nicolabs/Open-Sora
1
0
Fork 0
mirror of https://github.com/hpcaitech/Open-Sora.git synced 2026-05-21 11:59:01 +02:00
Code Issues Projects Releases Packages Wiki Activity Actions 1

[docs] update docs for bs search

Browse source
This commit is contained in:
zhengzangw 2024-05-15 13:13:17 +00:00
parent 0592723022
commit b73a8050fa
1 changed files with 27 additions and 29 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

56
docs/commands.md
View file

@ -169,44 +169,42 @@ You can modify corresponding config files to change the training settings. See m
To search the batch size for buckets, run the following command. To search the batch size for buckets, run the following command.
```bash ```bash
torchrun --standalone --nproc_per_node 1 scripts/search_bs.py configs/opensora-v1-1/train/benchmark.py --data-path YOUR_CSV_PATH -o YOUR_OUTPUT_CONFIG_PATH --base-resolution 240p --base-frames 128 --batch-size-start 2 --batch-size-end 256 --batch-size-step 2 torchrun --standalone --nproc_per_node 1 scripts/misc/search_bs.py configs/opensora-v1-2/misc/bs.py --data-path /mnt/nfs-207/sora_data/meta/searchbs.csv
``` ```
If your dataset is extremely large, you extract a subset of the dataset for the search. Here, your data should be a small one for searching purposes.
```bash To control the batch size search range, you should specify `bucket_config` in the config file, where the value tuple is `(guess_value, range)` and the search will be performed in `guess_value±range`.
# each bucket contains 1000 samples
python tools/datasets/split.py YOUR_CSV_PATH -o YOUR_SUBSET_CSV_PATH -c configs/opensora-v1-1/train/video.py -l 1000
```
If you want to control the batch size search more granularly, you can configure batch size start, end, and step in the config file.
Bucket config format:
1. `{ resolution: {num_frames: (prob, batch_size)} }`, in this case batch_size is ignored when searching
2. `{ resolution: {num_frames: (prob, (max_batch_size, ))} }`, batch_size is searched in the range `[batch_size_start, max_batch_size)`, batch_size_start is configured via CLI
3. `{ resolution: {num_frames: (prob, (min_batch_size, max_batch_size))} }`, batch_size is searched in the range `[min_batch_size, max_batch_size)`
4. `{ resolution: {num_frames: (prob, (min_batch_size, max_batch_size, step_size))} }`, batch_size is searched in the range `[min_batch_size, max_batch_size)` with step_size (grid search)
5. `{ resolution: {num_frames: (0.0, None)} }`, this bucket will not be used
Here is an example of the bucket config: Here is an example of the bucket config:
```python ```python
bucket_config = { bucket_config = {
"240p": {
"240p": { 1: (100, 100),
16: (1.0, (2, 32)), 51: (24, 10),
32: (1.0, (2, 16)), 102: (12, 10),
64: (1.0, (2, 8)), 204: (4, 8),
128: (1.0, (2, 6)), 408: (2, 8),
},
"480p": {
1: (50, 50),
51: (6, 6),
102: (3, 3),
204: (1, 2),
}, },
"256": {1: (1.0, (128, 300))},
"512": {1: (0.5, (64, 128))},
"480p": {1: (0.4, (32, 128)), 16: (0.4, (2, 32)), 32: (0.0, None)},
"720p": {16: (0.1, (2, 16)), 32: (0.0, None)}, # No examples now
"1024": {1: (0.3, (8, 64))},
"1080p": {1: (0.3, (2, 32))},
} }
``` ```
It will print the best batch size (and corresponding step time) for each bucket and save the output config file. You can also specify a resolution to search for parallelism.
```bash
torchrun --standalone --nproc_per_node 1 scripts/misc/search_bs.py configs/opensora-v1-2/misc/bs.py --data-path /mnt/nfs-207/sora_data/meta/searchbs.csv --resolution 240p
```
The searching goal should be specified in the config file as well. There are two ways:
1. Specify a `base_step_time` in the config file. The searching goal is to find the batch size that can achieve the `base_step_time` for each bucket.
2. If `base_step_time` is not specified, it will be determined by `base` which is a tuple of `(batch_size, step_time)`. The step time is the maximum batch size allowed for the bucket.
The script will print the best batch size (and corresponding step time) for each bucket and save the output config file. Note that we assume a larger batch size is better, so the script use binary search to find the best batch size.