Yongrae Jo

Annotation of logs

1

I'd like to annotate on logs, especially what epoch a log belongs to. Iteration-level (or each nth-iteration) logging makes it hard to know what epoch a specific data point is....

Hello, I'm trying to reproduce your training results using the train.py script with the same data. I think using clip scores for filtering data is essential, as I found out...

To do beam search, I set cond_scale=1.0, turned off do_sample, and set num_beams > 1, then called model.generate(). However, I got the following error. ``` TypeError: reshape total size must...

Is it meaningful to put both input vectors and output vectors on the same plot

1

While playing and learning with your visualization, I noticed input vectors and output vectors are placed in the same space, and plotted on the same plot (after applying PCA). Input...

[Vid2Seq] Reproduction of the paper results

18

I followed the instructions in [README](https://github.com/google-research/scenic/tree/main/scenic/projects/vid2seq#training) to evaluate the released checkpoints, but I could not reproduce the results on the paper. The paper says a fully fine-tuned Vid2Seq achieves 7.9...