History

Hongxin Liu 3bf09efe74 [booster] update prepare dataloader method for plugin (#3706 ) * [booster] add prepare dataloader method for plug * [booster] update examples and docstr		2023-05-08 15:44:03 +08:00
..
README.md	[example] add train resnet/vit with booster example (#3694 )	2023-05-08 10:42:30 +08:00
data.py	[booster] update prepare dataloader method for plugin (#3706 )	2023-05-08 15:44:03 +08:00
finetune.py	[example] add finetune bert with booster example (#3693 )	2023-05-06 11:53:13 +08:00
requirements.txt	[example] add train resnet/vit with booster example (#3694 )	2023-05-08 10:42:30 +08:00
test_ci.sh	[example] add train resnet/vit with booster example (#3694 )	2023-05-08 10:42:30 +08:00

README.md

Finetune BERT on GLUE

🚀 Quick Start

This example provides a training script, which provides an example of finetuning BERT on GLUE dataset.

Training Arguments
- -t, --task: GLUE task to run. Defaults to mrpc.
- -p, --plugin: Plugin to use. Choices: torch_ddp, torch_ddp_fp16, gemini, low_level_zero. Defaults to torch_ddp.
- --target_f1: Target f1 score. Raise exception if not reached. Defaults to None.

Install requirements

pip install -r requirements.txt

Train

# train with torch DDP with fp32
colossalai run --nproc_per_node 4 finetune.py

# train with torch DDP with mixed precision training
colossalai run --nproc_per_node 4 finetune.py -p torch_ddp_fp16

# train with gemini
colossalai run --nproc_per_node 4 finetune.py -p gemini

# train with low level zero
colossalai run --nproc_per_node 4 finetune.py -p low_level_zero

Expected F1-score will be:

Model	Single-GPU Baseline FP32	Booster DDP with FP32	Booster DDP with FP16	Booster Gemini	Booster Low Level Zero
bert-base-uncased	0.86	0.88	0.87	0.88	0.89