skypilot

1
# Runs ResNet in inference mode on the fake_imagenet dataset
2
#
3
# Usage:
4
#   sky launch -c infer resnet_inference_app.yaml
5
#   sky down infer
6

7
name: resnet-inference
8

9
resources:
10
  accelerators:
11
    V100: 1
12

13
file_mounts:
14
  /tmp/resnet-model-dir: 
15
    source: s3://mluo-resnet-model-dir
16

17

18
setup: |
19
  git clone https://github.com/concretevitamin/tpu || true
20
  cd tpu
21
  git checkout gpu_train
22

23
  . $(conda info --base)/etc/profile.d/conda.sh
24
  pip install --upgrade pip
25

26
  conda activate resnet
27

28
  if [ $? -eq 0 ]; then
29
    echo "conda env exists"
30
  else
31
    conda create -n resnet python=3.7 -y
32
    conda activate resnet
33
    conda install cudatoolkit=11.0 -y
34
    pip install tensorflow==2.4.0 pyyaml
35
    pip install protobuf==3.20
36
    cd models
37
    pip install -e .
38
  fi
39

40
run: |
41
  cd tpu
42
  . $(conda info --base)/etc/profile.d/conda.sh
43
  conda activate resnet
44

45
  export XLA_FLAGS='--xla_gpu_cuda_data_dir=/usr/local/cuda/'
46
  python -u models/official/resnet/resnet_main.py --use_tpu=False \
47
  --mode=infer --data_dir=gs://cloud-tpu-test-datasets/fake_imagenet \
48
  --model_dir=/tmp/resnet-model-dir --amp --xla --loss_scale=128 \
49
  --infer_batch_size=8 --infer_steps=10000
50