Duo Zheng
Duo Zheng
Sry, we previously tried to adopt `opt-3b` as the backbone, but we did not store the corresponding weights. For your question, I believe that with specific optimization such as quantization,...
Could you please explain in detail where the error occurs?
The actual training step is equal to `num_steps_per_epoch * num_epochs / gradient_accumulation_step `. If you set `gradient_accumulation_step` to 32, you need to scale either `num_steps_per_epoch` or `num_epochs` by 4 times.
The results on the CVDN Val set results are 6.16 and 5.91 in the chart and log, respectively, which is within a reasonable range.
Yes, sorry for this mistake. Actually, for CVDN, Goal Progress (`dist_to_end_reduction`) is adopted as the primary metric.
Hi, what are the specific numbers on REVERIE? Have you tested the results on other datasets?
This result seems a bit weird, but your evaluation script is working fine. I’m not yet able to pinpoint where the issue might be.
Here is one sample for the submission. ``` { "instr_id": 0, "trajectory": [ { "path": [ [ "073656b4030d4389ba7f304298be1d73", 0, 0], [ "06addff1d8274747b7a1957b2f03b736", 0, 0], [ "3be66dc4f0fb46449d63b07d92c74f3f", 0, 0], [ "548d13fef345430ba0ca030f8edf4543",...