Haoqiang Guo

`training throughput` may not equal to `time per step`

1

[One](https://github.com/google-research/tuning_playbook#determining-the-feasible-batch-sizes-and-estimating-training-throughput) of the sections mentions that `training throughput` is equivalent to `time per step`. There is a doubt here. Suppose there are two kinds of `batch size`: `64` and `128`,...

> - All benefits of using a larger batch size assume the training throughput increases. If it doesn't, fix the bottleneck or use the smaller batch size. > - Gradient...

### 🐛 Describe the bug When I initialize a local git repository, executing the code will raise `ValueError: Remote named 'origin' didn't exist`. I checked the code and found the...

- File tree ``` . ├── components.py ├── config.jsonnet ├── debug.py ``` - The agent file called `debug.py`. ```python import sys import runpy import os base_path = os.path.dirname(__file__) os.chdir(base_path) args...

https://github.com/Das-Boot/scite/blob/1e082be765810f40e37c328f452d5ded69b06fa8/models/tag2triplet.py#L205 For `record.append(([np.abs(e[0]-e[1]) for e in i], list(i)))`, I think the code should be like that: `record.append((sum([np.abs(e[0]-e[1]) for e in i]), list(i)))`

Add another example that does not require downloading data from the Internet. ## 📚 Context In the caching example documents, one example needs to download data from GitHub, which can...