-
Notifications
You must be signed in to change notification settings - Fork 1
/
beaker.yml
44 lines (44 loc) · 2.03 KB
/
beaker.yml
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
version: v2-alpha
# NOTE: change 'description' each time you submit a new job
description: Same as 029 but smaller LR
tasks:
- name: train
image:
# NOTE: Update this when new commits are pushed.
# Images are built from this GitHub repo:
# -> https://github.com/epwalsh/torch-beaker-image/
# and pushed to this Docker Hub repo:
# -> https://hub.docker.com/repository/docker/epwalsh/allennlp-beaker
docker: "epwalsh/allennlp-beaker:1bf7fc7c6dae865a92bccbde97f3c3b94446cc84-cuda101"
result:
path: "/output"
envVars:
- name: CONFIG
# NOTE: Update this everytime you push a new config. If you're using the same file,
# all you have to do is replace the commit SHA with the new one.
value: "https://raw.githubusercontent.com/epwalsh/vilbert_vqa/b4915c76af299c5490024f2319acdeed3a95faea/vilbert_vqa_from_huggingface.jsonnet"
- name: OVERRIDES
# NOTE: Update this to change the random seeds.
value: '{"random_seed":4,"numpy_seed":4,"pytorch_seed":4}'
# This is a beaker hack to enable shared memory on the container.
# It's needed to use multi-process data loading, which needs shared memory
# to send tensors between processes.
- name: BEAKER_FEATURE_SHARED_MEMORY_OVERRIDE
value: "true"
datasets:
# This dataset contains the images and the image feature cache.
# -> https://beaker.org/ds/ds_zbkh08o8ujxu/details
- mountPath: "/data/vqa"
source:
beaker: "ds_zbkh08o8ujxu"
context:
# NOTE: This runs on Google Cloud, which currently uses CUDA 10.1. That's why the
# tag for the Docker image above ends with 'cuda101'.
cluster: "ai2/shared-v100-1x-8x"
# Replace with this for on-prem servers.
# NOTE: On-prem servers use CUDA 10.2 or 11.0, which means you'll need to use
# a different Docker above. Just replace 'cuda101' for example with 'cuda102' or 'cuda110'
# in the tag.
# cluster: "ai2/on-prem-ai2-server"
resources:
gpuCount: 1