+--------+---------+---------+---------+---------++Fusion Epoch 124-----+----------+----------+----------+----------+----------+----------+----------+----------+-------+
|  Rank@1  |  Rank@1  |  Rank@5  |  Rank@5  | Rank@10  | Rank@10  | Rank@50  | Rank@50  | Rank@100 | Rank@100 | mIoU  |
| mIoU@0.3 | mIoU@0.5 | mIoU@0.3 | mIoU@0.5 | mIoU@0.3 | mIoU@0.5 | mIoU@0.3 | mIoU@0.5 | mIoU@0.3 | mIoU@0.5 |       |
+----------+----------+----------+----------+----------+----------+----------+----------+----------+----------+-------+
|  16.70   |  11.13   |  34.90   |  22.69   |  34.90   |  22.69   |  34.90   |  22.69   |  34.90   |  22.69   | 11.33 |
+----------+----------+----------+----------+----------+----------+----------+----------+----------+----------+-------++Proposal Epoch 124---+----------+----------+----------+----------+----------+----------+----------+----------+------+
|  Rank@1  |  Rank@1  |  Rank@5  |  Rank@5  | Rank@10  | Rank@10  | Rank@50  | Rank@50  | Rank@100 | Rank@100 | mIoU |
| mIoU@0.3 | mIoU@0.5 | mIoU@0.3 | mIoU@0.5 | mIoU@0.3 | mIoU@0.5 | mIoU@0.3 | mIoU@0.5 | mIoU@0.3 | mIoU@0.5 |      |
+----------+----------+----------+----------+----------+----------+----------+----------+----------+----------+------+
|  13.96   |   9.50   |  31.62   |  20.70   |  31.62   |  20.70   |  31.62   |  20.70   |  31.62   |  20.70   | 9.44 |
+----------+----------+----------+----------+----------+----------+----------+----------+----------+----------+------++Matching Epoch 124---+----------+----------+----------+----------+----------+----------+----------+----------+------+
|  Rank@1  |  Rank@1  |  Rank@5  |  Rank@5  | Rank@10  | Rank@10  | Rank@50  | Rank@50  | Rank@100 | Rank@100 | mIoU |
| mIoU@0.3 | mIoU@0.5 | mIoU@0.3 | mIoU@0.5 | mIoU@0.3 | mIoU@0.5 | mIoU@0.3 | mIoU@0.5 | mIoU@0.3 | mIoU@0.5 |      |
+----------+----------+----------+----------+----------+----------+----------+----------+----------+----------+------+
|  11.05   |   7.10   |  26.64   |  16.93   |  26.64   |  16.93   |  26.64   |  16.93   |  26.64   |  16.93   | 8.89 |
+----------+----------+----------+----------+----------+----------+----------+----------+----------+----------+------+

I’m not sure if this is caused by some incorrect parameter settings. Could you please provide the opt.json of your best result on Ego4d NLQ without Naq?

[Tanveer81]

Hi, Thank you for your interest in our work. We use the provided training scripts, and the checkpoints we provided were trained using these scripts as well. During training the opt.json file is automatically generated.