skypilot

1
name: resnet50-randaug
2

3
resources:
4
  candidates:
5
  - {accelerators: T4:4}
6
  - {accelerators: V100:4}
7

8
workdir: ./examples/benchmark/timm
9

10
setup: |
11
  conda create -n timm python=3.8 -y
12
  conda activate timm
13

14
  # Install SkyCallback
15
  pip install "git+https://github.com/skypilot-org/skypilot.git#egg=sky-callback&subdirectory=sky/callbacks/"
16

17
  # User setup
18
  git clone https://github.com/rwightman/pytorch-image-models.git timm
19
  cd timm
20
  git checkout v0.5.4
21
  pip install -r requirements.txt
22

23
  # Apply the patch to enable SkyCallback
24
  git apply ../callback.patch
25

26
  # Apply the patch to use a dummy ImageNet dataset to avoid data downloading
27
  git apply ../dummy_dataset.patch
28

29
run: |
30
  conda activate timm
31
  cd timm
32
  python3 -m torch.distributed.launch --nproc_per_node=4 train.py \
33
    -b 64 --model resnet50 --sched cosine --epochs 200 --lr 0.05 \
34
    --amp --remode pixel --reprob 0.6 --aug-splits 3 \
35
    --aa rand-m9-mstd0.5-inc1 --resplit --split-bn --jsd \
36
    --dist-bn reduce
37